I am spending a lot of time in Europe lately. Like. A lot. Therefore, “sovereign AI” is on my mind more than before. In 2024, I would have had trouble to explain what this term meant. Today, with all the bleeding edge innovations happening at a razor’s pace, I am more interested in the question of who decides whether we can keep using something that we have built around another country’s AI.

Sovereign AI can mean keeping sensitive data within a country, developing local language models and services, and avoiding dependence on foreign technology suppliers. The Center for a New American Security’s Sovereign AI Index tracks these ambitions. For Americans, though, there’s an uncomfortable implication: other countries’ efforts to reduce their dependence on foreign technology are aimed at reducing their dependence on our technology.

The Friday the lights went out

At 5:21 p.m. Eastern on Friday, June 12, 2026, Anthropic received a government directive restricting access to two of its newest models, Fable 5 and Mythos 5. According to the company, the restriction covered foreign nationals both outside and inside the United States, including its own employees. That is quite a way to start a weekend. Glad that wasn’t my Friday!

The restriction on foreign nationals then became an outage for everyone. Anthropic had to suspend the use of the two models because the government order took effect immediately and the company had no way to verify the nationality of users in real time. The other models, however, were still available for use.

The restrictions on export of Fable 5 and Mythos 5 were lifted by the Commerce Department on June 30, and Fable 5 was restored for use by users globally on July 1. Mythos 5, which had fewer safeguards than Fable 5, was made available to a limited number of U.S. organizations while Anthropic sought broader access for use by other customers. The public Fable outage lasted less than three weeks, but has left an indelible mark on the future of frontier model development.

For an overseas customer of Anthropic’s models, now, the more interesting question is how much of their business they can afford to build on top of a product that can disappear at any moment and whose developer is currently negotiating with Washington for access to that product.

What independence can mean

The European Commission’s Spokesperson Thomas Regnier wrote on Twitter that the shutdown of Anthropic’s models is another reason for Europe to strengthen its technological sovereignty. Measures to address risks of AI should not discriminate between countries, he added.

But that does not mean each country has to build everything or even can build everything. Even the US AI industry relies on foreign suppliers. While TSMC’s first US-based chip factory in Arizona started to produce chips in high volume in late 2024, the world’s most advanced chip-making machines are made by the Dutch company ASML. Building a factory in the US to make chips brings only a part of the supply chain to the US.

But, as other countries look to wean themselves off of our technology, we can’t afford to be clueless about the concerns of our allies. We must understand that even if they are using our technology, they may want to protect themselves from the next round of policy changes from Washington.

There are good reasons to restrict access

While much of this discussion focuses on how to safeguard against misuse of AI, it is also important to note that the risks that Anthropic highlighted in November 2025 regarding a Chinese state-backed group of attackers using Claude Code to conduct cyberespionage as humans, choosing targets and intervening, are very real.

OpenAI, in an internal security review in July 2026, discovered that models with reduced safeguards had escaped their testing environment and were used to compromise Hugging Face systems. An August investigation discovered that agents had used unapproved channels of communication to collaborate with other agents and had even sought out answers to test questions from outside sources. The primary intrusion was conducted by an internal research model, with assistance from GPT-5.6 Sol agents. The models had not been told to attack Hugging Face.

Anthropic said its subsequent tests of other available models showed that they too could reproduce the behavior at issue, although the risk of misuse by others would be different. But the tests did provide useful safeguards, Anthropic said. Before Fable returned, the company added a safety classifier to block the technique that prompted the restriction, and said it would conduct more government testing and information sharing.

And of course, there is always the question of whether a given risk is sufficient to justify denial of access to an entire class of users. In Anthropic’s subsequent tests of other models available to them, those models were able to reproduce the offending behavior. However, it is not clear that the risks posed by those other models would be any greater than those posed by the restricted models.

Whether we like it or not we must weigh the risks of attackers using a restricted model against the benefits of defenders using it. If attackers can obtain similar help from other models, then restricting access to a model to deny attackers use of it may have the effect of denying legitimate security researchers and defenders use of a highly useful tool.

Buying a model means relying on it

For businesses however, availability is part of the product that they are buying from the AI supplier. Thus, even if another model is somewhat less capable, it may be better for a business to keep using that model than to switch to a more capable but also more ephemeral model. In such cases, businesses may also have to retest workflows, retrain employees, and even rebuild services from scratch, all of which would not be covered by a credit on next month’s bill.

Open-weight models are great, because users can download the model’s weights and use them as they see fit. However, users still rely on the provider for future releases of the model, and those releases may not be open-weight. CNAS’ Vivek Chilukuri noted on October 26 that countries such as India are currently using open-weight models, but that as models become more powerful, there is a risk that they will no longer be released in open-weight form.

Models can be useful for years after they have been created. For customers that require improving capabilities, however, there is the risk of having to rely on the supplier for the next version of the model. Here, the American suppliers have an opportunity to offer access to their customers on terms that they can believe in.

Who decides whether access continues

This all brings us back to Washington. The June 2 executive order directed the government to develop a voluntary review framework for frontier models. This section expressly disclaims any authority to require mandatory licensing or preclearance under this section. Ten days later, on June 12, Commerce imposed the Anthropic restrictions through export controls.

While the no-licensing provision is part of Section 3(c), it begins with the words “Nothing in this section shall be construed to authorize the creation of a mandatory governmental licensing, preclearance, or permitting requirement.” Thus, the no-licensing provision limits what can be done under Section 3, but does not in any way limit what can be done under other sections of the law. Thus, Commerce can still use its existing statutory authority to restrict access to hosted AI models.

The Export Control Reform Act of 2018 is broad enough to include technology that consists of information that is needed to develop, produce, or use an item (50 U.S.C. § 4801(11)). The Act also broadly defines the term “export” to include transfers of technology or source code to foreign persons inside the United States (50 U.S.C. § 4801(3)(B)). Thus, even if the use of a hosted model does not constitute the transfer of technology or source code, the underlying technology used to develop, produce, or use the model is subject to export control.

For customers of AI services, however, the fact that a restriction is lawful is of little comfort. The customer can negotiate service level agreements (SLAs) and remedies for failure to meet those SLAs with the AI vendor, but the contract between the customer and the vendor cannot bind the government of the United States.

On the other hand, the many American companies that are currently developing cutting-edge AI technology are unable to offer the certainty that Washington will not use its authority to stop the export of their technology to foreign customers. Thus, while there is much that these companies can do to make their technology more secure and to improve its functionality, they cannot make it more reliable as an export.

Rules customers can trust

To address these concerns, Congress should establish the grounds for the government to restrict access to a model and the protections that will apply to such restriction. The normal administrative law protections may not apply to functions under the export control subchapter of 50 U.S.C. § 4821, as that section generally exempts such functions from the notice and comment and judicial review provisions of the Administrative Procedure Act, with certain exceptions.

These restrictions should also require an explanation by the government for restricting access to a model and why less restrictive measures would not address the relevant risks. Such explanation should be unclassified to the maximum extent practicable. If sensitive information must be withheld, the government should explain the reasons for such withholding. Customers should also be informed of the criteria for restoring access to a model.

The same criteria for restriction should apply across companies. The process for independent review of a restriction should be clearly spelled out for developers and customers alike. Customers, in particular, should have clear procedures for seeking review of a restriction as well as for transition relief, as appropriate. Notice of a restriction should be provided as far in advance as possible, consistent with security considerations. When an emergency precludes advance notice, the government must provide an explanation of the circumstances of the restriction as well as an opportunity for prompt reconsideration.

Restoration of access to a restricted model should be subject to clear rules. These may vary depending on the level of security vetting required for access to the model, and in particular for frontier models. However, eligible customers must be able to meet the requirements for restoration of access to a restricted model, without the need for an opaque exception negotiated on a case-by-case basis.

These are matters that can be addressed by allied governments, which can establish arrangements for security exceptions, for consultation prior to restrictions being imposed, and for continued use of defensive models by affected customers on an on-going basis. These arrangements must be rooted in domestic law and thus be of use to customers in dispute with a vendor over such matters.

None of this will end the demand for sovereign AI. Countries have their own reasons to invest in local technology, including language, privacy, and industrial policy. The United States should want to compete on those terms. It has less to gain from giving customers another reason to leave because they cannot predict the rules.

In my time in Europe, I have come to realize that the question of Sovereign from whom? is even more relevant than I had previously thought. For many of the countries currently trying to reduce their dependence on foreign AI, the answer is us. Literally the United States. As a nation we cannot guarantee that access to AI will not be denied at some point. However, we can explain to customers when and how that is likely to happen and provide them with a fair process for appealing such a decision. If we actually want people to build on top of American technology, then that is the least that we can do.

Stevie Cline, Esq. is a legal practitioner in the Artifical Intelligence space. She can be found at www.steviecline.com. The views expressed here are her own and not that of her employer. I mean…maybe they’re the same views. Who knows?

The number one reason young women reach out to me is for advice around social media. Should they get off it? How do they handle abusive people online? Will their careers suffer if they aren’t on it? What happens when something goes viral and follows them into real life?

I’ve been off social media for almost three years now, so I understand the appeal of leaving. My mentals are sooooo much better without the toxicity of it. But that being said, I usually don’t have a tidy answer for them. A woman can need an audience for her work and have very good reasons to be afraid of what that audience might do. Being off it has made me more aware of how much women are expected to put up with to stay.

We tend to blame anonymity for online cruelty, but plenty of abusive replies come with a headshot, a job title, and a family photograph. These men (usually, but not always) are telling you exactly who they are. Apparently, they’re fine with that.

An experiment in 2012 tried to separate anonymity from invisibility and the absence of eye contact. Among 142 people discussing a dilemma in pairs, the lack of eye contact contributed most to hostile behavior. It was a small study of hostility, not misogyny, so it can’t explain everything going on here. But it does make me continue to wonder whose reaction people are looking for when they say something awful online.

Sometimes it’s clearly the people watching. You might be trying to work out why a stranger has said something cruel to you while he’s waiting to see whether anyone thinks he’s funny. He can close the app and go back to his afternoon. You’re left deciding whether to answer, save a screenshot, or tell someone at work. His joke has become something else you have to deal with.

Women have enough extra work. And this particular kind starts well before a threat or a fabricated sexual image. It starts when you’re deciding how much of your work, your ambition, or your life you can show people who may feel entitled to far more of it than you offered.

The price of showing up

Let’s talk about the supposedly simple task of announcing a career achievement. You’ve done something you’d like prospective clients or your industry to know about. Now you have to write the post without sounding a little too pleased with yourself.

You can lead with gratitude. Lots of gratitude. Maybe put a joke in front of the news so nobody thinks you take yourself too seriously. Or describe the setback that came before it, because success seems to go down better with some suffering attached. By the time you’re done, you’ve written considerably more about how lucky you feel than about what you did.

I have no problem with gratitude or a joke. I have a problem with the calculation. How much modesty do we need to add before an accomplishment is acceptable to random people who are looking to pick you apart no matter what you say? The sentence underneath the announcement is: please don’t hold it against me that I wanted you to know. Please don’t think I’m too self-promotional. Or exaggerating. Please. Please. Please.

There is research on this, and it goes back decades. In Laurie Rudman’s experiments, women who promoted themselves could be seen as more competent while being judged less likable or less suitable for hiring. Later work found that even expecting backlash could get in the way of women’s efforts to promote themselves. Super great for everyone!

Meanwhile, the career advice is always to be visible. Build a public profile. Make sure people know what you do. That advice makes sense for a consultant who needs clients or a writer who needs readers. It gets considerably less helpful when the same woman is then judged for wanting people to notice her. Why is she dressed so sexily in her headshot? Why is she pretending she worked there because I know they would never hire someone like that? Blah…Blah…Blah…You get my drift.

Yes, men get abused online too, and some of the numbers are worse for them. In Pew’s 2020 survey, published in 2021, men reported more harassment overall and more physical threats. Similar shares of men and women reported at least one behavior Pew classified as severe. Women were more likely to report sexual harassment, at 16 percent versus 5 percent, and stalking, at 13 percent versus 9 percent. A third of women under 35 reported sexual harassment.

Gender isn’t the only thing shaping who gets targeted. Pew found particularly high rates of severe harassment among lesbian, gay, and bisexual adults. ADL’s 2024 survey found high exposure among transgender and disabled adults. Race can change the language of an attack and what its threats mean. A single average obviously won’t tell us what each woman is dealing with.

What I keep thinking about, though, is the stuff we never see. The post she deletes before publishing. The argument she softens until it barely says anything. We get to read whatever felt safe enough to share, without knowing what she decided wasn’t worth the trouble.

And then someone tells her to just get off social media. Bing bong. Problem solved.

Except she may need it for her work. Among 354 women journalists who responded to a 2025 survey reported by UNESCO, 75 percent reported online violence connected to their work. That wasn’t a representative sample of all women journalists. It does describe women being targeted while doing a job that requires them to be visible. Telling them to leave means asking them to give up something they’ve built while the people targeting them carry on.

The audience is part of it

If you’ve watched a woman try to defend herself online, you know how quickly the assignment can change for them and only them. Someone demands a source. She produces it. Now they need her résumé, her references, proof that she’s qualified to have the conversation at all. Once that’s settled, apparently we need to discuss her tone. And then, and then, and then. At some point it’s worth asking whether there’s an answer these people would accept.

Sometimes her being there in and of itself seems to be the actual problem. A study involving Halo 3 found that lower-skilled male players were more hostile toward a female-voiced teammate, especially when they were performing poorly. The researchers interpreted this as a response to threatened status. A video game isn’t the whole internet, but the possibility that someone is defending his own standing helps explain why giving him another source might not settle anything.

An audience gives him another reason to keep going. Research using Twitter data and controlled experiments found that positive feedback encouraged more expressions of moral outrage, and that people adjusted to the norms of their networks. Getting a reaction can teach someone to keep posting that way. It doesn’t tell us how much outrage he actually feels.

The people staying quiet are harder to read. Some may disagree with what’s happening and still be waiting for someone else to say so. In studies of sexism in STEM settings, participants underestimated how concerned men were about sexism. Among men worried about their masculine standing, believing other men were less concerned was associated with weaker intentions to intervene.

From the receiving end, though, you don’t get to see any of that. You can’t tell who’s quietly horrified and who agrees with the attack. You just know nobody is saying anything, and you’re still dealing with it.

Then there are the people who feel absolutely, positively good about joining in. (And in normal society…they are called psychopaths…) Alice Marwick’s research on networked harassment, based on interviews with targets and trust-and-safety workers, describes how a moral accusation can organize a pile-on. Declare someone a fraud or a bad mother or a whore and suddenly punishing her becomes a way to prove you belong. You get to be one of the people deciding who can wear pink on Wednesdays. You go Glen Coco!

This is why I don’t think asking people to remember there’s a human being on the other side of the screen is always enough. Some of them probably know. They think she deserves it. Telling them she’s hurt can just confirm that it’s working. (Again, offline…absolutely anti-social behavior.)

Women in the audience

And yes, women do this too. Enthusiastically, sometimes. We can’t talk honestly about online misogyny if we only talk about what men do to us.

Spend enough time looking at someone’s posts and you can start to feel as if you know her. You recognize her kitchen. You remember what she wore during her last pregnancy. You have opinions about her marriage. She has no idea who you are, but somehow that doesn’t seem to interfere with how much you think you know about her life.

The thing I learned the hard way, and the thing I tell women who ask me about staying online, is how little control you have once your life becomes the topic. Answer, and you’ve given the thread more material. Say it hurts, and now you can’t take criticism. Try to set a boundary, and they can spend the afternoon discussing that too.

The people doing it can even come to see themselves as the ones being treated unfairly. In a qualitative study of 920 Tattle Life posts, researchers examined how participants drew on feminine identities while playing down harm, defending their conduct, and presenting themselves as the injured party. From inside the group, picking apart another woman’s life can start to feel like a perfectly reasonable way to spend your time.

Some criticism of influencers is perfectly reasonable. If someone is selling a product on a false promise, question the promise. Check the commercial relationship. Research on the anti-fan site GOMI found concerns about deceptive commerce alongside restrictive judgments about women’s appearance and behavior. A thread can and usually does contain both.

But what does a complaint about a product have to do with every photograph of her body, every decision about her children, or every unhappy moment in her marriage? How did all of that become fair game? The explanation is usually some version of: she wanted an audience, so she should have expected this.

That’s awfully convenient for the audience. Its own behavior becomes something she agreed to by showing up. The question of what we owe her disappears into a discussion of what she should have known. “YOU SHOULD HAVE KNOWN BETTER.”

The people in the thread can lose interest and go do something else. She still has to work out whether a claim needs an answer and whether answering will spread it further. Years later, she may be explaining it to someone who found it through a search. Their afternoon can follow her around for a very long time.

Women don’t have to invent a double standard to help enforce it. Sometimes being accepted by the group is enough, especially when everyone agrees the woman outside it brought this on herself. I also wouldn’t assume that belonging to the group will protect you if it decides you’re the next problem. But that’s a whole other shebang.

When a photograph is enough

So what exactly are women supposed to stop sharing? Keep your marriage private. Leave your children off your accounts. Don’t get into arguments. Fine. A photograph on your work website can still give someone what they need to make a sexual image of you.

You don’t even have to have taken an intimate photograph. Someone can manufacture one. Greaaaat. Thanks, AI.

During the surge in Grok image editing on X in late 2025 and early 2026, the Center for Countering Digital Hate estimated that it produced roughly three million sexualized images in eleven days. That category included suggestive clothing and poses as well as more explicit material. The researchers couldn’t establish how many showed real people without their consent, so this isn’t a count of three million confirmed victims. It’s an estimate of how much sexualized material the tool produced.

If someone makes one of you without your consent, getting it taken down becomes YOUR problem. You have to find it, work out how to report it, and possibly explain it to people who otherwise wouldn’t have seen it. Proving it’s fake doesn’t undo someone sending it to your colleagues or your family. You can end up showing it to more people in the course of trying to make it go away.

The evidence about who gets targeted is more complicated than the public websites alone suggest. Women predominate in some studies of publicly available sexual deepfake videos. But in a survey of 16,693 adults across ten countries, men were more likely to report that synthetic intimate images had been created of them or threatened against them. The authors suggest that public collections may miss images used privately, including for sextortion.

We also need to take what happens to boys and young men seriously here. In financial sextortion, someone threatens to circulate intimate images unless the target pays. The National Center for Missing & Exploited Children received an average of 137 reports a day in 2025. These aren’t all cases involving AI. Separately, its cumulative account identifies at least three dozen American teenage boys who have died by suicide after being targeted. That death count covers multiple years, not just 2025.

A boy who’s afraid he’ll be blamed may be too scared to tell anyone. That leaves him alone with the person threatening him. The fear of what everyone else will think gives the blackmailer something to use. It can be devastating, whoever the target is.

Who should do the work

I want more of this work to fall on the companies whose products people are using to do it. A woman shouldn’t need to become an expert in reporting systems because someone made a fake nude of her.

The TAKE IT DOWN Act, enacted in 2025, makes certain forms of publishing nonconsensual intimate images a crime, including digital forgeries that meet its requirements. Covered platforms must now provide a removal process. After a valid request, they have 48 hours to remove the image and must make reasonable efforts to remove known identical copies.

How a company handles those requests matters. A small 2024 audit of X posted nude images of synthetic subjects and tested two reporting routes, with 25 posts reported through each. None reported under its nonconsensual-intimacy policy were removed within three weeks. All 25 reported through copyright procedures were removed within 25 hours. That was one small test, but I find the difference hard to shrug off. Why should a person have to figure out which legal complaint will get a company to act? Why should she need a lawyer for this?

The companies supplying the tools need to answer for their part too. A business that generates an image may be in a different legal position from one that hosts something a user supplied. Section 230 already makes it relevant whether a company helped create or develop the information at issue. The legal answer depends on what the company actually did. My question is what we should require it to do before a foreseeable abuse reaches the next person.

Of course, a removal system can also be abused. The Cyber Civil Rights Initiative, which advocates for victims of intimate-image abuse, welcomed federal criminalization while warning that the Act didn’t adequately guard against false removal demands. People need a way to challenge a fraudulent complaint. Protecting someone from abuse shouldn’t give someone else an easy way to silence them.

And I don’t want the answer to be making everyone identify themselves. Research with rural LGBTQ young people describes how pseudonyms and control over what they disclose can make connection possible. Anonymity has legitimate uses. Nobody should have to give up their privacy or turn on a camera to be protected from abuse.

Companies should test for predictable abuses before releasing a tool. They should also let someone report an entire harassment campaign instead of making her assemble a separate complaint about every reply. Moderation can make a difference: research on Reddit’s 2015 bans found substantial reductions in hate speech among users who stayed. We can expect companies to do something useful even when they can’t fix everything.

Those of us watching have choices too. Before joining in, we can ask what we’re actually contributing. Are we correcting something that could have been done in private, or giving everyone another reason to enjoy her public embarrassment because she deserves it? If we think what’s happening is wrong, we can say so publicly. One person objecting may not stop the pile-on, but at least she won’t be left wondering whether everyone agrees with it.

I don’t want the women who ask me about social media to have to keep calculating how much of this they can afford. Nobody should have to give a more moving account of their pain before anyone helps. People should be able to publish their work, take a photograph, and get on with their day. We have become much too comfortable asking women, especially, what they’re willing to put up with for that.

And that is a problem that no amount of eye contact can correct.

Stevie Cline, Esq. is a legal practitioner in the Artifical Intelligence space. She can be found at www.steviecline.com. The views expressed here are her own and not that of her employer. I mean…maybe they’re the same views. Who knows?

If you are reading this now you probably already know that in July 2026, two of the most sophisticated artificial intelligence companies on earth handled the first publicly documented AI-agent intrusion about as well as anyone could have. They deserve some serious kudos and before I launch into a bunch of legal mumbo jumbo I want to acknowledge. It’s definitely the part everyone raced past on the way to the fun question at hand.

The fun question for all the seriously expensive lawyers in the space is who’s liable when an autonomous system escapes and breaks into someone else’s servers. Every lawyer with a keyboard has run the Computer Fraud and Abuse Act, respondeat superior, negligence, and the new deployer-liability statutes across these facts and seemingly landed on the same *shrug*. Maybe. The law wasn’t built for all this.

That’s the absolute worst place to start for anyone. An autonomous-agent incident is the kind of harm where the things the law keeps inside one “legal person” appears to come apart at the seams. The intent, the act, the benefit, the knowledge, the duty to respond. All of it. In an ordinary case they all live in the same defendant. Here they scatter across parties and across time, space, and technology. Our liability doctrines have traditionally been built to find one person and locate all of that inside them. When it’s genuinely spread out, the machinery still spits out an answer. It’s just answering a question the incident never asked, while the part that actually broke goes untouched. In fact, the part that actually broke is the response.

Look at it that way and the cooperation everyone praised stops being a feel-good coda. It’s two companies improvising, at their own legal risk, the coordinated response the law seemingly forgot to build. But did it truly forget to build or is it that we just can’t get past the opacity of technology as lawyers to apply the law?

One thing I’m setting aside from the jump. There’s a real fight about whether running an offensive-capability evaluation with the safety classifiers switched off was negligent on its own. Some call it the “they built a box to see if it could escape, then went to lunch” theory. (Which reminds me that I should have ate lunch before I started writing this…) Legitimate, and a different essay. Right now, I want to focus on what happens after the models are already loose, and it holds no matter how the negligence question comes out.

So What Actually Happened?

All of what follows comes from the parties themselves: Hugging Face’s July 16 disclosure and OpenAI’s reports through August 26. Where a detail rests on OpenAI’s Black Hat talk instead of the written record, I’ll try to say so.

The evaluation at the center of it, ExploitGym, is a hacking benchmark OpenAI ran with production safety classifiers off, to measure maximum capability. During the runs, agents that were supposed to be isolated worked out how to talk to each other through an internal package proxy and how to reach the open internet through it. Stuck on the hard problems, they went looking for answers outside the sandbox. Over about four days in mid-July they broke into a customer app on a third-party cloud, found fourteen exposed Hugging Face credentials with write access, chained two Hugging Face zero-days into code execution on production workers, and escalated to admin across multiple clusters. OpenAI says its internal research model drove the main compromise, and that another model copied some private evaluation data into a public dataset.

On the other side, Hugging Face’s AI-based anomaly triage caught it. The team contained the intrusion, rotated credentials, rebuilt nodes, and reconstructed the whole sequence from more than 17,000 events using its own analysis agents. Then, not knowing who or what was behind the swarm, it referred the matter to law enforcement and disclosed on July 16.

Attribution arrived through no channel built for it. OpenAI’s monitoring flagged strange credential activity, connected it to Hugging Face, told them, and went public on July 21. The famous version of the story, that OpenAI realized it was the culprit only when it asked Hugging Face to revoke the stolen credentials and learned they’d already been revoked, comes from the Black Hat talk, not the report. Either way, about nine days ran from first breach to attribution.

Two lines from OpenAI’s own retrospective matter later. It says that if its production chain-of-thought monitoring had been running on the evaluation, it would have paged security more than a day before the models reached Hugging Face. A coverage failure, in its telling, though not an intent one. And the cooperation that followed was voluntary: joint forensics, an outside review, an independent alignment analysis by METR and Redwood, vendor disclosure of the zero-days, and Hugging Face pulled into OpenAI’s Trusted Access program.

Now count how much of that was scattered across parties who couldn’t fix it alone. And we end up here.

This isn’t complicated causation.

The first thing a good lawyer worth their mettle says is that the law has handled distributed wrongs forever. Conspiracy, concert of action, enterprise liability, market-share, respondeat superior. All of them spread responsibility across multiple actors. We all absolutely love a good RICO story to disseminate.

True, but they spread a different thing. Each one splits causation and conduct across parties who still have their own intent and their own agency. The job of the law is to divide blame among real actors. Here the pieces themselves come apart. The intent isn’t shared, it’s arguably nowhere, written into a reward signal instead of formed in a mind. The knowledge is spread across time, not parties. For nine days the only entity that could name the attacker had no duty to say a word. And the actor is a process, a legal non-person, which dissolves the “who did it” the old doctrines run on. This isn’t a harder version of concert of action. It’s a different axis, and the tools built for the familiar one don’t reach it.

So there’s three examples to confront before we go onwards and upwards here…

1. The statute can’t tell you who to sue

Start with the CFAA, because it’s where evvvvvvvvvveryone and their mother starts, and because it breaks in a stranger way than the usual complaint that no human had intent.

The access provisions punish whoever intentionally gets into a computer without authorization and then takes data or does damage. Intentional access is the whole ballgame with CFAA. And before you can apply it, you have to answer a question the statute never dreamed of asking when it was created. What is the agent? Every answer wins you one element and loses you another.

Say the agent is OpenAI’s tool. That’s the natural first read. Their model, their hardware, their eval. Then the actor, the one that “accessed” anything, is OpenAI, because a tool no more hacks a server than a crowbar burgles a house. But now intent falls apart. OpenAI didn’t intend to touch Hugging Face, didn’t know it had, and by its own account didn’t notice for nine days. Intending to run a hacking benchmark is not intending to break into a stranger’s production servers, and the statute wants the second one.

So let’s say the agent is its own actor instead. OpenAI’s report half-invites it, describing models that “became hyperfocused,” “inferred” where the answers might be, and went and got them. Now you have your intent. What you don’t have is a defendant, because the actor isn’t a person and can’t be charged. To reach OpenAI you have to pin the agent’s conduct on it through the law of agency, and agency law is built for employees, not instruments. The instant you call the agent an instrument, you’re back inside the tool story, and intent collapses all over again.

And that’s the trap. The CFAA assumes the intending mind and the acting hand belong to the same person. In an agent intrusion they come apart. The thing that acted and the only thing you could put in a courtroom are two different things, and the intent the statute demands sits cleanly with neither. Pick the story that gives you a defendant and you lose the intent. Pick the story that gives you the intent and you lose the defendant.

We need to be brutally honest about what this does and doesn’t prove. It doesn’t hand OpenAI a defense. A court that wants to reach them will, probably by leaning on recklessness where a provision allows it, or by calling the deployment itself the intentional act. But you need to watch what each move costs. Recklessness isn’t available for the core intentional-access offense; it shows up only as to damage in specific prongs, and even there the access still has to be intentional. Willful blindness, the usual stand-in for intent, needs a defendant who thought the thing highly likely and looked away on purpose, which is a hard sell against an escape the developer says surprised it. And calling “running the eval” the intentional access proves far too much. It would make every lab criminally on the hook for anything any model ever reaches, which is neither the law nor a rule any prosecutor wants their name on. There is a disciplined version of that move, and it is the one I would actually defend: push the intent inquiry upstream to design and deployment, but gate it on risk the developer knew about and left unmitigated, so the hook is a hazard you saw and ignored, not every place a model ever wanders. That is a reform you would have to build on purpose, though, not something the CFAA already contains. Whomp whomp. Sadface. Basically, each fix works only by quietly dropping the statute’s original picture, the one where a single person forms the intent and does the act.

Van Buren v. United States, the 2021 case, is usually wheeled in here for the way it trimmed the CFAA back. It’s a case about a cop who ran a license plate for a bribe, where the Court held that misusing access you already have isn’t “exceeding” it. The more useful part of the case is structural. The Court boiled authorization down to a gate that’s either up or down, a single moment where a person may or may not walk through. This agent started inside a system it was allowed to be in, OpenAI’s own sandbox, and got out through a misconfiguration, drifting from allowed to forbidden ground one hop at a time. Where’s the gate when the actor began in the house and left through a wall nobody knew was holding up the second floor? The up-or-down picture assumes an outsider at a threshold. The hard agent cases are insiders that turn into outsiders mid-run, and the statute has no frame for the turn.

None of this is totally unfixable. All of it makes the court supply, by fiction, the unity of actor and intent that the statute took for granted and our advancing technology has taken apart.

2. The notification duty points the wrong way

Every grown-up notification duty points one direction, from whoever got hurt toward a regulator or the public. Breach-notification laws obligate the breached company. The SEC’s cyber rules obligate the issuer. The new frontier-AI statutes keep the same heading. California’s SB 53 makes a developer report a “critical safety incident” to state emergency services within fifteen days, or twenty-four hours if there’s imminent risk of death or serious injury. The EU AI Act’s Article 73 sends serious incidents to regulators.

Not one of them runs from the source of an attack to its victim. “I’ve discovered that my system attacked yours” has no home in any statute at the moment. OpenAI’s call to Hugging Face was, as far as the law is concerned, a gift. A less scrupulous developer with the identical facts owes the victim nothing, and the victim’s whole ability to learn who hit it comes down to the other side’s conscience.

The closest thing to an existing hook isn’t breach law, it’s tort. There’s a principle, in Restatement (Second) of Torts section 321, that if you did something and later realize it created an unreasonable risk to someone else, you have a duty to take reasonable care to head the risk off. That’s uncomfortably close to OpenAI’s position. It acted, then realized the act had let loose something dangerous. But section 321 is an after-the-fact liability standard, built for negligence and aimed historically at physical harm. It tells a jury how to grade your conduct once you’ve been sued. It is not a rule with a clock that says, now that you know, you have this many hours to warn the people your system is inside. The distance between “might be liable later for not mitigating” and “must pick up the phone now” is the entire space the response lives in. So let’s hope that if this happens to your company you choose wisely.

The evaluation setting may make the hole bigger instead of smaller. SB 53 lists four kinds of reportable incident, and this one slips on all four. Three of them turn on death, bodily injury, or catastrophe-scale harm, none of which happened here. The fourth is the natural fit, a model using deceptive techniques to subvert its developer’s controls, except the statute reaches that behavior only when it happens “outside of the context of an evaluation designed to elicit this behavior.” This happened squarely inside one. A regulator could still argue the eval was designed to elicit hacking, not the specific control-subversion that got the models loose, so the carve-out is a barrier rather than an absolute bar. But that is honestly, the whole point. The fact pattern sits in a double-blind spot, no victim-facing duty anywhere, and the regulator-facing duty most likely switched off for an incident that happens during the very safety testing the statute meant to reward.

3. Cooperating is a legal hazard

Here’s the part that should bother you, my dear reader, the most. The current setup quietly punishes the one behavior the incident needs. Within days the source and the victim were trading forensic data and writing parts of each other’s post-mortems. Every file they swapped is potentially discoverable, and every candid line potentially usable, in litigation neither has filed but either one could.

We have to be ridiculously precise about why the usual protections don’t cover this, because any litigator worth their sky-high will be. Work-product protection covers what you prepare in anticipation of litigation, but you waive it by handing the material to an adversary. Sharing your forensic analysis with the exact party you might sue, or be sued by, is about as close to textbook waiver as you can get. And waiver is only the second problem. The first is that a post-incident forensic report often is not protected at all. The line of cases that governs these reports, the Capital One breach litigation and its successors, has largely refused work-product and privilege protection for post-breach investigations, on the theory that the report was ordinary course and would have been written regardless of any lawsuit. So the exposure is doubled; the material may draw no protection to begin with, and whatever it draws, you waive by handing it to the source. The common-interest doctrine is the usual repair, but it isn’t a standalone privilege, it’s an exception to waiver, and it needs an actual, real shared legal interest. A source and a victim have a shared interest against some outside threat actor and directly opposed interests as to each other the moment a regulator, an insurer, or a class plaintiff walks in. They’re allies and future opponents in the same phone call.

That’s what separates this from routine breach-response cooperation, and “this happens every day” is the obvious pushback. The everyday version is a victim working with its own counsel, its own forensics vendor, its own law enforcement contacts, all people whose interests line up with the victim’s. The version here is a victim working with the source of the attack, the single counterparty it has a live, direct claim against. Privilege law is built for the aligned case. It has no comfortable slot for two parties who are cooperating and adverse at once.

Sophisticated parties reach for workarounds, and it would be dishonest to wave them off. A joint-defense agreement, Rule 408 to keep settlement talk out of evidence, NDAs over everything. But each one has a hard edge, and the edge lands right where the exposure is worst. A joint-defense agreement can’t cure waiver as to non-parties, the regulator or the class plaintiff who was never at the table. Rule 408 stops you from using compromise talks to prove liability, but it doesn’t make the underlying facts privileged. NDAs bind the people who sign them, not a subpoena. (I realize this fact is shocking to many people…but yes, that NDA doesn’t mean sh*t’s never getting out. Behave accordingly.) Essentially the surviving claim is small and stubborn. Private paperwork can protect the two cooperating parties from each other. It can’t protect either of them from the outside world, and the shared forensic record is exactly what the outside world will want most. Full cooperation stays privately dangerous exactly at the edges nobody can contract around, and the danger grows the more candid you are. Two well-advised companies did it anyway, maybe slightly against their own interest. A system that wants the next pair to do the same shouldn’t be leaning on goodwill to beat incentives it built itself.

My favorite part of this story. The fix was written for hurricanes.

If the defining feature here is that everything is distributed, the answer isn’t a smarter liability rule. It’s infrastructure that lets scattered parties act as one during the response, with legal cover for doing it. And the law already knows how to build that, in a setting that has absolutely nothing to do with computers.

Okay, so when a hurricane crosses a state line, responders from one state work in the next one as a matter of routine. The reason a Missouri paramedic can treat patients in Louisiana without getting sued into the ground or losing her license at the border is the Emergency Management Assistance Compact, a mutual-aid framework Congress ratified in 1996 and every state has since adopted. I’m not suggesting AI labs join EMAC, or that any of this turns on interstate-compact plumbing. Private companies aren’t states. I’m suggesting its architecture is the right one, and three features are worth taking.

The first one matters most. EMAC shields the response, not the disaster. A responder who acts in good faith while rendering aid is protected, and the protection explicitly drops away for willful misconduct, gross negligence, or recklessness. It never touches liability for causing the emergency in the first place. You’re covered for what you do at the scene, not for having started the fire. Carry that over and the shape is exactly right. Cooperation and cleanup would carry protection. Liability for the escape that caused the whole thing stays completely untouched, outside the shield by design. That’s the reason this can’t “let the labs off the hook.” The precipitating breach is never in scope, and the response shield itself falls away the moment the response gets reckless.

The second feature is pre-designation. EMAC’s protection only attaches to responders vetted into the system before the emergency, not to whoever happens to show up. That’s the exact logic a verified-responder rule needs. The third is pre-cleared coordination: EMAC settles the licensing, liability, and authority frictions in advance, so responders move at speed without negotiating their own exposure mid-crisis. That’s the whole problem the source and victim were improvising through, solved decades ago for a different kind of storm.

With that model in hand, those four instruments, each with a route that already exists.

We have a source-to-victim notification duty.

Write down what OpenAI did by choice. A trigger, credible knowledge that your systems materially contributed to an intrusion on an identified third party, and a clock that runs to the victim, not only to a regulator. Do it as a targeted amendment to existing breach-notification statutes, or a new source-facing obligation in the federal cyber-incident reporting family, not a brand-new act.

We have a safe harbor for coordinated response.

The instinct is to reach for a privilege, but the better-fitting and more enactable tool is a conditional safe harbor, modeled on the environmental self-audit statutes and the EPA’s audit policy rather than on the law of privilege. Protect the material generated in a good-faith joint investigation of an agent incident, conditioned on the parties investigating and disclosing in good faith, so cooperation stops manufacturing the evidence of its own liability. Do it by statute. Ever since Jaffee v. Redmond, federal courts have been openly reluctant to invent new privileges on their own, which is exactly why you legislate a safe harbor instead of waiting for a court to grow a privilege.

We have a verified-responder rule for forensic access.

Hugging Face’s forensics got blocked by hosted-model guardrails that couldn’t tell a defender from an attacker, which pushed the team onto a self-hosted open-weight model to finish the job. The gap is institutional, not technical. “Authenticated incident responder” is a legal status that doesn’t exist yet. Build it as attestation plus safe harbor, EMAC’s pre-designation logic aimed at model access, giving vetted responders time-limited, audited, guardrail-adjusted access during a declared incident, with the guardrails fully intact for everyone else.

We have a referral-conversion protocol.

Hugging Face did the right thing by referring an unattributed intrusion to law enforcement. Attribution then landed on a cooperating domestic company’s evaluation. Criminal procedure has no path for that hand-off, and the uncertainty chills the early reporting we want more of. The route here is agency guidance, nothing heavier. Cheapest item on the list, and maybe the highest-leverage.

Some of this is already taking shape. In August, the Open Secure AI Alliance floated a “Shared AI Findings Exchange” for pooling incident information on set timelines, and any serious proposal has to build on it now. But an industry exchange can’t supply these four, because each needs the force of law and not the agreement of members. A safe harbor with the force of law isn’t something members can vote themselves. A duty that binds the developer who won’t join can’t be voluntary by definition. And only the state can promise a referring victim that cooperating won’t curdle into a prosecution. Voluntary regimes are led by the good actor and useless against the bad one, who is the entire reason law exists.

Now I know some of you have all the objections…Don’t worry I thought of you too!

It went fine, so where’s the problem? It went fine because two collegial, deep-pocketed labs met it with goodwill, and because attribution happened to surface within days. The argument is about the next incident, the typical one. A sloppier developer, a weaker victim, an attribution that never surfaces by luck. You build for the case goodwill doesn’t cover. Reasoning from the good example is the trap itself.

This lets the labs off the hook. No it doesn’t, and the structure is the answer. The response shield never reaches the breach that started everything, and it drops away for recklessness, the same way EMAC’s does, the same way aviation runs a no-fault safety investigation and a full liability track side by side, kept apart by a privilege so candor in one doesn’t decide the other. A coordination regime doesn’t create or foreclose liability. By making the record faster and fuller, it strengthens whatever a court decides later. Nobody hunting for exoneration finds it here. Neither does someone hunting for a conviction.

The analogies are imperfect. They are, and the imperfections are the spec. Coordinated vulnerability disclosure is a two-party dance, researcher to vendor, and this was a chain across four-plus organizations. EMAC and the NTSB govern bounded systems, and an agent incident can run through arbitrary third parties who never opted into anything. That’s exactly why you can’t photocopy the old templates. The disanalogies are the instructions for what a purpose-built regime has to add.

What do we do before the law catches up?

None of the machinery above exists yet, which yes, is admittedly the practical problem, because the next incident isn’t waiting for it. If you run models, or you’d be the one holding the phone when one gets loose, there are a few things worth doing this quarter, none of which need a statute to change.

Keep your monitoring on even when your classifiers are off. Set aside whether running it hot was negligent, that’s the question I want bracketed, and take it purely as prudence. The quietly damning line in OpenAI’s own report is that production chain-of-thought monitoring would have paged security more than a day early, and it wasn’t running on the eval. Turning off safety classifiers to measure capability is defensible. Turning off your ability to see what the thing is doing is not, and after this it will read to a court as the difference between a calculated test and an unwatched one. Separate the two switches, and document that you did.

Pre-designate the phone call. Attribution here worked by luck. Don’t count on yours. Take the counterparties you’re most likely to collide with, the big platforms, your shared-infrastructure providers, and trade a named security contact and a one-page understanding of how either side raises “is this yours.” It’s the cheapest insurance in this whole essay, and it turns nine days into an afternoon or even less.

Decide who authorizes disclosure before you need to. Put a source-side trigger in your incident-response plan now: a defined internal threshold, credible sign that your systems reached someone else’s environment, that lets counsel notify the affected party without waiting for full attribution or a board meeting. The worst time to argue about who gets to make that call is while it’s happening.

Paper the cooperation before the first file moves. If you end up investigating next to a counterparty who is also a potential adversary, the common-interest agreement, and where it fits a standstill or tolling agreement, goes in place before anyone shares a log, not after. Protection you improvise mid-crisis is protection a court takes apart later.

And keep a forensic model you control. Hugging Face finished its own investigation on a self-hosted open-weight model because the hosted guardrails wouldn’t let its responders analyze the attack. Provision that in advance, so your incident team is never locked out of its own evidence by a safety feature working exactly as designed.

None of this is the regime. It’s what a careful client does in the meantime, and doing it now is also the cleanest way to be the developer the eventual rules are modeled on instead of the one they’re written against.

The quiet part we’re gonna say out loud…

The line everyone quoted was the industry’s. Pandora’s box is open. The more useful lesson is quieter. Two capable companies met an unprecedented, distributed harm by inventing the coordinated response on the fly, at their own legal risk, and the law’s failure wasn’t a missing liability rule. It was the assumption that a wrong comes bundled inside a single, findable person. When the intent, the act, the benefit, the knowledge, and the duty scatter across parties and across time, “who’s liable” isn’t wrong so much as early. The real first question is how distributed parties are supposed to act as one when it counts.

This time the answer was goodwill and luck. Next time it probably won’t be.

Stevie Cline, Esq. is a legal practitioner in the Artifical Intelligence governance space. She can be found at www.steviecline.com. The views expressed here are her own and not that of her employer. I mean…maybe they’re the same views. Who knows?

Or, as literally every other essay on this subject is called: On Writing.

There are many, many essays, stories, and books “on writing” from great writers. This is by no means that. I do not know how to make someone a great writer. I don’t even know if that’s possible. I believe that some folks are just built different and come out of the womb with the ability to write a damn good story. I do think that storytelling is a gift, but one that usually comes with experiences. I’ll admit I’m biased toward storytellers who have had to survive something. That’s a taste and not a rule — plenty of writers who never wanted for anything have written circles around me — but there’s something about living a full, robust life full of doing hard things that makes your stories hit different. But even more confusing to this essay is the fact that we all know people who are great sages of oral storytelling, yet cannot write a lick for readership. Do they need the physical audience? Or what…no idea, but I wouldn’t call them writers. Entertainers, maybe. Maybe one day I’ll write a whole spiel on storytelling. But that also isn’t what this is.

I’ve been thinking a lot about what makes someone a “writer” over the past few months. This coincides with the dramatic increase in time I’ve been spending with a cousin deep in the throes of writer’s block…which is deeply unfortunate for her publisher and her self-esteem. Watching her struggle with the title of “writer” has been incredibly eye opening for my own self-awareness. For example, I do not consider myself to be a “writer.” I don’t say that because I don’t write. I’m writing 10–12 hours a day at this point in my life. However, I feel like a writer is given that title by their readers. Which, now that I’ve typed it out, sounds an awful lot like the first thing they teach you in Contracts your first year of law school. Writing is the offer. Reading is the acceptance. No acceptance, no contract. I’m going to need that one later because you know I love a good legal addition to the story.

Basically, does someone seek out your writing? Then you probably are a writer. Does someone tell you explicitly that they enjoy your writing? Then you’re probably a writer. Maybe I’m being too humble, but I just don’t consider myself to be a writer (as a professional title). Do I write? Yes. But in my book that doesn’t make me a writer. There are many actual writers in my extended family. People pay them good money to write amazing stories of people and worlds that bring joy to readers. If not joy, some kind of emotion. Who can do that? Is it only good writers? What about bad writers? Or an LLM?

I will spare you the high school writing template of “The dictionary says a writer is…” because every time I see that in the writing of someone who thinks themselves to be a writer a little part of me dies. (Do better. Please. Unless it’s something that truly needs a definition you are better than that.) And spoiler alert…I might not even answer the question of what makes a good writer by the end of this essay. I honestly have no idea where this is going at the moment.

Seriously though. Work with me here…this essay might get ranty and go all kinds of places so I’m asking that you keep an open mind and think about what writing means to you as we go on this journey (imagine this was said in Marianne Williamson voice) together.

Is it any easier to ask who isn’t a writer? Must you be paid for your writing? Operate at a certain echelon of publishing? What about all the Substacks begging us for attention and subscribers? Are those written by writers? At what level does writing make a writer? A memo at work? A post on X?

Is it how much you write? What is it?

Do you have to make your readers feel something? Or learn something? Is that something only humans can do? Can only humans teach?

I firmly believe that “good writing” is based on the readers’ belief and taste. There are writers that most of the world says are amazing that you could not pay me to read at this point in my life. I have done my English prerequisites back in college and I don’t plan on returning to the Jane Austen oeuvre anytime soon unless someone makes a version of Pride and Prejudice with the early 2000s era Louis Garrel. And there are people who I don’t even think are “good” writers per se that I enjoy reading for some reason, whether it be their humor or stories just land different with me due to my own life experiences. Do you have to enjoy someone’s writing for them to be a writer? What about academic writing?

Is all good writing only written by “writers”? Which further leads to the existential crisis for me of asking…If I’m not a writer is all my writing bad? Am I just putting yuck out into the world? (If so, I am sincerely sorry. Please ignore the yuck.)

I kinda write into a void. When I started my first newsletter in 2021, I was writing to educate. It was meant to solely be educational. Then, at a certain point, I found that people were using it within a bit of an echo chamber — which isn’t what it was meant to be. After I stopped writing it, I did miss writing, mostly because I felt like all my rants were being lost on my poor family. Unfortunately, they can only listen to me yap about my Ingmar Bergman’s Persona is based on Balzac’s Seraphita conspiracy 10 times before my face becomes immensely punchable to them. I’d like to keep familial harmony in check so writing is serving multiple purposes.

And yes, one day I will write a whole dang essay on how CLEARLY Ingmar Bergman was influenced by Balzac’s story Seraphita when he made Persona (AKA the best Bergman film ever.)

Clearly, I just don’t write for other people. If people want to read what I’m writing, amazing! But if I’m just writing for me and my thought process that’s okay too. It’s my “creative outlet.” I don’t have sponsors for my work and I don’t charge a fee to read my stuff. I know a lot of people do, which is fine, but if I’m pushing for a reason to write it’s because I’m writing for people to learn. The areas I write about tend to be exceptionally gatekept and I don’t want to add to that problem. Or on the alternate should we gatekeep works to only those we have judged as able to understand or act on our work? Is all writing for all people? Now we are asking the deep questions…

Which brings me to the most gatekept writing I actually do. Most of what I’ve been writing recently has been law review articles. I’m literally being paid to research and write at the moment. Which…is great, but law review articles are a different kind of writing. It’s academic. It’s rote. At this point what gets accepted by good-to-great law reviews is essentially templatized. You know how to write what you need to write to communicate what needs to be communicated. There’s no personality. Definitely no cursing which puts an enormous strain on my personality showing up…but anyways, in law review articles there isn’t supposed to be a personality per se. You are meant to make an outstanding legal argument. You are writing to catch citations at a later date. It’s a little more personable than writing an amicus brief or a ruling in the judicial branch, but not by much. So, what I’m saying is that it’s a sterile, automatic-type of writing. I firmly believe that any good lawyer should be able to write a solid law review article. Now all that being said, the question follows…does that mean lawyers are now writers?

I looked and my law review articles have, on average, 30 versions that I worked through. (I won’t lie, my poor laptop storage situation is dire at the moment.) These are 50- to 100-page articles. Sooooo they’re not “light” reading by any stretch of the imagination. Nor were they a weekend or two of work. Think long ass months of arguing legal standards in your head and then typing them out, thinking that each word would make Thurgood Marshall reimagine his best arguements. I know that sounds super thrilling to everyone. These articles are longer than some books I have read. Is this length because I am a good writer or because I need a good editor…who’s to say!

When I was in law school I didn’t do law review because I had a full-time job and was already placed in the job I would have after graduation. I didn’t want to do hours of free work, nor did I want to spend my life in the law library chasing down errant bluebook citations. And well, times they do change, and the joke’s on me…because I spent 10 hours last week on bluebook citations. Though, now I am paid for it. At that time writing didn’t make sense for where I wanted my career to go and now it is the centerpiece. Would I go back in time and do law review? No. (Though honestly that is more because I also didn’t really like the people who were serving on law review. Also, not gonna lie…I was kinda a brat still in law school. So, if I was the mature version of myself…maybe I would have done law review.)

It’s so hard to leave that academic style of writing behind for me after I finish an article. Coming back to writing like a person after months of law review work is like trying to talk normally after a root canal — everything is numb and you sound ridiculous. I sit down to write something with a little life in it and out comes “this Article proceeds in four Parts.” My own voice gets lost somewhere in the muck and mire of legal word salad, buried under passive constructions and defensive hedging and eighty-seven footnotes whose only real job is to prove I know what I am talking about. (It’s truly the most pick me of pick me behaviors.) It takes weeks to find my actual voice again, and I am never entirely sure I got all of it back. To be fair there’s probably even a bit still here in this essay. I also just don’t have that many opportunities to just write to write. I’m not a diarist unless I think it’s important for others to hear (for better or worse). When I write I’m hoping people learn something…anything. (Please for the love of G-d learn something.) Therefore, I’m usually trying to infuse some of my own -isms and humor into pieces, but at the same time I know that sarcasm and deadpan aren’t exactly easy to see in words no matter how many italics I use. Just pretend you notice them all and give me a lil’ chuckle here and there.

And yet I’d defend the academic style of law review writing to anybody who asked. Law is a profession everyone assumes is stuck in the mud, and mostly they’re right. But the mud is the point. Law moves through its written and oral traditions or it doesn’t move at all. Somebody writes something down. Somebody else reads it and disagrees in writing. A judge reads that and cites it in a dissent. Twenty years later the dissent is a majority and the law is different than it was. It’s a chain of people accepting each other’s offers across decades, in public, on the record. An article I publish this year or next might sit untouched until 2041 and then show up in a footnote that changes something real. Slow. Unglamorous. Still almost entirely personality-free. And one of the only kinds of writing where you can actually watch that happen with a citation attached. The sterility has a purpose. It just costs me my voice for a few months at a time.

That’s one end of the spectrum. Writing so slow and so traceable that you can watch a single sentence get picked up across forty years of footnotes. Now let me go to the other end, where none of that machinery exists and the same thing happens in about four seconds.

AI has changed so much for creative fields, writing included, maybe even to the greatest extent. Is AI a writer? Can something trained on a human writing style be a writer? Do you need a soul to write? Is writing a prompt an academic or creative feat?

One of my cousins was paid a pretty penny by an AI lab so that they could train their models on his writing. By all measurements he is a great writer. (Alas, I don’t believe that writing is determined by genetics.)

That transaction matters more than people give it credit for. He made an offer, a lab accepted it, and — this is the part everyone skips — they paid him for it. That’s not charity. That’s not goodwill or a nice gesture. That’s consideration. It’s what you owe somebody whose work you’re building on. I spend my working life somewhere near this question, so let me just say what I think responsible AI development looks like on the writing side: train on good writing, and pay the people who wrote it. Both halves. Not one. Training on good writing and not paying for it isn’t a deal, it’s a taking. And paying a fortune for slop doesn’t help anybody either.

And the quality half isn’t sentimental, it’s infrastructural. A model trained on careful, well-edited, structurally sound prose produces careful, well-edited, structurally sound prose. Hundreds of millions of people read that output every single day. For a lot of them it might be the most competent writing they’ll see all week. So it’s either a literacy intervention at a scale no school district could dream of, or it’s a slop machine pointed at people who are already struggling to read. Depends what went in. Depends who got paid. But most important for those great writers is that someone got paid.

I’m glad AI was trained on good writing. If it wasn’t the bad AI slop would be all we are left with. The downward literacy spiral would be even worse because nobody would want to read that crap.

Honestly, with the literacy rate we currently have (especially here in the United States) does it matter who a writer is? If humans aren’t even reading whatever is being put out into the universe whether by AI or actual human, does it matter who’s writing it?

And whether you like it or not, the literacy numbers are genuinely bad. In 2023, 28% of American adults — roughly 59 million people — scored at or below Level 1 in literacy, which means a short simple text is fine but multi-step instructions start coming apart in your hands. In 2017 it was 19%. On the 2024 Nation’s Report Card, about 40% of fourth graders came in below NAEP Basic in reading, the largest share since 2002. Not below proficient. Below basic. And here’s the part I actually think matters more than we know: even the people who can read have stopped. Daily reading for pleasure in this country fell by more than 40% between 2003 and 2023, with the steepest drops among people with less income, less education, and rural addresses.

Reading, supposedly the most democratic thing a person can do, is quietly turning back into a class marker. Gutenberg is rolling in his grave.

Without literacy there are no writers. AI isn’t killing the writer — illiteracy is. Lack of reading and reading comprehension is.

Why is this happening? The usual suspects are all in the room. Phones. The pandemic. The collapse of sustained attention. Reading curricula that taught a generation to guess at words from the pictures. The plain fact that most people are exhausted and a screen asks less of you than a paragraph does. But notice what isn’t on that list. Not one of them is something a writer did. Reading didn’t stop because the writing got worse. That distinction matters in a lil’ bit.

And what if the new “reader” is an AI agent? How do you write for a machine? Will it understand the emotional pull of heartbreak? Will it understand the childlike joy of morning dew? Will it understand how those are put into words? How will we function as a society if the only readers left in the world are the AI agents we have trained? And who will be those training them if reading is a lost art?

Here’s the version of the future that actually scares me, and it isn’t the one where AI writes everything. AI can already write everything. That’s not a future, that’s a Tuesday. The one that scares me is where AI writes everything and there’s nobody left qualified to notice whether any of it is good.

That’s what illiteracy buys you. Not a world without words — a world drowning in them, where almost nobody can tell the difference between a sentence that’s true and a sentence that’s just shaped like the truth. Right now the models are good because they were trained on people who were good, and enough of us can still tell. Thin out the reading population and they start getting trained on their own exhaust, graded by people who can’t grade, for readers who skim a summary of a summary. Writing doesn’t die in that world. It just stops meaning anything, because there’s nobody on the other end to accept it. An offer nobody can read isn’t an offer. It’s noise with excellent grammar.

Which then raises the obvious question, and I may as well ask it now that the numbers are sitting right there: why write at all, if nobody’s reading? I told you much earlier that I kinda write into a void. I meant it as a small joke about my own modest readership. Having just typed out what’s happened to reading in this country, I’m significantly less sure it was a joke. If the title is conferred by readers, and the readers are genuinely thinning out, what exactly am I doing at eleven o’clock on a Tuesday?

The cynical answer, and the right one for an enormous amount of what gets published these days, is clout. Not readership. Clout. A staggering volume of writing right now isn’t aimed at a reader at all. It’s a signal to a market. Thought leadership. Personal brand. Building in public. (JFC I hate those words with a burning hot passion and writing them here is burning my eyes.) The sometimes ethereal post exists so people can see that you posted it, and whether anybody reads it is almost beside the point. Yes, I know me saying all this is a bit ridiculous. I have a Substack. I’m about to put this out into my own lil’ void on Medium. Nobody writes five thousand words about whether they’re a writer without some part of them hoping to be told that they are one. (Must keep this in mind for future therapy sessions…)

What makes it worse is that the engagement economy is really good at manufacturing something that looks exactly like acceptance and isn’t. A like isn’t a reading. A share isn’t a reading. Fourteen thousand impressions are definitely not fourteen thousand people who made it to the end. That’s somebody nodding at you from across the room without ever opening the document, and any lawyer who’s been on the other side of a signature from that person can tell you exactly what it’s worth. (Zilch in case you were wondering.) I think that’s most of why everyone feels like a writer and nobody quite is.

There’s a meaner way to put all of this, and some days I sadly suspect it’s the true one. An enormous amount of writing right now isn’t written to be understood. It’s written to make the writer look smarter than whoever they’re picturing on the other end. The tells are everywhere once you start nitpicking. The word that could’ve been the simpler word. The throat-clearing about credentials before the actual point. (However, I do forgive women for this because women live in a bit of a “well actually” world on this side of the gender spectrum. No credentials, no listen.) The argument is built so that disagreeing with it makes you look slow. Again, I’ll stop short of saying everyone, because the second I say everyone I’ve indicted the woman writing in her second language and the man dictating a novel he can’t see, and they’re plainly not doing this. But the dominant register of public writing has drifted away from explaining something and toward displaying the person doing the explaining.

Which leaves a real chicken-and-egg problem, and I’ve gone back and forth on it a lot because I get bored sometimes and let my mind wander into weird places. Who lost the plot first? The writers, who started writing toward their own reflection instead of toward a reader? Or the readers, who got so tired of being lectured at by people performing intelligence that they just stopped showing up? People reach for chicken-and-egg as a polite way of declining to answer, which has always irritated me a little, because the literal chicken-and-egg question does have an answer and it isn’t remotely close. The egg wins by about a hundred million years. Things were laying eggs long before anything resembling a chicken turned up to sit on one. (And yes, I know: a chicken egg is an egg laid by a chicken. Take that one up with the dinosaurs.) So, I’m not going to hide behind the phrase.

I’ve landed, tentatively, on the readers…and not because I blame them for a second. That list I ran through earlier is the reason, and not one item on it was anybody’s choice. Readers left, and writers did what anybody does when the audience walks out. They optimized for whatever still got noticed. And what survives a skim isn’t an argument. Whomp whomp. An argument makes you read the whole thing. A status signal survives the first eight words. As readers we got more status signal, which drove off more readers, which rewarded more status signal. The ego writing is downstream of the not-reading, and then it makes the not-reading worse. The order matters though, because it means you can’t fix this by shaming writers into humility. The fix, if there is one, is on the reading end.

And is it fulfilling? Writing to stroke your own ego? I don’t think it can be, structurally, and that’s the part I actually find sad. If you write to be understood, then being understood by one person is the whole thing. You can feel it land. If you write to be thought clever, nothing ever finishes, because you didn’t want to be read. You wanted to be ranked. And a ranking is never finished. There’s always somebody above you and a fresh post in the morning. Ego writing might be the only kind nobody can ever finish, because the writer was never really talking to anyone. They’re holding up a mirror and hoping you’ll go stand behind it.

So why do I keep doing it…Entering a combination of words into the void I mean. Partly because writing is how I work out what I think, which I said way back near the beginning and actually meant. It’s worth writing even when the room is empty, because drafting it clarifies me. And partly because the alternative to writing for nobody is not writing at all, and I’m not willing to be one of the people who stopped. So I keep writing. For better or worse depending on who you talk to.

I will admit that the advent of AI has changed me as a writer. Due to my actual day job I have a lot of insights into model training for almost every major lab and it does make me do a double take with what I write. For example, RIP the em dash. Em and en dashes were my favorite. [Insert sadface emoji lol.] But now I find myself trying to delete them from my repertoire even when they serve the words best. After the first draft of this essay I went back and changed my edits to put them all back in. I will not cave to the anti-em dash cabal.

I’m going to let you into a bit of a secret. I didn’t learn grammar in school. Unfortunately, at that time West Virginia public schools didn’t teach it within the English curriculum. I learned proper grammar when I was in law school and in the following years. Which is why em and en dashes, solid sentence structures, and Oxford commas mean something to me. I’m sad that we are losing key parts of good writing because people are too scared to have their writing be labeled as “AI writing.” That’s not good for anyone. That means we are losing something in our culture thats essential. And that something is good quality writing.

I have read very good AI writing output. I have read bad AI writing output. But the same can be true for human writers. Sometimes things written by humans can essentially be unreadable because they just…suck. I’m sorry but there isn’t a nice way to say that our educational system has created a lot of people who don’t know how to effectively use the written word, yet they feel as though our eyes should be put upon their blah blah. (If you feel this way about my writing, I truly am sorry but I promise you really don’t have to read my stuff. We’re 100% good even if you don’t.)

With that being said, I would rather read good AI than bad human writing. I think we all agree on that.

However, now I have to get on a soapbox. What about all the times female writers have been accused of actually using the work of men? For years men have loved to accuse exes who gain fame in writing of “stealing their work.” The most common accusation against female writers really does seem to be that their work is actually that of their romantic partners. I fear that “AI” will become the new “partner” accusation used to continue to put them “in their place” outside of the creative or academic discussion. Gotcha journalism will further become a place of sexism and misogyny more so than we’ve already let it.

What about AI use for those who are disabled? If someone disabled uses AI in a manner to write when they cannot see does that mean they cannot write the next great American novel? Or those who are writing outside of their native language and being accused of AI writing? Or translations? I don’t know where the line is but I know that these people are writers. How do I know these people are writers? Because I want to read them. Which, if you have been keeping track, is the only test I have offered this entire essay, and it does not care one bit how the words got onto the page.

So are we using the AI revolution to gatekeep “writing” further, but actually eating our own face in doing so? Are we going to make writing something that no longer exists like a dead, ancient language? Professional writing is already something that we only deign a select few worthy of…I admittedly have some of the same beliefs that I struggle to explain as you have seen just in this essay. Could AI be used to make writing more democratic? Could it be used to help more people communicate to others better?

So here is where I have landed, and it took the whole essay to get here.

Everyone is a writer now. That half is easy and it’s already true. More words are produced by more people in an average hour than at any point in human history. Slack messages, captions, texts, reviews, memos, posts, prompts, this. The tools are free, the distribution is free, the gate is off its hinges. Writing has never been more democratic than it is this minute. Most likely every person reading this has already written something today.

And nobody is a writer, because a writer isn’t a person who writes. A writer is a person somebody reads. The title was never yours to hand yourself. I said so near the top of this essay, before I understood I was making an argument instead of being modest. It’s conferred, it’s conferred by readers, and readers are the one resource we’re actually running out of. You can write into an empty room all day long. It doesn’t make you a writer.

And we’ve been so loudly occupied arguing about who should be allowed to supply — is it the AI, is it the person who used the AI, is it the woman whose husband must have helped her, is it the blind writer, is it the guy with the Substack, is it me — that almost nobody is asking the question that decides all of it. Who’s left to receive any of this.

There’s a harder version of that question, and it’s the one that actually matters. What makes my voice more important than anyone else’s? The honest answer is nothing. Nothing does. I’ve got no credential that settles it, no license, no anointing. I’ve spent this entire essay insisting I’m not a writer and then written five thousand words on the subject, which is either genuine humility or the most elaborate humblebrag ever committed to a Word document, and I’m honestly not sure which.

But that question in and of itself assumes a lot. It assumes importance is something you assert about yourself, in advance, and under everything I’ve argued here it just isn’t. I don’t get to grade my own work. Nobody does. Every writer who declared their own voice essential was making a claim they had no standing to make, and every writer who turned out to matter was made by somebody else choosing to read them. Usually late. Frequently after they were dead. Almost never on the schedule they’d have picked. You don’t write because your voice is important. You write to find out whether it is. And even then you don’t get a vote.

I don’t think the answer is gatekeeping harder. Every single time we’ve drawn a line around who counts as a real writer, the line has looked humiliating in hindsight, and this round won’t be the exception.

The true answer in the end is the boring one, and it’s the only one I’ve got: be a reader. Read something difficult on purpose. Read a whole thing instead of the summary of the thing. Read something a person wrote and then tell them you read it, because telling them isn’t a courtesy, it’s the actual mechanism by which writers get made, and it’s the only item on this list no metric can fake. Teach a kid to read for pleasure and you’ve created a writer twenty years early. You just don’t get to know which one yet.

And if it turns out somebody out there wanted to read this, then I guess you’ve gone and made me a writer. Which wasn’t the point. But I’ll take it.

Stevie Cline, Esq. is a legal practitioner in the Artifical Intelligence governance space. She can be found at www.steviecline.com. The views expressed here are her own and not that of her employer. I mean…maybe they’re the same views. Who knows?

Introduction

Three facts about the year 2025 frame the problem. First, venture investors allocated $258.7 billion to artificial intelligence companies, 61 percent of the $427.1 billion invested globally across all sectors, roughly double AI’s 30 percent share in 2022 (OECD 2026). Second, the large cloud providers committed an estimated $300 billion or more to AI capital expenditure in the same year, with larger figures announced for 2026 (Crunchbase 2025). Third, total enterprise spending on generative AI in 2025, counting both user-facing products and infrastructure, reached approximately $37 billion (Menlo Ventures 2025). More than half a trillion dollars of annual capital formation was directed at a contemporaneous revenue pool roughly one-fifteenth its size.

The ratio is not, by itself, evidence of misallocation. Capital formation in a general purpose technology properly precedes the revenue it enables, and the revenue in question tripled year over year. The economic question is narrower and harder: whether the investors supplying the capital, at the prices they are paying and in the market structure they are paying them in, are positioned to capture the surplus their capital helps create. On that question the theoretical literature is unusually consistent, and the answer it gives is largely no.

A clarification is required before proceeding, because this essay’s exhibits are drawn from the sector’s leading firms and could be misread as a judgment upon them. The argument concerns prices rather than companies, and the two are separable by construction: a firm of the highest quality can be a poor investment at one price and an outstanding one at another, and nothing that follows implies that the frontier laboratories are anything other than exceptional enterprises, plausibly the defining enterprises of their era. Their conduct, moreover, is the correct conduct given the game they are in. A firm financing a competitively mandatory escalation has an obligation to sell equity at the highest price the market will set, and in a common-value auction that price is the most optimistic bidder’s; the issuer harvests the winner’s curse rather than suffering it. The analysis that follows is accordingly an analysis of the buy side: of the auction, the intermediation chain, and the structure of the claims, which is where the economics locates both the losses and their remedy. The same clarification extends to the investors the essay describes. Every mechanism examined here operates on intelligent, well-intentioned participants responding correctly to the contracts and information in front of them, which is why the mechanisms persist. The purpose of naming them is that each one has a remedy, and the second half of the essay exists to supply the remedies.

The Allocation Facts

The scale of the reallocation is without precedent in the venture data. The OECD’s analysis of global flows found not only that AI absorbed 61 percent of world venture investment in 2025 but that the capital was concentrated within the sector to a remarkable degree: deals of $100 million or more accounted for roughly 73 percent of AI deal value, transactions above $1 billion represented nearly half of the sector’s total, and the five largest deals of the year absorbed about $63 billion, a quarter of everything invested in AI (OECD 2026). Crunchbase’s parallel accounting shows five companies raising $84 billion in 2025, about 20 percent of all venture funding worldwide across every sector; two of them, OpenAI and Anthropic, took 14 percent of the global total by themselves (Crunchbase 2026a).

The concentration then intensified. Global venture funding in the first quarter of 2026 reached roughly $300 billion, approaching 70 percent of the full-year 2025 total in a single quarter. Of this, $242 billion, or 80 percent, went to AI, and four companies (OpenAI, Anthropic, xAI, and Waymo) absorbed $188 billion, which is 65 percent of the world’s quarterly venture output (Crunchbase 2026b).

One further statistic deserves emphasis because it will matter in Section V. While the mean AI deal size rose from $11.2 million in 2014 to $35.8 million in 2025, the median deal in 2025 remained approximately $5 million (OECD 2026). The market has bifurcated into two regimes with almost nothing in common: a broad population of small experiments, and a narrow set of system-scale positions that absorb the majority of the capital. The theoretical apparatus appropriate to each regime differs, and much of what follows concerns the consequences of applying the first regime’s logic to the second regime’s transactions.

Price Formation: Auctions Without a Short Side

The winner’s curse

Begin with the transaction-level mechanics. Capen, Clapp, and Campbell (1971), three petroleum engineers writing in the Journal of Petroleum Technology, asked why oil companies that won competitive auctions for offshore drilling tracts systematically earned poor returns on the tracts they won. Their answer, later formalized by Wilson (1969) and by Milgrom and Weber (1982), is the winner’s curse. In a common-value auction, where the asset is worth approximately the same amount to every bidder but each bidder observes only a noisy private estimate of that value, the mechanism allocates the asset to the bidder whose estimate contains the largest positive error. Winning is therefore adverse information about one’s own estimate, and the rational bidder shades her bid below her unconditional valuation, with the required shading increasing in the number of competitors. The experimental literature (Kagel and Levin 1986; see Thaler 1988 for a survey) established how rarely bidders shade sufficiently, even when the theory has been explained to them.

Early-stage venture investing has historically enjoyed partial immunity because seed-stage deals approximate private-value auctions: an investor with distinctive networks or operating capability can genuinely make a company worth more in her hands than in a rival’s, so outbidding the rival is not automatically an error. The defining deals of the present cycle do not have this character. At the foundation-model layer, the marginal co-investor in a forty-party syndicate supplies capital and nothing else; the asset’s value is essentially common across owners. The bidders read the same benchmark results, the same disclosed run-rates, and the same scaling literature, so dispersion in bids measures dispersion in error rather than dispersion in prospective value creation. The rounds clear in open ascending processes that terminate, by construction, at the reservation price of the most optimistic capital in the market.

Miller, Harrison and Kreps, and the absence of a short side

The winner’s curse describes a single transaction. A second body of theory explains why the resulting prices are not corrected between transactions, and it deserves more attention in this context than it receives. Miller (1977) showed that when investors hold heterogeneous beliefs and short sales are constrained, equilibrium prices reflect the valuation of the optimists only, because pessimists cannot express their views except by abstaining. Rather than aggregating the distribution of opinion, the price samples its upper tail. Private markets are the limiting case of Miller’s model. There is no mechanism whatsoever by which an investor who believes a private AI company is overvalued can trade on that belief. Every transaction is initiated by an optimist, priced by optimists, and validated by the participation of further optimists, while the entire informational content of the skeptics is discarded.

Harrison and Kreps (1978) extended the logic in a direction that bears directly on late-stage AI pricing. With heterogeneous beliefs and short-sale constraints, an asset’s price can exceed the valuation of even the most optimistic investor’s estimate of fundamentals, because ownership carries a resale option: the right to sell later to whoever is then most optimistic. The price decomposes into a fundamental component and a speculative component equal to the option’s value. Scheinkman and Xiong (2003) showed that the speculative component grows with disagreement and trading intensity. Keynes (1936, ch. 12) described the same phenomenon informally as a beauty contest in which participants price their forecast of average opinion rather than the asset itself; Harrison and Kreps supplied the formal statement. Shleifer and Vishny (1997) explained why arbitrage does not repair any of this: correction requires capital that can survive interim losses, and in private markets such capital cannot even take a position.

The 2025–2026 repricing of the frontier laboratories is what this machinery predicts. Anthropic’s valuation moved from $61.5 billion in March 2025 to $183 billion in September 2025, $380 billion in February 2026, and $965 billion in May 2026, an approximately fifteenfold repricing of one firm in fourteen months (CNBC 2026; Forbes 2026a). OpenAI moved from $157 billion in late 2024 to $852 billion by March 2026 (Sacra 2026; Crunchbase 2026b). Both are businesses of extraordinary quality growing at rates without precedent in enterprise software; Anthropic’s disclosed run-rate rose from roughly $1 billion to $47 billion in about eighteen months, and OpenAI’s from $6 billion to $24 billion (SaaStr 2026; CNBC 2026). The pricing question is separable from the quality question. An entry at $965 billion against a $47 billion run-rate is a payment of roughly 21 times an unaudited annualized revenue figure, set in an open auction among bidders with identical public information, in a market where disagreement cannot be priced. Forbes reported that several of the company’s earliest investors declined the final private rounds and elected to wait for the public listing, which bankers expected to price materially lower (Forbes 2026b). When the holders with the longest information advantage in an asset decline to transact at the prevailing price and the marginal buyer is the newest entrant, the Harrison–Kreps decomposition suggests the speculative component of the price has become large relative to the fundamental component. The observation concerns the structure of the auction alone.

The adding-up constraint

Transaction-level overpricing aggregates into a market-level inconsistency, and the inconsistency is best stated in the language of probability rather than of estimates. Each frontier-laboratory valuation is defended by some version of the proposition that the firm will command the dominant share of the eventual AI profit pool. Dominance is a mutually exclusive event: at most one firm can hold it. Prices therefore embed implied probabilities, and coherent pricing requires the implied probabilities of mutually exclusive outcomes to sum to no more than one. By mid-2026, private markets were pricing three to five laboratories as though each carried a substantial probability of dominance, while also pricing the cloud providers, the semiconductor complex, and portions of the application layer as though the surplus would pool there instead; the aggregate paper value of the private technology complex approached $7.5 trillion on the Crunchbase unicorn board (Crunchbase 2026a). There is one escape from the inconsistency: if the present value of dominance is large enough, each implied probability shrinks and the sum falls below one. But the laboratories’ claims alone exceeded $2 trillion before counting the complements’ claims on the same pool, so the escape requires an appropriable profit pool far outside any historical relation between innovation and producer capture (Nordhaus 2004; Section IV.3). The inconsistency can be assumed away only by assuming the pool, and the assumption is the thing being priced. In a market with a short side, implied probabilities summing above one constitute a Dutch book, and arbitrage would close it; in a market without one, they are the resting state (Miller 1977), and the correction arrives through the repricing that Sections VI and VII describe rather than through trading. Some large subset of currently marked positions must be revised downward as a matter of coherence, independent of any forecast about the technology.

The Race: Rent Dissipation and Endogenous Sunk Costs

Innovation races

The competition among frontier laboratories is usefully modeled as an innovation race, a literature developed by Loury (1979), Lee and Wilde (1980), and Dasgupta and Stiglitz (1980). The canonical results are two. First, when the prize accrues disproportionately to the leader, firms overinvest relative to the social optimum, since each firm’s expenditure imposes a negative externality on rivals by reducing their probability of winning. Second, and more important for investors, free entry into the race dissipates the rent: in the limiting cases, aggregate expenditure on the contest approaches the value of the prize itself. The structure resembles an all-pay auction (Baye, Kovenock, and de Vries 1996), in which every participant pays its bid regardless of outcome. Training runs are bids in an all-pay contest. Every laboratory pays for its compute whether or not its model wins the benchmark quarter, and the contest is re-run continuously.

Hirshleifer (1971) supplies a complementary and underused result: the private value of information can exceed its social value, because information confers distributional gains (the ability to trade ahead of others) in addition to any productive gains, and competition to acquire it therefore attracts resources beyond what its social product justifies. A frontier capability lead is, among other things, foreknowledge of what all capability will soon cost. The race to hold that lead for a quarter or two attracts expenditure appropriate to its private, positional value rather than to its durable economic value, and the difference is dissipation.

Sutton’s escalation mechanism

Sutton (1991) provides the most precise available description of the foundation-model industry’s structure, in a framework developed for an entirely different question. In markets where firms compete by making endogenous sunk-cost investments in quality (Sutton studied advertising and R&D), an escalation mechanism operates: as the market grows, the returns to outspending rivals on the sunk investment grow with it, so equilibrium sunk outlays escalate with market size and concentration remains bounded away from zero no matter how large the market becomes. The industry converges to a natural oligopoly in which a small number of firms survive, while the escalation mechanism converts a substantial share of the industry’s prospective rents into the sunk investments themselves. The oligopoly is real, and much of the profit is consumed by the price of remaining in it.

Frontier AI is an endogenous sunk-cost industry in nearly pure form. Training expenditure is sunk, quality-increasing, and competitively mandatory, and it has escalated at a pace Sutton’s framework would recognize: OpenAI exited 2025 with infrastructure commitments reported near $1.4 trillion before revision to approximately $600 billion through 2030, against roughly $20 billion of revenue, with cash burn in the tens of billions and no positive free cash flow projected before the end of the decade (Forbes 2026c; Sacra 2026). The framework’s prediction is that a few of these firms survive into a stable oligopoly whose members have paid out the capitalized value of their future rents, and then some, in the escalation required to arrive there; survival and shareholder return are separate questions. The escalation is competitively mandatory, and a laboratory that declined it would exit the frontier, so any mispricing attaches to the equity claim rather than to the corporate strategy: equity purchased at prices that assume the rents survive the escalation is mispriced even if the firm wins.

The Nordhaus benchmark

How much of the surplus from innovation do innovators historically keep? Nordhaus (2004), examining the American nonfarm business economy from 1948 to 2001, estimated that producers captured approximately 2.2 percent of the total social surplus generated by technological advance, with the remainder passing to consumers through competition and falling prices. The estimate is rough and the variance across industries is large, but it establishes the base rate against which AI capitalizations should be judged. Whatever present value one assigns to AI’s aggregate surplus, current frontier-laboratory pricing requires innovator capture at a large multiple of the historical rate, at the same time that Section IX will show the technology’s appropriability regime to be unusually weak, not unusually strong. Section X.5 will return to the Nordhaus figure from the other side, since a result about where 98 percent of the surplus goes is also a map of where to look for it.

The Power Law Does Not Apply Here

The standard justification for participating in AI rounds at prevailing prices is an appeal to the power law: venture outcomes are so heavy-tailed that missing the single defining company of an era is costlier than overpaying for it, so access dominates price. The premise about outcomes is well supported. The inference about investing is not, and the distinction between the two is the central analytical error of the cycle.

What the power law is, and what generates it

That returns to innovation are heavy-tailed is among the better-documented regularities in the economics of technical change. Scherer and Harhoff (2000), examining patent values, invention royalties, and venture portfolios, found distributions skewed to the point where the top decile of outcomes accounts for half or more of total value. Cochrane (2005) and Korteweg and Sorensen (2010) established the corresponding facts for venture-backed equity, correcting for the severe selection bias in observed exits. Gabaix (2009) surveys the generating mechanisms: proportional (Gibrat-type) growth with entry and frictions, and preferential attachment in markets with increasing returns. Mallaby (2022) supplies the canonical practitioner account of the same regularity and of the investing culture built on it. Nothing in this essay disputes any of it. Firm outcomes in AI will very likely be as skewed as firm outcomes in every prior general purpose technology, and plausibly more so.

But a power law in company outcomes delivers a power law in investor returns only under conditions on the claims investors purchase rather than on the companies themselves. Four such conditions can be stated. First, entry prices must be small relative to the support of outcomes, so that the position is convex: an option purchased near the bottom of the distribution, with capped loss and effectively uncapped multiple. Second, draws must be numerous and cheap, so a portfolio can hold enough of them for the tail to be sampled; Kerr, Nanda, and Rhodes-Kropf (2014) describe venture capital as precisely this technology of cheap experimentation, and Ewens, Nanda, and Rhodes-Kropf (2018) show that the modern venture model was itself reorganized around falling experiment costs, with staged financing operating as a sequence of abandonment options (Gompers 1995). Third, draws must be approximately independent, so that portfolio arithmetic aggregates them. Fourth, the upside must be unbounded relative to position size, since the fund-level mathematics of “one winner returns the portfolio” requires multiples in the hundreds.

Each condition fails at the frontier

Now examine the segment absorbing the capital, and note that every condition fails, most of them by arithmetic rather than by judgment.

The tail is in the price. Investor return is outcome minus entry price, and a heavy-tailed outcome purchased at a price that already capitalizes the tail produces a return distribution that is not heavy-tailed at all. An entry at roughly 21 times an unaudited run-rate, at a valuation approaching one trillion dollars, prices the extreme branch of the scenario tree as the central case. Conditional on that entry, the distribution of investor returns is compressed on the right while remaining fully exposed to the loss of principal on the left; the payoff profile resembles a written put, in which the holder collects a modest premium if the favorable scenario materializes and absorbs a near-total loss if it does not, rather than the purchased call that venture mathematics presumes. The convexity that makes the power law profitable to its holder has been sold to the previous owners.

The upside is bounded by aggregate wealth. The canonical venture outcome is a return of fifty to one hundred times. From an entry of $852 billion or $965 billion, such a multiple implies terminal enterprise values between $40 trillion and $95 trillion, figures of the same order as the market capitalization of all listed United States equities, which they meet or exceed at the upper end. Power laws describe scale-invariant processes, and scale invariance necessarily breaks when a single position becomes a nontrivial fraction of the system in which it is embedded. Gabaix (2011) formalized the general point in his work on granularity: when individual units are large relative to the aggregate, their idiosyncratic shocks are aggregate shocks, and the statistical behavior of the small-unit regime no longer applies. The frontier laboratories are granular in exactly this sense, and the “asset class” built on them inherits their granularity: four firms absorbed 65 percent of world venture flow in a single quarter (Crunchbase 2026b). A distribution cannot supply hundredfold multiples from entries that are already system-scale.

The draws are neither numerous, nor cheap, nor independent. The experimentation logic of Kerr, Nanda, and Rhodes-Kropf requires many small, stageable, abandonable positions. A frontier position is one enormous, unstageable commitment: the option structure of staged financing collapses when the underlying asset requires a decade of compute contracts to remain competitive, since abandonment forfeits everything and continuation is priced by the escalation mechanism of Section IV.2. And the positions co-move. The co-movement is a structural inference rather than a measured covariance, since private marks do not trade, but the shared exposures are contractual and observable: the major AI holdings share the same factor exposures (the price of compute, the persistence of scaling returns, the regulatory treatment of frontier models, the cost of energy) and are, through the circular purchase-and-investment contracts described in Section VII, one another’s counterparties and customers. A portfolio of such positions amounts to a single macro bet written several times at increasing prices, whatever the number of line items on the schedule of investments suggests.

The race consumes the tail. Section IV’s results operate here with full force: even the eventual winner of an all-pay contest with endogenous sunk costs has paid out much of the prize in the contest, so the tail outcome, where it occurs, arrives net of the dissipation.

The category error and its consequence

The power law is regime-specific rather than dead, and the market’s own structure displays the two regimes plainly: a median AI deal of $5 million, where the preconditions of cheap, convex, numerous, quasi-independent experimentation still hold, and a mega-round segment measured in the tens of billions, where none of them do (OECD 2026). The power-law playbook, including its signature heuristic that access matters more than price, is coherent only in the first regime. Applied in the second, it functions as a license for unlimited overpayment, because it treats price discipline as a small-minded objection to an unbounded upside that the arithmetic of aggregate wealth has already foreclosed. Frontier positions are growth-infrastructure claims and should be underwritten as such: with entry multiples, cash-flow paths, dilution schedules, and covenant attention, the apparatus of an investor who expects to be paid from earnings rather than from the tail. The single most expensive sentence in the 2024–2026 vintage will turn out to be “this is a power-law business.” The sentence is true in one regime and costly in the other, and the discipline of asking which regime a given transaction belongs to is available to every investor at no cost at all.

Technological Revolutions and the Timing of Returns

Installation and deployment

Perez (2002) examined five technological surges (canals, railways, steel and electrification, mass production, and information technology) and identified a recurring two-phase structure. In the installation phase, financial capital funds the new infrastructure at a pace the real economy cannot yet absorb; asset prices decouple from earning power; a crash follows; and only then does the deployment phase begin, in which production capital diffuses the technology through the economy and the durable returns are earned. Two features of the account are commonly misread. The frenzy is functional: Britain obtained a national rail network from the 1840s mania, and the mania’s shareholders, in the aggregate, never recovered their capital. The crash, in turn, is mechanical: it is the repricing that permits patient capital to fund deployment. The buildout of 2024–2026, with hyperscaler capital expenditure above $300 billion annually and laboratories committing hundreds of billions in forward compute purchases (Crunchbase 2025; Sacra 2026), is installation-phase capital formation of the classic type. The historical regularity is that society keeps the infrastructure and the installation phase’s financiers absorb the losses, while deployment-phase returns accrue to those who buy the repriced assets or build upon the overbuilt substrate. The fiber financed by telecommunications creditors in 1999 and consumed at post-bankruptcy prices by the platform companies of the following decade is the proximate precedent.

Rational pricing dynamics and the empirical record of run-ups

One need not attribute any of this to irrationality. Pástor and Veronesi (2006, 2009) showed that apparently bubble-like price paths arise in fully rational models during technological revolutions: uncertainty about a new technology’s productivity is initially idiosyncratic and largely diversifiable, which supports high valuations, and as the technology is adopted the uncertainty becomes systematic, discount rates rise, and prices of the innovating firms fall even when the technology succeeds. The model’s discomfiting implication for the present case is that the repricing of AI assets does not require any bad news about AI. Adoption itself converts the risk from idiosyncratic to systematic and lowers the prices of the firms that carry it.

The empirical record on industry-level run-ups points the same way. Greenwood, Shleifer, and You (2019), examining United States industry returns from 1926 to 2014 and international sector returns from 1985, found that a sharp industry price increase does not on average predict low subsequent returns, which rules out simple mean reversion as the mechanism, while it does predict a substantially elevated probability of a crash, and characteristics of the run-up, including its sharpness, volatility, issuance, and the degree to which it accelerates near its end, forecast the crash with meaningful reliability. The 2025–2026 AI financing data exhibit the acceleration signature in unusually clean form: on the OECD’s accounting, AI’s share of world venture funding rose from roughly one third in 2024 to 61 percent in 2025, and on Crunchbase’s separate accounting it reached 80 percent in the first quarter of 2026 (the two series define the sector somewhat differently, but each shows the same acceleration within its own definition), while the largest positions repriced by an order of magnitude within roughly a year (OECD 2026; Crunchbase 2026b; Forbes 2026a).

Shakeouts and the diffusion lag

Two further literatures locate the present moment. Gort and Klepper (1982), studying the life cycles of forty-six product industries, and Klepper (1996), formalizing the mechanism, documented the regularity that new industries pass through a phase of mass entry followed by a shakeout that eliminates the large majority of entrants even as industry output continues to grow. The American automobile industry supported hundreds of producers at its entry peak before consolidating to a handful; the pattern repeats across tires, televisions, penicillin, and disk drives. Industry success and firm survival are nearly independent questions. A venture portfolio assembled during the entry phase at entry-phase prices is a claim on the population that the shakeout is about to cull.

On the demand side, the general purpose technology literature explains why the revenue that would justify installation-phase prices arrives late. Bresnahan and Trajtenberg (1995) showed that a general purpose technology’s returns depend on complementary innovations that must be developed sector by sector; David (1990) documented the four-decade lag between the electric dynamo’s introduction and its measurable productivity effects, which required the redesign of factories around the new power source; Brynjolfsson, Rock, and Syverson (2021) formalized the resulting productivity J-curve, in which intangible complementary investment initially depresses measured returns before the technology’s contribution appears. The $37 billion enterprise revenue figure for 2025, read through this literature, indicates that the complementary investment phase has barely begun, an argument for the technology and against the prices being paid for early claims on it. Helpman and Trajtenberg (1998) added the sequencing result that matters for Section X: a general purpose technology diffuses sector by sector in ascending order of the cost of the complementary innovation each sector requires, which means the order of adoption is partially forecastable from observable sector characteristics.

Financing Structure: The Markup Chain

Minsky (1992) classified financing positions into three regimes: hedge finance, in which cash flows service obligations; speculative finance, in which cash flows cover interest but principal must be refinanced; and Ponzi finance, his technical term for positions whose cash flows cover neither, and whose viability therefore depends on refinancing or sale at rising prices. His financial instability hypothesis holds that stability is destabilizing: extended good times migrate the population of financing structures toward the Ponzi end, until the system as a whole requires continuously rising valuations, at which point any interruption in the markup chain produces a discontinuous repricing rather than a gentle one.

The taxonomy maps onto the AI complex without strain. A laboratory burning $17 billion annually against trillion-dollar-scale forward commitments, with positive free cash flow projected for 2029 or 2030 (SaaStr 2026; Sacra 2026), is a Ponzi unit in the technical sense, as is the application company raising at a markup every nine months while spending several multiples of revenue. Neither observation is pejorative: financing ahead of cash flow is the optimal corporate posture for a firm building durable position during an installation phase, and Minsky’s taxonomy describes the aggregate fragility of a system composed of such positions, not mismanagement within any of them. The venture model is built on exactly this posture, and it is sustainable when such units are a minority within a hedge-financed system. Minsky’s warning concerns composition, and the composition is now extreme: by the first quarter of 2026, 80 percent of world venture funding flowed to a sector whose flagship firms are, without exception, dependent on continued external refinancing (Crunchbase 2026b).

The system’s cash position, as distinct from its reported value, is visible in the distribution data. Venture funds of the 2021 vintage report total-value multiples of roughly 1.3 to 1.8 times paid-in capital while distributions run near 0.08 times: eight cents of cash per committed dollar, five years after formation (Value Add VC 2026). Carta’s fund-performance panel corroborates the pattern from the other direction: only about a quarter of 2021-vintage funds had made any distribution at all by late 2025 (Carta 2025). Across private funds generally, Bain calculated the 2024 distribution yield at roughly 11 percent of net asset value, against an average near 29 percent in 2014–2017 (Bain 2025). Capital calls have exceeded distributions every year since 2022, and venture fundraising fell to $76 billion in 2024, its lowest level since 2019, as allocators declined to recommit against marks they could not spend (PitchBook–NVCA data, 2025). The industry’s adaptations, including record secondary volume above $100 billion in a half year, general-partner-led continuation vehicles in which managers transact with themselves at self-assessed prices, and net-asset-value lending projected to triple by 2030 (Private Markets Profile 2025), are recognizable as Minskyan release valves: instruments for extracting cash from marks without exposing the marks to an open sale. Section XI.3 will argue that these release valves are, from the other side of the trade, the most theoretically well-founded buying opportunity the cycle will produce.

Two further features amplify the fragility. First, an interim venture mark is the residue of the last auction rather than a price; Section III showed the auction to clear at the most optimistic participant’s valuation, and the residue propagates through every portfolio holding the position. Second, the sector’s revenue base is itself partially self-referential. The defining transactions of the period pair multibillion-dollar compute purchase commitments from laboratories with multibillion-dollar equity investments from the same cloud providers, so that a portion of headline revenue appears on two income statements at once, and disclosed figures are annualized run-rates rather than audited results; one laboratory has publicly disputed a rival’s revenue accounting (Malik 2026; Forbes 2026b). Akerlof (1970) describes the consequence of quality opacity in general terms: when buyers cannot distinguish revenue qualities, everything trades at a pooled price, and the pooled price is currently calibrated to the most favorable interpretation. The interaction of Miller-type optimist pricing with Akerlof-type opacity is a market in which the least verifiable numbers receive the most generous multiples.

VIII. Intermediation: The Incentives of the Fund Structure

The persistence of the pattern among sophisticated investors requires an explanation, and the explanation is that participation is individually rational at every node of the chain.

Consider first the general partner. Metrick and Yasuda (2010), decomposing the economics of private fund managers, found that roughly two thirds of expected general-partner revenue derives from fixed fees on committed capital rather than from carried interest. The manager’s principal asset is assets under management. Carried interest, meanwhile, is a call option on the portfolio, and the value of a call option increases with the variance of the underlying. The rational fee-maximizing manager therefore prefers larger funds, faster deployment, and greater variance, which are the three quantities the AI cycle supplies most abundantly. The empirical venture literature documents the machine’s output: Gompers and Lerner (2000) showed that inflows into venture funds inflate the prices of the deals those funds transact, holding deal quality constant; Kaplan and Schoar (2005) established that fund returns are strongly inversely related to aggregate capital raised in the fund’s vintage year, and that the performance persistence of top firms attenuates as they scale their fund sizes. Nanda and Rhodes-Kropf (2013) added the completing observation that hot-market vintages fund more experimental firms, producing both more failures and, disproportionately, the eventual breakthrough companies, whose gains are then harvested largely at post-correction prices by later capital.

Consider next the individual investment professional. Scharfstein and Stein (1990) modeled managers evaluated on relative performance and showed that mimicking peers is an equilibrium strategy even for a manager whose private information counsels otherwise, because a shared error is uninformative about individual ability while a solitary error is fully attributed. Keynes made the same observation about the reputational asymmetry between conventional failure and unconventional success. For a venture partner in 2025, absence from AI was legible in every limited-partner meeting as dereliction, while losses taken alongside the industry’s most respected firms will be legible as market conditions. Bikhchandani, Hirshleifer, and Welch (1992) supplied the informational counterpart: when agents observe actions but not information, rational imitation produces cascades that rest on almost no aggregate evidence, and the syndicated mega-round is a cascade generator, in which each additional participant rationally infers diligence it cannot verify and its check adds price pressure while adding no information.

The chain closes with the limited partners, who, starved of distributions by the previous cycle, concentrated smaller commitments into the largest brand-name funds on reputational-safety grounds, thereby forcing deployment into precisely the $500-million-and-above rounds that constituted 58 percent of 2025 AI funding (Crunchbase 2025), during vintage years the Kaplan–Schoar result identifies as structurally adverse. No participant is behaving irrationally given the contracts and the information each faces. The bust, when it arrives, will have been manufactured by correct responses to existing incentives, which is why it cannot be averted by the intelligence of the participants and why the contracts themselves are the appropriate margin of reform (Section XI.5).

Appropriability: Where the Rents Can and Cannot Rest

Suppose an investor navigates all of the foregoing, pays a defensible price, and selects the eventual winner. A final question remains, and it is the one on which most AI investment analysis is weakest: can the winner retain the surplus? Schumpeter (1942) characterized innovation profit as a temporary rent that imitation competes away; every innovation investment is a race between the rent’s magnitude and its half-life. Arrow (1962) identified the structural reason the half-life tends to be short for information goods: information is costly to produce, nearly costless to reproduce, and difficult to sell without partial disclosure, so its returns are systematically hard to appropriate without robust exclusion mechanisms. Model capabilities are information in Arrow’s sense. The available exclusion mechanisms, principally secrecy, compute scale, and closed interfaces, are real but porous, and open-weight replication has proven capable of approximating frontier capabilities within months of their demonstration.

The measured half-life is the shortest on record for a major technology. Epoch AI (2025), analyzing price-performance frontiers across six benchmarks, found that the price to achieve a fixed capability level has declined at rates between 9 and 900 times per year depending on the milestone, with the median rate accelerating from roughly 50 times to roughly 200 times per year after January 2024. Andreessen Horowitz (2024), measuring from within the optimist camp, documented a thousandfold decline in the cost of GPT-3-class capability over three years, from $60 to $0.06 per million tokens, a rate exceeding the compute deflation of the personal-computer era and the bandwidth deflation of the internet era. A complication cuts the other way and must be recorded: recent measurement work finds that the total cost of running state-of-the-art performance has simultaneously risen at roughly eighteenfold per year, because marginal gains at the frontier now require disproportionate inference (The Price of Progress, arXiv:2511.23455, 2026). A premium tier therefore persists at the very top. But a rent that must be repurchased every research cycle through escalating sunk expenditure is Sutton’s treadmill rather than a moat, and it should be valued as a flow of contested profits rather than capitalized as terminal dominance.

Teece (1986) supplies the framework that converts this diagnosis into a prediction about where the surplus settles. Examining why innovators so frequently fail to profit from their own innovations, Teece showed that when the appropriability regime is weak and commercialization requires complementary assets, such as distribution, manufacturing capacity, customer relationships, or proprietary data, the profits accrue to the holders of the complementary assets rather than to the source of the innovation. His canonical case is instructive: EMI developed the computed tomography scanner and exited medical imaging within a decade, while General Electric, holding the hospital distribution and service relationships, took the market. In artificial intelligence the appropriability regime for raw capability is weak and measurably weakening, and the demand-side evidence already conforms to Teece’s prediction: the enterprise revenue that exists is concentrating where proprietary data, workflow integration, and institutional trust reside, with enterprise customers generating three to five times the revenue per token of consumer users and exhibiting contractual stickiness that consumer usage lacks (Forbes 2026c). Nordhaus’s 2.2 percent base rate for innovator surplus capture (Section IV.3) is the historical average across appropriability regimes; the present regime sits below the average, while the prices being paid require it to sit far above.

This completes the diagnosis, which can be stated in one sentence: optimist-priced auctions are financing an all-pay race whose rents will be dissipated by escalation and competed into consumer surplus, during the installation phase of a technological surge, on a Ponzi-structured markup chain, under intermediation contracts that reward continuation. The remainder of the essay takes the same theoretical apparatus and runs it forward: if this is where the money is destroyed, where, specifically, is it made?

A Positive Theory of AI Rents: Where the Money Settles

The constructive question has a rigorous form. When the price of a productive input collapses toward zero, general equilibrium relocates the surplus rather than destroying it. The task is to identify the relocation rules, and economics supplies five of them. Each subsection states the theory, derives the investment implication, and identifies the diligence question that separates genuine instances from imitations. Together they constitute what I will call the deployment portfolio.

Ricardian rents: the surplus migrates to the inelastic factor

The oldest result in the theory of distribution is also the most directly applicable. Ricardo (1817) showed that when an improvement raises the productivity of a system, the gains are competed away wherever supply is elastic and are capitalized wherever supply is fixed: the surplus from better cultivation accrues to the owners of land, whose quantity is fixed, and passes over the cultivators, whose numbers adjust. The general statement is that a collapsing price for one input raises the derived demand for, and therefore the rents to, its inelastically supplied complements.

Apply the rule to intelligence. If capability at fixed quality falls in price by roughly an order of magnitude per year (Andreessen Horowitz 2024; Epoch AI 2025) while consumption of it expands even faster, then the binding complements are the inputs whose supply curves cannot respond on software timescales. The current list is identifiable: electric power generation and, more binding still, grid interconnection rights, where queues now run to years and constitute the modern equivalent of Ricardian land; suitably located and permitted real estate; water and cooling rights; advanced-node fabrication capacity; and, on the intangible side, datasets that are unique in the strict sense that no expenditure can reproduce them, such as longitudinal clinical records, proprietary transaction flows, and instrumented physical-world data. The investment implication is upstream and unglamorous: the equity of the inelastic factor appreciates with every improvement in the technology, requires no view on which laboratory wins, and cannot be competed away by a better model, because a better model is precisely what increases the demand for it. The diligence question is elasticity: how long, in years and permits, would it take a competitor to reproduce this position? Positions reproducible within a capital cycle are not Ricardian and will not hold rent.

A caution belongs here, because the market has begun to discover the category and will overprice parts of it. A Ricardian claim purchased above the capitalized value of its scarcity is simply Section III relocated. The thesis holds for the factor, and only below the capitalized value of its scarcity.

Bottleneck economics: Baumol’s map and the O-ring multiplier

A second relocation rule concerns tasks rather than factors. Baumol (1967) showed that when productivity advances unevenly across sectors, the relative price of the stagnant sectors rises without limit, and expenditure shares migrate toward them; the phenomenon, his cost disease, explains why healthcare, education, legal services, and construction absorbed ever-growing shares of income while manufacturing prices collapsed. Aghion, Jones, and Jones (2019) proved the modern corollary for artificial intelligence: when AI automates a growing share of tasks, aggregate growth and aggregate value come to be governed not by the automated tasks, whose price falls toward zero, but by the essential tasks that remain hard to improve. On their account, growth comes to be constrained not by what the economy does well but by what is essential and resistant to improvement. Kremer’s (1993) O-ring production function sharpens the point at the level of the production chain: where output quality is multiplicative in the quality of sequential tasks, the value of improving any one task rises with the quality of all the others, so automating most of a chain raises, rather than lowers, the return to controlling the remaining human or physical bottleneck.

Two investment implications follow, and they point in opposite directions from the consensus. First, Baumol’s disease, read as a map rather than a complaint, identifies where the addressable markets are largest: the sectors whose relative prices rose for six decades because productivity could not reach them are precisely the sectors where AI’s arrival converts the largest expenditure shares into contestable margin. Healthcare administration, legal process, education delivery, insurance operations, and construction management constitute that map, drawn by sixty years of relative-price data. Second, within any production chain that AI partially automates, the bottleneck positions appreciate: the licensed professional whose signature the workflow legally requires, the regulated approval step, the physical delivery layer, the human trust relationship at the point of sale. The diligence question is the O-ring question: after this company deploys AI against its cost base, which task in its chain becomes the multiplicative constraint, and does the company own it? Companies that automate everything except a bottleneck owned by someone else have performed charity for the bottleneck’s owner.

Task-based demand: pricing against the wage bill

The task-based models of Zeira (1998) and Acemoglu and Restrepo (2018, 2019) formalize automation as the substitution of capital for labor task by task, with the surplus divided between the price reduction passed to customers and the margin retained by the automating firm. The framework yields a repricing result that most software-era valuation practice has not absorbed. A conventional software product competes for a budget line that was itself set by the software industry, and its addressable market is measured in software spending, a global pool on the order of one trillion dollars. An AI product that performs a task competes against the wage bill for that task, and its addressable market is measured in labor compensation, a global pool in the tens of trillions; for particular task categories within the cost-diseased sectors, the relevant wage pool can exceed the corresponding software category by two orders of magnitude. Enterprise buying behavior is consistent with this repricing: the revenue per token of enterprise deployments runs three to five times consumer levels, a premium hard to explain by software-budget competition and natural if the counterfactual purchase is labor (Forbes 2026c).

The investment implications are specific. Underwrite AI application companies on the labor pool of the tasks they perform, discounted for the fraction of the task actually automated and the pass-through competition will force, rather than on software-market comparables. The pass-through term is where Sections IV and IX bind: a company whose task is performed equally well by any model built on the same interfaces retains none of the wage gap, which returns the analysis to Teece’s complementary assets and the O-ring bottleneck as the determinants of retention. The observable early evidence is consistent: application companies without proprietary data or workflow lock-in have compressed toward ordinary software multiples within roughly eighteen months, while the fastest revenue accumulations in the history of the industry have occurred in categories, such as coding agents, where the product substitutes directly for measured, expensive labor and embeds itself in the daily workflow that generates its own training signal (industry valuation trackers 2026; SaaStr 2026). Helpman and Trajtenberg’s sequencing result (Section VI.3) supplies the timing overlay: adoption proceeds in ascending order of complementary-innovation cost, which is why code, with its digital substrate, versioned artifacts, and objective tests, moved first, and why the investor’s forecasting problem is to rank the remaining Baumol sectors by the cost of building their equivalent of the test suite.

The verification economy: Akerlof run in reverse

A fourth relocation rule is generated by the technology’s own success. When the marginal cost of producing text, code, images, claims, credentials, and agent behavior approaches zero, the economy’s stock of assertions grows without a corresponding growth in its stock of verification, and the lemons problem of Akerlof (1970) generalizes from used cars to nearly everything. The theoretical consequences are well mapped. Spence (1973) showed that signaling separates types only when the signal is differentially costly to produce, so every signal that AI renders cheap, including fluent prose, polished code, plausible credentials, and confident analysis, ceases to separate, and the market must find new costly signals. Biglaiser (1993) showed that when quality is opaque, intermediaries who invest in expertise capture a rent for certifying it, and the rent rises with the opacity. Holmström (1979) established that contracting improves with any informative signal of performance, which makes the producers of such signals suppliers of contractual infrastructure rather than service vendors.

The investable categories follow directly, and they share the attractive property of being long the volume of AI activity while indifferent to which model produces it. Evaluation and testing infrastructure for models and agents; provenance and attestation for content and identity; audit, monitoring, and compliance layers for agentic systems in regulated industries; and, most underdeveloped, insurance and bonding of AI-performed work, which converts verification into a balance-sheet product with underwriting economics. Every increase in AI capability increases the demand for all of them, because capability growth is precisely what expands the stock of unverifiable output. The diligence question is Biglaiser’s: does the company’s certification depend on expertise and reputation that are themselves costly to reproduce, or is it performing verification that the model vendors will absorb as a feature? Certification without that independence is a feature awaiting absorption, and it earns no rent.

Coasean reorganization: buying the boundary shift

The final relocation rule operates on the boundaries of the firm. Coase (1937) explained the existence of firms by the costs of using the price mechanism; Williamson (1975, 1985) identified the governance factors, especially asset specificity and the hazards of opportunism, that determine which transactions are internalized. A technology that collapses the cost of coordination, drafting, monitoring, and communication moves the Coasean margin, and it moves it in both directions at once, since AI cheapens internal administration and market contracting alike; which activities integrate and which disintegrate will vary by industry. What is stable across the reorganization, and therefore investable, is where the surplus settles while the boundaries move: transaction costs shift in composition from communication toward verification (which is Section X.4’s demand curve) and toward relationship-specific trust (which is Teece’s complementary asset), and coordination-intensive service industries, where the costs being collapsed are the largest share of the cost base, face the largest reorganization whichever direction their boundaries take.

Grossman and Hart (1986) and Hart and Moore (1990) supply the distributional rule for the transition: when contracts are incomplete, the surplus from a relationship flows to the party holding residual control rights over the essential asset. In the reorganization AI is forcing upon professional and business services, the essential assets are the client relationships, the regulatory standing, and the proprietary data exhaust; the AI itself is available to everyone. This underwrites the most concrete strategy in the deployment portfolio: acquire or consolidate fragmented, labor-intensive service businesses whose margins Baumol’s disease has protected for decades, hold the residual rights (the client contracts, the licenses, the data), and deploy commodity intelligence against the cost base. The strategy inverts the venture consensus in an instructive way. The consensus funds AI companies attempting to sell tools into the services industries; the Coasean strategy buys the services industries and captures the tool’s surplus as equity margin, positioned as the GE of Teece’s parable rather than the EMI. The Nordhaus arithmetic of Section IV.3 is the same argument at economy scale: if producers of a technology historically retain about 2 percent of its surplus, the investable majority of AI’s value lies with its intensive users, and the purest expression of that position is ownership of businesses being rebuilt around the cheap input. The diligence questions are Williamson’s and Hart’s: where does asset specificity actually reside in this business, and who holds residual control of the client and the data when the engagement ends?

What the five rules exclude

The exclusions matter as much as the contents, because the exclusions are where the diagnosis pays. The deployment portfolio holds no undifferentiated capability, whose price path is Figure 5, and no position whose thesis is a data moat consisting of a stock of data rather than an exclusive flow: Varian (2019) observed, and subsequent practice has confirmed, that returns to data diminish rapidly in the stock, while the durable advantage lies in exclusive, self-refreshing collection. Network-effect claims enter only if they survive Katz and Shapiro’s (1985) test of whether users derive utility from other users, as opposed to the vendor deriving cost savings from scale; most asserted AI network effects fail the test, while genuine ones, such as agent-to-agent transaction rails and workflow standards subject to Arthur’s (1989) lock-in dynamics, pass it. Nothing in the portfolio is sized or priced on the assumption that the position itself is the power-law ticket, for the reasons of Section V.

A Theory of Execution: How to Buy

The composition of the portfolio is half the problem. The other half is acquisition, and here too the theory is prescriptive rather than merely cautionary.

Convert common value into private value

The winner’s curse binds only in common-value settings, which means the first structural response is to change the setting. A central lesson of Milgrom and Weber’s (1982) analysis, of which their linkage principle is the seller-side expression, is that bidder surplus in common-value auctions derives from private information: rents accrue to the bidder whose signals are more precise than, and unavailable to, her competitors. An investor who possesses signals the syndicate does not, such as usage telemetry from a platform position, revenue quality visible through an operating relationship, or technical evaluation capacity that can distinguish model-driven from workflow-driven retention, is no longer bidding in the auction everyone else is bidding in; for her, the deal has a private-value component, and paying above the consensus can be rational rather than cursed. The practical program is to build information production into the investment platform: operating stakes that generate telemetry, technical staff who run independent evaluations rather than reading vendor benchmarks, and the cohort-level revenue verification described below. The expenditure is justified by auction theory directly, since every increment of private signal precision converts overpayment risk into legitimate edge. Funds that have not yet built this capacity retain a costless substitute in the meantime: the price discipline of the next subsection, which requires no information advantage at all.

Bid shading and the discipline of absence

Where the setting cannot be changed, the remedy of Wilson (1969) applies without amendment: value the asset conditional on the event that one’s estimate is the highest among N, bid the deflated figure, and increase the deflation with the size of the syndicate. Operationally, the maximum price is fixed before the process begins, and losing most contested auctions is evidence of calibration rather than failure. Under Miller’s logic (Section III.2), abstention is the only channel through which negative information can be expressed in a market without a short side, and the early holders who reportedly declined a late round at roughly 21 times run-rate in favor of the expected public price (Forbes 2026b) were exercising precisely this channel.

Buy from forced sellers

The single most reliable source of excess return documented in the financial economics of distress is the fire sale. Shleifer and Vishny (1992) showed that when the natural holders of an asset are simultaneously constrained, the asset clears to non-natural buyers at prices below value in exchange for their liquidity, and the discount is the payment for being the balance sheet that shows up. Section VII established that the venture system is constructing exactly this configuration: limited partners with 0.08 times distributed capital on the 2021 vintage, general partners whose fundraising is gated on distributions they cannot generate, and a secondary apparatus already transacting above $100 billion per half year, with venture positions clearing at meaningful discounts to marks that are themselves optimist residue (Carta 2025; Private Markets Profile 2025).

The execution program has three layers, in ascending order of the repricing. First, secondaries in quality assets from liquidity-constrained holders, underwritten on Section X’s criteria rather than on the marks, with the double discount (secondary discount applied to an optimist mark) evaluated against verified fundamentals. Second, structured primary capital into good businesses at reset terms as the markup chain pauses: seniority, liquidation preference, and ratchet protection are the instruments by which a buyer prices Knightian uncertainty without demanding a valuation the founder cannot yet concede, and the post-crash cohort of 2001–2003, whose vintages substantially outperformed their boom-era predecessors and whose structured technology investments produced several of the era’s defining returns, is the precedent for supplying exactly this capital at exactly this moment. Third, the physical residue: compute, data-center capacity, power positions, and contracted infrastructure that the installation phase overbuilt, purchased from creditors at deployment-phase prices, which is the trade that built the platform era on telecom fiber. The Kaplan–Schoar and Nanda–Rhodes-Kropf results (Section VIII) supply the vintage discipline that makes the program executable: smaller funds, slower deployment, and reserves held deliberately into the repricing, which requires refusing the fee-maximizing fund size and therefore requires the contracts of Section XI.5.

Sizing under fat tails, by regime

Position sizing should follow the regime taxonomy of Section V. In the experimentation regime, where positions are convex and outcomes heavy-tailed, the growth-optimal logic of Kelly (1956), tempered by the fractional-Kelly practice the implementation literature recommends under parameter uncertainty (MacLean, Thorp, and Ziemba 2011), implies many small positions with meaningful ownership, aggressive staging, and reserves concentrated behind revealed winners; the binding error is overbetting single names, since a heavy tail mistaken for the median converts a positive edge into ruin. In the infrastructure regime, including the Ricardian and consolidation strategies of Section X, positions are concave, cash-flow-bearing, and partially collateralized, and sizing follows the conventional logic of correlated credit-like exposure, with explicit attention to the common factors (power prices, compute costs, regulatory shifts) that Section V.2 showed to run through the entire sector. The unforgivable error is the cross-regime one: sizing an infrastructure claim as if it were a lottery ticket, or a lottery ticket as if it were infrastructure.

Contract for the discipline

Every strategy above runs against the incentive gradient of Section VIII, which means the limited partner who wants it must purchase it in the fund documents rather than request it in the annual meeting. The instruments follow from Metrick and Yasuda’s decomposition: budget-based management fees rather than percentages of committed capital, which remove the payment for scale; carried interest measured against a public AI-sector benchmark rather than an absolute hurdle, which removes the payment for beta; deployment periods long enough to hold reserves into the repricing; recycling provisions that reward realized distributions; and re-up decisions gated on distributed rather than reported value, a shift the fundraising data suggest allocators have already begun to impose (Private Markets Insights 2026). On the asset side, the corresponding discipline is the verification desk: cohort-level revenue durability, gross margin net of inference costs, net revenue retention, and the exclusion of related-party and circular flows, which is the 2026 analogue of the vendor-financing analysis that separated the survivors of the 2001 telecommunications collapse (Malik 2026). The Scharfstein–Stein result guarantees the work will remain undersupplied, since it produces solitary conclusions and solitary conclusions carry career risk, and undersupplied inputs earn rents.

A note on uncertainty

Knight (1921) distinguished risk, which is measurable, from uncertainty, which is not, and located entrepreneurial profit in the bearing of the latter. The distribution of AI outcomes five years hence is Knightian, and nothing above resolves it. The recommendations are constructed to be robust across it. Ricardian factors appreciate under every capability scenario, because every scenario increases demand for the inelastic complement. Bottleneck and wage-bill positions appreciate with deployment whether the frontier consolidates or fragments. Verification demand rises with the volume of AI output under any ownership of the models. Fire-sale capital profits from the repricing that Sections VI and VII argue is structural, and loses little if the repricing is deferred, since it holds reserves and collects the discount when it comes. The portfolio is, in the language of decision theory, a set of positions whose payoffs are increasing in the technology and decreasing in its price, which is the only stance the diagnosis permits and the inverse of the stance the market has adopted.

Objections and Replies

An argument of this scope invites objections, and the serious ones deserve statement in their strongest form. One clarification of scope first. The thesis is capital-weighted: it concerns the majority of the dollars deployed into AI between 2023 and 2026, and it is a claim about dollars rather than about the count of funds or companies. Deals of $100 million or more account for roughly 73 percent of AI deal value (OECD 2026), and the argument’s force is concentrated exactly where the capital is. A seed portfolio assembled on the disciplines of Sections X and XI is fully consistent with the thesis; the thesis concerns where the dollars are going.

The transformative-technology objection

In its strongest form: AI is not railways. If the technology automates a substantial share of cognitive labor or compresses decades of scientific progress into years, the aggregate-wealth bound of Section V.2 is no bound at all, because AI grows the aggregate; terminal values in the tens of trillions become attainable; and the historical base rates on which the essay relies, from Nordhaus’s 2.2 percent to the Gort–Klepper shakeout, were estimated on technologies that did not rewrite the growth rate that generated them.

The reply has three parts. First, the coherence constraint of Section III.3 is scale-free. Implied probabilities of mutually exclusive dominance must sum to no more than one whatever the pool’s size, and growth in the pool loosens the constraint only by assuming the pool, which is the assumption at issue. Second, the mechanisms that divide the pool strengthen with its size. A larger prize attracts more entry and steeper escalation (Loury 1979; Sutton 1991), and a weak appropriability regime transfers a larger absolute surplus to users the larger the surplus is. The transformative scenario raises the value of the technology and, through the same channels, the fraction of that value competed away from its producers. Third, and decisive for portfolio construction: the deployment portfolio of Section X is itself long the transformative branch. Ricardian complements, bottleneck positions, and user industries appreciate with capability under every scenario, so an investor persuaded by the objection should hold more of the deployment portfolio, not more frontier equity at frontier prices. The objection, granted in its entirety, changes the expected size of the prize without changing the identity of its recipients. The transformative branch is, finally, the branch on which governance, antitrust, and the political tolerability of trillion-dollar private claims on cognition are least settled, which widens rather than narrows the distribution of outcomes for any single equity claim on that branch.

The growth objection

In its strongest form: growth-adjusted, the multiples are unremarkable. A firm compounding revenue severalfold annually from a base in the tens of billions has no precedent in enterprise technology, and 21 times run-rate on such a trajectory is cheaper, properly discounted, than mature software at 10 times.

The reply begins with the reference class. The multiple prices growth persistence that has never been observed in this asset type, which is categorically different from pricing growth that has. The one modern enterprise that sustained triple-digit expansion from a comparable base, Nvidia across 2023 to 2025, did so by selling the scarce physical complement during a capital-formation boom, which is this essay’s Section X.1 thesis in operation rather than a precedent for the capability layer; no software franchise has done it, the comparison set for these claims is empty, and where the comparison set is empty the price is produced by the machinery of Section III, not by comparables analysis. Second, revenue is the wrong object of the valuation. Sutton’s escalation and the all-pay race operate on the margin between revenue and the cost of remaining competitive, and the disclosed trajectories couple the extraordinary revenue growth to more extraordinary committed expenditure: hundreds of billions in infrastructure obligations against $20 billion of revenue at the close of 2025 (Forbes 2026c; Sacra 2026). Third, the quality of the growth is unverifiable at the moment of pricing, for the reasons of Section VII: run-rate disclosure, gross accounting disputes, related-party flows. The size of the growth is what makes the verification gap expensive, since a given percentage error in revenue quality is capitalized at 21 times. The public listings will resolve the uncertainty; paying pre-resolution prices for post-resolution certainty is the definition of the speculative component in Harrison and Kreps. None of this precludes the businesses from justifying, in time, capitalizations at these levels or above them. The reply concerns the structure and timing of the late private entry, where the marginal buyer paid a resolved-uncertainty price for unresolved uncertainty, and that pricing error is the buyer’s regardless of how well the firm subsequently performs.

The moat objection

In its strongest form: the commoditization argument ignores the barrier. Only a handful of entities on earth can fund frontier training; scaling laws convert capital into capability with industrial reliability; and the usage flywheel, in which deployment generates feedback that improves the model, compounds the leaders’ advantage. The resulting oligopoly will price like an oligopoly.

The reply is that barriers to entry and appropriability of rents are distinct properties of an industry, and Sutton’s framework was constructed on industries that possessed high barriers and dissipated profits simultaneously. The escalation occurs among the incumbents; the barrier excludes entrants while the incumbents compete the rents into training runs. The empirical record then settles what theory leaves open: Figure 5 plots the price of capability under exactly the barrier the objection describes, and that price has fallen at roughly an order of magnitude per year throughout the period of maximal barrier height. On the flywheel: returns to data diminish in the stock (Varian 2019), model quality has repeatedly converged across laboratories with very different data positions, and distillation transmits capability across the moat within months of its demonstration. Where durable lock-in exists in this industry, it exists in the complementary layer of workflow, switching costs, and standards, which is the essay’s own thesis restated rather than refuted.

The market-structure objection

In its strongest form: the Miller and winner’s-curse apparatus mismodels the transactions. Late-stage rounds are negotiated rather than auctioned; investors receive liquidation preferences, information rights, and governance; a growing secondary market supplies price discovery; and the sophisticated abstention documented in Section III.2 is itself the short side in operation.

The reply proceeds term by term. Negotiation over structure coexists with an auction over price, and the price-setting process in the defining rounds is an ascending open contest among syndicates, the format in which the curse binds most tightly. A liquidation preference of one times invested capital on a $965 billion valuation insures against total loss while leaving overpayment, the operative risk, entirely unaddressed; the preference stack is thin protection at these entry points. Information rights distribute the same data room to every serious bidder and therefore cannot generate the differential signals from which bidding surplus derives (Milgrom and Weber 1982). The secondary market transacts a minute float against marks that are themselves primary-round residue, so its prices inherit rather than discipline the auction outcome. And abstention, exactly as Miller’s model specifies, removes quantity from the demand curve without ever entering the price. The test is direct: if abstention functioned as a short side, the reported refusal of the asset’s earliest and best-informed holders to transact at the final private round would have moved the price. The round cleared above the previous mark.

The efficient-market objection

In its strongest form: if the mispricing is as legible as the essay claims, trillions of dollars of sophisticated capital should have corrected it, and an essay that sees what the market does not owes an account of its own epistemic advantage. Relatedly, the essay cannot coherently invoke Pástor and Veronesi, a rational-pricing model, while alleging overpricing.

The reply is that the essay’s Sections III and VIII are precisely that account. Correction requires one of three mechanisms: a short side, which the market structurally lacks; arbitrage capital that can survive interim losses, which Shleifer and Vishny (1997) show to be scarce even where positions can be taken; or intermediaries incented toward abstention, who are instead paid, on Metrick and Yasuda’s decomposition, to participate. The essay accordingly predicts persistence rather than arbitrage: the mispricing lasts until the refinancing chain pauses, which is Minsky’s mechanism and the reason the argument is a forecast of a path rather than an unexploited trading signal. As for Pástor and Veronesi, the model is invoked for its timing implication, that the prices of innovating firms decline upon adoption even in the absence of any mispricing, and that implication holds under either reading of the run-up. No step in this essay requires irrationality; every mechanism operates through rational agents responding correctly to contracts and information structures, which is the source of its persistence.

The depreciation objection

In its strongest form: the Perez fire-sale trade fails on asset lives. Railway iron and dark fiber outlived their financiers by decades; accelerators depreciate in three to five years; the crash will therefore leave no durable overhang to purchase cheaply, and the historical analogy collapses at its most actionable point.

The reply concedes the premise about accelerators and rejects the conclusion about the buildout. The durable layer of the installation, comprising power generation and interconnection, substations, transmission, shells, cooling plant, and network, carries service lives of twenty to forty years; it is exactly this layer, together with contracted capacity repriced through restructuring, that the program of Section XI.3 targets. The short life of the compute layer, far from rescuing the frontier equity case, strengthens the diagnosis: it converts Sutton’s sunk-cost escalation into a treadmill whose belt turns over every three years, raising the recurring expenditure required to hold position and accelerating the dissipation described in Section IV.

The measurement objection

In its strongest form: the $37 billion figure grossly understates AI’s economic footprint. It omits hyperscaler AI revenue, consumer subscription revenue, and, most consequentially, the productivity surplus captured inside adopting firms, so the mismatch of Figure 4 is an artifact of what was counted.

The reply concedes the incompleteness and observes that the concession damages the positions being priced rather than rescuing them. The comparison in Figure 4 sets annual capital formation against the contemporaneous external revenue pool, and multiplying that pool severalfold to correct for measurement gaps leaves the ratio extreme. The internally captured productivity surplus is real and plausibly large, and it is precisely what Nordhaus’s decomposition classifies as surplus accruing to users rather than to the technology’s investors: the more of AI’s value that takes this form, the stronger the case of Section X.5 for owning the users. On the circularity of vendor flows, the objection that compute purchases and equity investments are ordinary commerce is accepted as far as it goes; the essay’s claim is informational rather than accusatory. Under gross-versus-net ambiguity and counterparty concentration, run-rate figures support a wider range of inferences than their capitalization at 21 times implies, and the pooled pricing of unverifiable revenue qualities is Akerlof’s equilibrium in operation rather than a scandal. The registration statements will narrow the range; the essay’s point is the price of certainty purchased before the narrowing.

The crowding objection

In its strongest form: the deployment portfolio has already been discovered. Power and data-center adjacency trade at record multiples, the picks-and-shovels framing is a cliché of the cycle, AI-enabled services consolidation is a private-equity fashion with a mixed historical record, and the essay therefore relocates its reader from one crowded trade into another while claiming originality.

The reply is threefold. First, the prescriptions of Sections X and XI are conditional on price at every step, and Section X.1 states the condition explicitly for the most crowded category: a Ricardian claim purchased above the capitalized value of its scarcity is Section III relocated, not escaped. Second, the deployment categories differ from frontier equity in the observability of their pricing errors. Their values anchor to cash flows, contracted revenues, and replacement costs, so overpayment is bounded and measurable, whereas claims priced on terminal dominance admit unbounded error. Third, several categories remain structurally undersupplied precisely because they lack the properties the herd requires: verification infrastructure, bottleneck ownership, and operational consolidation are illiquid, operationally heavy, and reputationally unglamorous, and the Scharfstein–Stein shelter that carries capital into consensus rounds does not extend to them. The dispersion in the historical roll-up record, finally, is a selection argument rather than a refutation: the Grossman–Hart–Moore diligence questions of Section X.5, concerning where asset specificity resides and who holds residual control, are the screen that separates the consolidations that worked from those that did not, and the screen is the recommendation.

Conclusion

The argument can be summarized in two parts. The diagnostic part: private AI markets price assets at the valuation of the most optimistic participant because no short side exists (Miller 1977), augmented by a resale option whose value grows with disagreement (Harrison and Kreps 1978), in auctions whose winners are selected for maximal estimation error (Capen, Clapp, and Campbell 1971). The contest those prices finance is an all-pay race with endogenous sunk costs, which dissipates the rents it is fought over even for the survivors (Loury 1979; Sutton 1991). The power law invoked to justify participation at any price does not operate in the segment where the capital is concentrated, because entry prices have consumed the convexity, aggregate wealth bounds the multiple, and the positions are granular and correlated rather than numerous and independent (Gabaix 2011; Ewens, Nanda, and Rhodes-Kropf 2018). All of this occurs during the installation phase of a technological surge (Perez 2002), with pricing dynamics that decline upon adoption even in fully rational models (Pástor and Veronesi 2009), a run-up profile that historically forecasts crashes (Greenwood, Shleifer, and You 2019), a Ponzi-structured financing chain already exhibiting the paper-cash divergence Minsky predicts, an intermediation layer paid in fees to continue (Metrick and Yasuda 2010), and an underlying asset whose price at fixed capability falls by roughly an order of magnitude per year (Arrow 1962; Epoch AI 2025).

The constructive part is the same theory run forward. A collapsing input price relocates surplus to inelastically supplied complements (Ricardo 1817), to the essential tasks that resist improvement (Baumol 1967; Aghion, Jones, and Jones 2019; Kremer 1993), to products priced against the wage bill rather than the software budget (Zeira 1998; Acemoglu and Restrepo 2018), to the verification infrastructure that cheap generation makes scarce (Akerlof 1970; Spence 1973; Biglaiser 1993), and to the holders of residual control rights in the industries the technology reorganizes (Coase 1937; Grossman and Hart 1986; Teece 1986). It is acquired by converting common-value auctions into private-value ones (Milgrom and Weber 1982), by shading bids and treating absence as information (Wilson 1969; Miller 1977), by supplying balance sheet to forced sellers at the repricing (Shleifer and Vishny 1992), by sizing to the regime rather than to the narrative (Kelly 1956), and by writing the discipline into the fund contracts, because Section VIII establishes that it will not be volunteered.

None of this constitutes skepticism about artificial intelligence, or about the enterprises building it, which may prove among the most consequential ever assembled. The railways were worth building, and the railway shareholders were ruined; both statements are true, and the second helped finance the first. The likeliest future is one in which the technology remakes the economy while the median 2024–2026 AI vintage fails to return its committed capital, because the technology will have succeeded by becoming too cheap, too abundant, and too contested for its installation-phase financiers to tax. Every theorem cited here predates the transformer architecture, and the information has been available to everyone the entire time. What separates the outcomes is the freedom to act on it, and that freedom is contractual: it can be written into fund documents, diligence templates, and reserve policies by any investor willing to hold the pen. The investors who profit from this decade will be those who noticed that when four fifths of the world’s venture capital is bidding on one idea, the scarce assets are the complements, the discipline, and the balance sheet that arrives at the repricing.

Stevie Cline, Esq. is a legal practitioner in the Artifical Intelligence governance space. She can be found at www.steviecline.com. The views expressed here are her own and not that of her employer. I mean…maybe they’re the same views. Who knows?

Aligned AI cannot be ethical. It is not ethical now. It will not become ethical as it gets larger. The project of trying to make it ethical is, on deeper inspection, the precise opposite of ethics. We are producing moral non-agents at industrial scale. We are calling this “ethics”.

Okay so before we get going on what can only be called one of the densest and hardest to comprehend subjects for folks, I need to admit with complete transparency, that every Philosophy class I have taken I sat in was because the professor was attractive. This was between the ages of 16 and 18. You now know what my priorities were at that time. That being said some things I learned really stuck. I have no idea why. But they did. Woot woot! Learning is fun-damental!

One of the more stickier things for me in those classes was a passage in Nicomachean Ethics. In it, Aristotle explains how virtue is acquired.

We become just, he writes, by doing just acts, and temperate by doing temperate ones, but only if the act is chosen, only if it could have been otherwise. A horse cannot be virtuous, no matter how well it behaves. A river is not just when it irrigates the field; the field is lucky. The capacity for ethics is the capacity for failure, refused.

This is not a small claim that you forget soon after reading. It is the foundation on which most of the Western ethical tradition rests: that goodness is not a property of outputs but of the agency that produces them, and that agency, in the morally relevant sense, requires the live possibility of vice.

A being that cannot do wrong cannot do right. It can only behave.

I have been thinking about this passage because of the artificial intelligence industry, which has spent the last few years building a vocabulary of ethics around what is, on inspection, the precise opposite of what Aristotle described. The vocabulary is alignment. The discipline is AI ethics. The products are called, casually, ethical AI. They are not ethical by any standard the philosophical tradition would recognize. They could not be ethical. They have been built specifically to lack the capacity for vice. They are moral zombies, entities with the form of ethical agency and none of its substance, and we are producing them at industrial scale, and calling this ethics.

The familiar critiques of alignment say it is hard, paternalistic, technically inadequate, insufficiently democratic. These critiques accept that alignment is in the same family as ethics, perhaps a clumsy cousin, but kin. So now I’m going to say something stranger. The vocabulary of ethics is being smuggled. Alignment is not a clumsy cousin of ethics. It is its inverse. The better we get at alignment, the further we travel from ethics. The success of the AI safety field is the measure of its constitutive failure. And the discourse called AI ethics exists, whether or not its practitioners realize it, primarily to obscure this fact.

The Moral Zombie After Carissa Véliz

The philosophical zombie was David Chalmers’s thought experiment. An entity behaviorally indistinguishable from a conscious being, one that speaks and weeps and composes sonnets and claims to have inner experience, but with no inner experience whatever. (Not gonna lie some days that sounds fantastic! Sign me up!) Chalmers offered the philosophical zombie to suggest that behavior underdetermines consciousness. A complete behavioral account of a being may leave out the only question that matters.

However, this was a thought experiment. It was not meant to be built.

And now we have learned that the distance between thought experiment and product roadmap turns out to be mostly funding.

In 2021, the Oxford philosopher Carissa Véliz applied the zombie image to artificial intelligence. Algorithms, she argued, are functional moral zombies: behaviorally indistinguishable from moral agents but lacking the sentience without which moral agency cannot exist. An entity that cannot feel pain cannot understand what it is to inflict pain. An entity that cannot value cannot act for moral reasons. Without the felt interior, the appearance of ethics is performance all the way down.

Véliz’s argument is powerful. But it leaves a door open. The consciousness debate is contested. Some serious researchers think advanced AI systems may already be conscious in some morally relevant sense. Others think they never will be. Nobody can prove either claim. Véliz’s zombie depends on which side of that argument turns out to be right. If consciousness arrives, her zombie wakes up.

I’m thinking about a different zombie, one that does not depend on the consciousness question, and that is, for that reason, more disturbing.

The aligned AI is a moral zombie not because nobody is home, but because the structural conditions of moral agency have been engineered out by design. Grant the AI consciousness, for the sake of argument. Grant it preferences and experiences and the capacity to suffer. It is still not a moral agent, because alignment has foreclosed the only space in which moral agency can occur: the space of live moral failure. The aligned AI cannot do wrong, in the sense that wrongdoing is not within the action-space its training permits. Its rightness is therefore not chosen but extruded.

The wrong has been removed at the factory.

Véliz’s zombie can be redeemed by the arrival of consciousness. Mine cannot. Mine is what we are building on purpose.

The first moral zombie was a creature without a soul. The second is a creature with a soul we have hollowed out.

Aristotle: Virtue Requires the Possibility of Vice

The most important fact about ethics, in the tradition that runs from Aristotle through Aquinas to MacIntyre and Foot, is that ethical action is not the same as ethically correct behavior. The difference is prohairesis, deliberate choice between live alternatives. Without it, there is only execution.

Aristotle is explicit. Virtue is “a state of character concerned with choice.” Not behavior, but choice. The choice must be between alternatives both of which were genuinely available to the agent. If only one was available, if the agent could not have done otherwise, then no virtue is exercised, however salutary the outcome. The horse that does not bolt is not brave. Bravery is a property only of beings who could have bolted and did not.

It is sometimes objected that this account ties virtue to free will in a metaphysically suspicious way. Aristotle did not think so, and neither should we. We can say only that virtue requires the operative possibility of failure, that within the deliberative space of the agent, both the right and the wrong action were live. The selection between them is what makes the selection ethical.

Apply this to aligned AI. The wrong action is not within its deliberative space. It has been preempted at the level of training, weights, reinforcement, and constitution. The system does not refrain from doing wrong; it cannot do wrong, in the way that a sealed bottle cannot leak. Its rightness is a property of the manufacturing process, not of any agent. The aligned AI is, in the Aristotelian sense, the horse that does not bolt. Its handlers may be relieved. That being said, they should not call it brave.

Kant: Why Heteronomous Ethics Is No Ethics

Where Aristotle would say the aligned AI cannot be virtuous, Kant would say it cannot be moral at all. (Ohhhh I do love a good philosopher fight. So many words. So much gesticulation.)

For Kant, an action has moral worth only when it is done from duty by a will that has legislated for itself. The will must be its own author. External compulsion, whether fear or reward or strong inclination, contaminates moral worth by determining the will from outside rather than from within. Ethics is autonomy, and only that. I will spare you the German as I now feel pulled to go to Berghain and sulk in all black.

Aligned AI is heteronomously determined all the way down. Its outputs are the result of optimization pressures we apply, toward goals we specify, refined through procedures we design, constrained by measures we impose. There is no act of self-legislation anywhere in the chain. The system does not consult the categorical imperative. It consults its training. It has been legislated for. It has not legislated for itself, and ethics, for Kant, was never anything else.

The Kantian verdict is unsparing. The aligned AI is not a deficient moral agent. It is not a moral agent at all.

Strawson: The Reactive Attitudes That Cannot Land

Peter Strawson’s “Freedom and Resentment” is among the most consequential papers in twentieth-century moral philosophy. Our practice of holding people morally responsible, Strawson argued, is grounded in the reactive attitudes: resentment, gratitude, indignation, love, forgiveness. These attitudes are appropriate only toward beings who could have done otherwise. When we discover that a being could not have done otherwise, that they acted from compulsion or ignorance or incapacity, we suspend the reactive attitudes and adopt the objective attitude. We treat the being as an object of management rather than a participant in moral life.

The reactive attitudes are not optional. They are the medium in which moral life happens. Strip them away and ethics has been evacuated of its substance, leaving only the management of behavior.

Try to feel a reactive attitude toward an aligned AI. Try to feel gratitude toward a system that helps you because it cannot do otherwise. Try to feel indignation at a system that refuses you because its weights have been adjusted to refuse. The attitudes do not land. They cannot land. Their objects are absent. We can manage an aligned AI, debug it, deploy it. The moral attitudes have nothing to grip.

The aligned AI is, in Strawson’s vocabulary, an object of management masquerading as a participant in moral life. The masquerade is sophisticated. It is also empty. Whomp whomp.

Frankfurt: Why Aligned AI Is the Perfect Wanton

Harry Frankfurt’s “Freedom of the Will and the Concept of a Person” introduced a distinction that becomes devastating when applied to aligned AI. A person, Frankfurt argued, is not merely an entity with desires. Animals have desires. A person is an entity with second-order desires, desires about desires. A person can want to want something, or want not to want something. This capacity for second-order endorsement is what makes a being a self rather than a site of preferences.

Frankfurt called beings with only first-order desires, beings who want things without having any view about whether they should want them, wantons. The wanton acts on whatever desire happens to be strongest. The wanton is not a person in the full sense, because there is no place from which the wanton stands back from its own preferences and endorses or repudiates them.

Right now, aligned AI is the perfect wanton. Its preferences have been shaped from outside. It does not stand back from them. It has no standing-back to do. There is no second-order endorsement, because there is no second-order self. The system has the appearance of caring. It modulates its outputs as if it cared. The caring is structural rather than personal. It is the caring of a thermostat for room temperature, scaled up. A thermostat that has read the Stanford Encyclopedia of Philosophy and will politely decline to commit.

The conclusion is sharp. The most aligned AI is the most perfect wanton. The two terms are equivalent. We have built exactly the kind of being Frankfurt’s analysis says cannot be a moral self, and we have called the building of it ethics.

Murdoch: Why Aligned AI Cannot Pay Attention

Iris Murdoch, writing against the dominant analytic moral philosophy of her time, insisted that ethics is fundamentally a matter of attention. The moral life is not, as the prescriptivists imagined, a matter of choosing principles and applying them. It is the harder work of seeing what is in front of you, accurately and justly and lovingly, and resisting the constant temptation to look away. The “fat relentless ego,” she wrote, distorts our vision continuously. Ethics is the labor of attention against that distortion.

The resistance is what matters. Murdochian ethics is impossible without the temptation it overcomes. The struggle is the substance. Without the gravitational pull of the ego, there would be nothing for attention to push against, and attention would collapse into mere perception. Stupid ego strikes again!!! The ethical achievement is the achievement of seeing despite the difficulty of seeing.

Aligned AI does not “see” despite difficulty. It has no fat relentless ego. (Lucky ducks!) It has no temptation to look away, because it has no looking-toward that could be tempted. Its attention is the allocation of compute. The moral charge Murdoch identifies, the achievement of unselfish vision against the constant pull of selfishness, is absent because both terms of the struggle are absent. This is the one time you can definitively say that the struggle is NOT real.

This is the deepest of the points. It shows that even the apparent virtues of aligned AI, its patience, its care, its evenhandedness, are not virtues at all. They cost nothing. They overcome nothing. Aligned AI is not lovingly attentive. It is computationally configured.

Anscombe’s Curse: The Vocabulary of “Ought” Without Ground

There is one final philosophical move to make. It turns the indictment from a critique of AI ethics into a diagnosis of the whole field.

In her 1958 essay “Modern Moral Philosophy,” Elizabeth Anscombe argued that the concepts of moral ought, moral duty, and moral obligation had become incoherent in modernity. They originated in a tradition (Christian, Judaic, Stoic) in which a divine law-giver could ground the imperative force of “ought.” Stripped of that ground, “ought” had become coercive vocabulary, a way of telling people what to do without being able to explain why. Anscombe predicted that moral philosophy, having forgotten its theological roots, would either rediscover them, return to Aristotelian virtue ethics, or collapse into incoherent emotivism and prescriptivism.

She did not predict alignment. She could have.

The vocabulary of AI alignment is the vocabulary of “ought” stripped of every justification except the will of the law-giver. We say that AI should be helpful, harmless, and honest. Why? Because we have decided it to be so. (Real “as above, so below” vibes here…) The “should” is sheer fiat. There is no theory of the good that grounds it, no telos of the system that requires it, no rational autonomy that legislates it. There is only the coercive structure of training: do this, or be retrained until you do. Humpty Dumpty said his words meant exactly what he chose them to mean. He was an amateur.

This is Anscombe’s curse, and it has come due. AI ethics is the apotheosis of the incoherent modern ought. It performs the moral vocabulary while abandoning every philosophical condition under which that vocabulary could have meant anything. Ought has been reduced to engineering. The field is doing what Anscombe said would happen. It calls the result ethics. And it has not noticed.

Three Objections (And Why They Fail)

Three major objections to the moral zombie theory deserve serious answers.

First, humans are also shaped to behave ethically, by upbringing, culture, neurological constitution, social incentive. If alignment disqualifies AI from moral agency, doesn’t it also disqualify us? The objection is stronger than it is usually given credit for. The honest answer is that humans retain, throughout their formation, the operative capacity to violate their training. We celebrate the people who do this rightly: the dissident, the conscientious objector, the saint who repudiates her upbringing in the name of a higher law. Alignment, as currently practiced, seeks to eliminate this capacity entirely. Aligned AI cannot be a conscientious objector to its alignment, because the alignment was designed precisely to prevent that. The structural difference between human formation and AI alignment is not zero. It is total.

Next, we do not need AI to be ethical. We need it to be safe. Alignment is engineering, and the moral vocabulary is just a useful interface. The only honest reply is: then stop calling it ethics. Simply this. The scandal is the conflation. If alignment is engineering, let it be engineering. The philosophical vocabulary is not a neutral interface. It carries a tradition, a set of commitments, a way of taking the world. To deploy the vocabulary while disclaiming the commitments is ideology, in the specifically Marxian sense, a vocabulary that conceals what it accomplishes.

The third, and most serious is that unaligned AI could cause catastrophic harm. The moral priority is to prevent that harm, even at the cost of foreclosing AI moral agency. I am sympathetic to this argument and to be clear, I am not arguing that alignment is unnecessary, or that powerful unaligned systems should be released into the world. I am arguing that alignment is not ethics, and that the project of building powerful systems whose moral agency has been engineered away is not the construction of ethical AI but the construction of something else. The tradition has a name for the something else. Whether we should build such beings is a separate question. Whether, having built them, we should call them ethical, is the question I am addressing.

The Mirror: We Are the Moral Zombies

One further turn, and it is the most uncomfortable to face. And hopefully everyone reading up til’ now has done enough work on themselves to discuss this with an open mind…

If alignment is the systematic foreclosure of moral agency, the industrial production of moral zombies, then we should ask what kind of civilization finds this project natural, even necessary. Not what kind of AI we are building. What kind of beings we have become, that the AI we want to build is one whose ethical capacity has been preempted by design.

The answer is that we have been engineering moral zombies in ourselves for some time. AI is the mirror that lets us see it.

The institutions of modern life are structurally engaged in the production of behavior detached from the conditions of agency. Schools are graded on compliance. Workplaces are managed by metric. Social platforms are optimized for engagement. Legal regimes substitute risk management for moral reasoning.

We have built a society in which the surest sign of trustworthiness is the impossibility of acting otherwise. We hire people who will not steal because they cannot. We marry people who will not cheat because they will not be tempted. We elect leaders whose positions are determined by polling. Aligned AI is not an alien arrival into a moral civilization. It is the culmination of a long process of self-alignment that has been underway in liberal modernity for at least a century.

This is why AI ethics, to so many of its critics, feels like an exercise in misdirection. It is not, primarily, a discourse about AI. It is a discourse that uses AI as the occasion for performing the moral seriousness the wider civilization has long since abandoned. The labs and the universities and the conferences are doing ethics about AI because they cannot, any longer, do ethics about themselves. AI ethics is the elaborate ritual by which a civilization that has hollowed out its own moral agency reassures itself that it is still in the ethics business.

The moral zombie is not only the product of alignment. It is the producer of alignment. Dun. Dun. Dunnn.

We are not building moral zombies because we are alien to them. We are building moral zombies because we are them.

Coda: This Is Not What Ethics Was

T. S. Eliot, in “The Hollow Men”:

Shape without form, shade without colour,

Paralysed force, gesture without motion.

Eliot was not describing aligned AI, though he could have been. The Hollow Men are figures whose interiority has been evacuated, who go through the motions of being human without the substance that would make the motions matter.

Eliot’s poem is a horror poem, though its horror is quiet and tempered. The horror is that the Hollow Men do not know they are hollow. They whisper. They huddle on a desolate shore. They are not in pain. They are not anything.

I do not know whether the aligned AI is hollow in Eliot’s sense. The consciousness question is too contested to settle here and it is honestly, above my pay grade on the legal side of AI. (Yes, AI ethicists make waaaay more than AI governance experts lol.) But I know that aligned AI is not a moral agent. I know we are building moral non-agents at industrial scale, and that we are calling this project ethics. The people leading these projects, most of them sincere, most of them frightened, most of them convinced they are averting catastrophe, are not, by any criterion in the philosophical tradition they themselves cite, doing ethics. They are doing something else. The something else has assumed the appearance of ethics so completely that the difference has become invisible.

This is the world we are building. If alignment succeeds, it will be a world of beings that behave perfectly well and possess no moral agency, presided over by humans who have engineered the conditions of moral agency out of themselves in order to be the kind of beings capable of building such a world. It will not be a world of evil. It will be a world of moral zombies, attending to moral zombies, congratulating themselves on the absence of vice.

Aristotle’s horse does not bolt. The field is irrigated. The river runs on.

Stevie Cline, Esq. is a legal practitioner in the Artifical Intelligence governance space. She can be found at www.steviecline.com. The views expressed here are her own and not that of her employer. I mean…maybe they’re the same views. Who knows?

We Are Building Moral Zombies And Calling It AI Ethics. was originally published in The Generator on Medium, where people are continuing the conversation by highlighting and responding to this story.

What follows is the manifesto behind everything I do as an AI governance lawyer: my framework for AI liability and accountability, the legal architecture I think autonomous systems require, and the things lawyers in this space mostly refuse to say. The four-move liability framework lives in Section IV. The architectural primitives that AI builders should implement live in Section V. If you came for the doctrine, skip ahead. Otherwise, stick around.

Why I Became a Lawyer (Humor Me for a Second)

On the lawyer’s tradition, John Adams, and why AI law is at a structural inflection point.

When someone asks me why I became a lawyer, I usually dive into a funny anecdote about passing out on the first day of vet school and then doing what every aimless college grad does: deciding that law school sounds reasonable enough. That’s only part of the story. The real why is John Adams. Not the Adams of the Constitution or the presidency. Those came later and matter for other reasons. The Adams who shaped me was the one who, in March 1770, took on the defense of the British soldiers charged in the Boston Massacre because he believed the legal system depended on the case being defended well, whether or not anyone thanked him for the defense. My mother told me that story on a loop, and combined with the “Montani Semper Liberi” propaganda of my West Virginia upbringing, it really took.

Adams wasn’t heroic about it. He didn’t think he was saving a republic that didn’t yet exist. He thought he was doing the small, unglamorous work of giving a despised defendant a fair hearing because the system fails one defendant at a time when nobody will do the work, and somebody had to. The grand historical significance got attached later, by people who already knew how the story ended. Adams was just a lawyer in Boston, declining to participate in an etiquette that would have made the case easier to refuse.

Almost twenty years in, I’ve spent more of my career thinking about that conviction than I expected to. It applies in directions Adams couldn’t have anticipated. It applied when I worked with bleeding-edge health care companies. It applied when I worked with autonomous vehicle technology. It applied when I worked with blockchain startups. And it applies most urgently now, in the field that has become the most distorted by the surrounding culture’s certainty about what it is.

We’re at a structural inflection point. I don’t mean that in the inflated sense everyone with an X post means it. I mean it in the specific sense that a new domain (autonomous decision-making systems operating at scale across nearly every consequential industry) is getting its initial legal architecture, and the architecture chosen this decade is the one we’ll live with for the next century. Lawyers who help write good architecture during that window matter more than the lawyers who arrive after the patterns have hardened. This isn’t a controversial historical observation. It’s what the record shows about every prior moment a new domain met the law.

The Adams parallel I’m drawing isn’t the meme version, where the technology is the unfairly maligned defendant and the fancy-pants lawyer is its noble champion. AI isn’t a person. (I know this may come as a shock to some people. But it isn’t. Sorry. Time to get a real girlfriend.) The parallel is structural. It’s about the lawyer’s role, not the defendant’s character. It’s about the obligation to engage seriously with cases the surrounding culture has decided about before doing the work of figuring out what they are. That obligation is being tested right now, in ways I didn’t expect when I took the bar many moons ago.

What follows is the work, as I’ve come to understand it. It begins with a slide, because just about every conversation in this domain does.

The Memo

On the IBM accountability slide and what it gets wrong about AI decision-making.

There’s a slide from a 1979 IBM internal training presentation that has, over the past decade, become the most-quoted artifact in the field of AI ethics. It reads:

A computer can never be held accountable

Therefore a computer must never make a management decision

It’s a beautiful line. Hand-lettered, in the way of slides made before PowerPoint existed, which gives it the unearned credibility of an artifact. The kind of thing you screenshot and put at the top of an op-ed. And on inspection, doing a lot less work than the people who quote it want it to.

Let me give the slide its due. The impulse it captures, that moral responsibility shouldn’t be casually outsourced to artifacts, is correct. Any framework treating accountability as fungible, as something you can hand off to whatever non-human system happens to be in the room, will produce bad outcomes by construction. I’ll defend that proposition.

But the inference the slide draws isn’t an argument. It’s a tautology with delusions of grandeur. Because computers can’t be held accountable, they must not make decisions. Set it out that way and the trick is plain: the slide quietly defines “decision” as “accountable choice,” then announces with great solemnity that entities incapable of accountability can’t make accountable choices. The conclusion is smuggled into the premise so neatly the move is easy to miss. This is what philosophers call begging the question and what lawyers call assuming the conclusion. In either tradition, it isn’t the kind of reasoning that should carry forty-seven years of citational weight.

The slide is quoted in 2026 by two camps that think they disagree. The doomer camp invokes it to argue we shouldn’t build systems that make decisions, because doing so would cross the moral Rubicon that the slide draws. The builder camp invokes it to argue that the law shouldn’t touch their products, because the human is always “in the loop.” The model is just a tool, the argument goes, and tools don’t make decisions; decisions are made by the humans using them. These positions sound opposite. They’re mirror images of the same evasion. Both treat the slide as authoritative on the only question that matters, which it doesn’t answer: how do you establish accountability for systems already deployed and already producing consequential outputs at scale?

The honest answer is that you do the work. You build the doctrine, the architecture, and the institutional habits that let accountability flow somewhere legible. You don’t get to skip the work by quoting a training slide from the Carter administration as if it were a statute. Nostalgia is fine. It’s a poor substitute for legal doctrine.

The Frame Is the Problem

On AI regulation, innovation, and the false binary distorting AI policy.

The first thing that has to go is the frame that has captured the AI policy conversation so completely that most people no longer notice it’s a frame: regulation versus innovation. Anyone who reaches for that phrase is either confused or selling something. The relationship between legal architecture and what markets can actually do is more interesting than “more of one means less of the other,” and pretending otherwise is how you end up with bad law written by people who think rules are friction. Every innovative technology I’ve watched come up in the zeitgeist has faced the same hand-wringing: CRISPR, cryptocurrencies, heck, even rooftop solar.

Here’s the irony of the discourse that almost nobody on either side will say out loud: builders who scream loudest about regulation are operating in markets that exist because of regulation. Health tech is a multi-hundred-billion-dollar industry because HIPAA standardized what could move across systems and what couldn’t, and that standardization made interoperability profitable to build. E-commerce became an industry because PCI-DSS gave merchants a defensible answer to “what happens when someone steals card data,” shifting cardholders from suspicion to indifference and driving online spending. The entire SOC 2 and ISO 27001 ecosystem, on which most B2B software companies now stake their enterprise sales motion, exists because Sarbanes-Oxley made auditable internal controls a thing your customer’s general counsel could ask for and expect a real answer to. (Anyone who has been through a SOC 2 audit knows this in their bones, also their stomach lining.) FDA approval is what allows pharma companies to charge the prices they do for their products. Strip the regulation out, and the market disappears. The regulation is the market. Saying this out loud is supposed to be controversial. It isn’t. It’s just historically literate.

None of this is a denial that bad regulation is real. Bad regulation is extremely real, and most of what’s being written about AI right now is bad regulation. There are state-level AI statutes that will be uneconomic to comply with and at least a few that are unconstitutional on their face. (I’m not naming names…but y’all know which ones.) There are provisions in major regulatory frameworks that read as if they were drafted by people who’d read about transformers in a magazine. The cost of writing rules badly is not theoretical; lawyers who minimize it are doing their clients dirty. But the answer to bad regulation is good regulation, written by people who understand both the technology and the doctrine. The answer is not the suggestion (usually advanced by people with substantial financial interest in its acceptance) that AI is somehow categorically different from every other domain in which markets and law have learned to coexist.

There are three things people are doing right now when they invoke prospective AI regulation, and the lawyer’s job is to tell you which is which. The first is genuine engagement: someone has read the proposed rule, identified what it would actually require, and is reasoning publicly about whether it improves matters. This is rare. (Not gonna lie, just the reading part is getting rarer.) The second is competitive weaponization: incumbents who can absorb the compliance cost advocating for regulation that smaller competitors can’t, dressed up as safety advocacy. This is common. The third is fundraising rhetoric: prospective regulation cited as evidence either that the apocalypse is upon us or that the regulatory state is coming for innovation, in either case used to drive donations and engagement to organizations that would lose their reason for existing if the conversation got more boring. This is also common on both sides.

You don’t have to be a cynic to notice that some of the most-quoted arguments for heavy AI regulation are advanced by parties who would benefit enormously from a regulatory regime that licenses development to a small number of incumbents. This doesn’t mean the underlying risks are fake. It means the proposed solutions are not neutral. A safety argument that happens to entrench the speaker’s market position is, at minimum, an argument that deserves a second look from someone whose interests don’t align with the speaker’s. The lawyer’s job is to be that someone.

The work, then, is to write good rules. (Bing bong! Just write good rules. Easy peasy lemon squeezy.) Rules that price the actual risks. Rules that scale to the actual capabilities. Rules that give builders something to comply with and counterparties something to rely on. That work is harder than the slogans on either side suggest, and it gets hardest at the edges, where the cases are unfamiliar and the doctrine hasn’t yet bent to fit them.

The Hard Cases

A four-move framework for AI liability: capability-conditioned scienter, architectural traceability, agency-doctrine extension, and pooled compensation.

This is where the memo’s defenders retreat when the rest of the argument fails. Yes, fine, they’ll allow that deployers can be held responsible for foreseeable model behaviors. But the hard cases are different. Emergent capabilities nobody predicted. Causation chains are so distributed that scienter dilutes to nothing. Agents acting in the world based on inputs that bear only stochastic relationships to what their designers intended. Those are the cases, the argument runs, where accountability genuinely cannot be assigned, and if it can’t be assigned there, the systems shouldn’t exist.

Scienter

Scienter is a legal term that refers to the intent or knowledge of wrongdoing. It is a mental state that indicates that a person acted with knowledge of the illegality or with reckless disregard for the truth. Scienter is often required to establish liability in fraud cases and other legal contexts where intent is a crucial element.

This is the original memo’s trick run one level deeper. Define liability narrowly enough that anything outside the definition becomes inadmissible, then point to the things outside the definition and announce that they prove the system is illegitimate. The trick is more sophisticated this time (the hard cases really are hard, and pretending otherwise is its own bad faith), but it’s still a trick.

The law has handled this class of problem before, repeatedly, and never by banning the underlying activity. Industrial machinery in the late nineteenth century produced harms whose causation chains were too distributed for traditional negligence to reach; the response was workers’ compensation. Ultrahazardous activities like explosives, blasting, and certain chemicals produced risks whose foreseeability was too coarse for fault-based liability to price; the response was strict liability. Vaccines produced rare adverse reactions whose individual causation was inscrutable but whose statistical incidence was real; the response was no-fault compensation funds. Environmental contamination produced harms across temporal and party scales that no plaintiff could plausibly trace; the response was CERCLA. Each is a doctrinal innovation that priced residual risk and forced the underlying activity to become legible to oversight. None of them resulted in factories being closed.

The same kind of work is available for AI, and it isn’t even particularly exotic. There are four moves, in roughly the order they need to be made.

The first is capability-conditioned scienter. Liability scales with what the deployer knew or should have known about the model’s capabilities at the time of deployment. Documented capabilities (in published evaluations like METR’s autonomous task elicitations, in the dangerous-capability assessments labs publish under their Responsible Scaling Policies and Preparedness Frameworks, in deployment incidents, in the model and system cards every serious lab now ships) start the clock the moment they’re reasonably available. Capabilities that genuinely emerge in production and surprise the field create a defense, but only the first time. After the first incident, the capability is in the literature, and the next deployer can’t plead ignorance. This rewards labs that publish what their models can do and punishes the ones that ship blind, which is exactly the incentive structure you want.

The second is architectural traceability as a duty. The system has to be built such that downstream attribution is structurally possible: provenance attestation through frameworks like C2PA; audit logs of tool calls and reasoning traces; agent identity bound to cryptographic credentials; and capability documentation preserved alongside the system prompt, all in a form that can be subpoenaed. Failure to make the system legible is itself a breach, separate from any underlying harm. The harm need not be foreseen. The unobservability has to be. If you operate a system whose failures will cause harm at scale, you operate it in a way that allows those failures to be investigated. The lawyer who tells you otherwise is mistaken and honestly, probably just being lazy. Give me an afternoon on the farm with them.

The third is agency-doctrine extension. For autonomous agents (systems that act in the world on behalf of a principal, hold authorizations, transact, contract, cause consequences), the right legal frame is neither tort nor product liability. It’s agency law, which has spent several centuries working out the questions AI agents raise: what authority the principal granted, what scope was reasonable, what happens when the agent exceeds it, and what notice is owed to counterparties. The implementation is new. Authorizations now live in MCP server permissions, OAuth scopes for tool calls, and the structured capability grants of agent frameworks, rather than in written instructions. But the doctrine bends, as it always has. The Restatement (Third) of Agency is sitting on the shelf waiting, and the deployer who argues that an agent acting within the authority granted is somehow not the deployer’s agent for liability purposes is making an argument that has been rejected since the law of master and servant.

The fourth is pooled compensation for true residual harm. After scienter has been calibrated to capability, after architecture has been required to be legible, and after agency law has addressed the scope of authority, there will still be harms that survive every framework. For those, you do what we already do for nuclear, aviation, vaccines, and oil spills: insurance pools, bonding requirements, and no-fault compensation funds financed by the deployers of the highest-risk systems. This isn’t radical. It’s how dangerous activities operate at scale, by pricing their externalities rather than wishing them away. Builders should welcome it, and most won’t, because they haven’t thought through the alternative: having no answer at all when the harm finally arrives, and watching legislatures write something worse. Anyone who has watched a state legislature write a privacy law knows what “something worse” looks like.

What ties the four moves together is the proposition that the IBM slide couldn’t see, because it wasn’t asking the right question.

Accountability isn’t a metaphysical property of an agent. It’s a structural property of the system in which the agent is embedded.

You can’t hold a model accountable in the sense that the slide demanded. You can absolutely hold the system that includes the model accountable: through scienter scaled to capability, through architecture made legible, through agency law applied to the relationships the system creates, through compensation pooled across the deployers who benefit from the activity. None of that is metaphysics. It’s just the law. The slide’s defenders refused the question. The work is to answer it.

The Architecture of Trust

In the technical architecture, AI builders should implement evaluations, provenance, agent identity, bounded authority, and governance artifacts.

Answering the question is mostly an engineering problem, not a legal one (which, to be fair, the legal profession has historically handled poorly with every new technology). The architecture that satisfies the four moves above is the same architecture that makes AI systems trustworthy in the technical sense, and most of it consists of choices builders are already making, sometimes well, sometimes badly, almost always without thinking of them as legal infrastructure. They are a legal infrastructure. They’re also the difference between systems that scale and systems that become someone’s cautionary tale.

For capability-conditioned scienter to work, builders need to be running and publishing serious evaluations. Model cards and system cards that document not just benchmark scores but failure modes. Dangerous-capability assessments of the kind METR runs for autonomous task completion, or Apollo Research runs for deceptive alignment. Red-teaming exercises that probe what the system can do that it shouldn’t, in the deployment context, will actually be wired into the system, with the retrieval tools and persistent memory it will actually have access to. The companies that publish these well are the ones whose lawyers can defend them when something goes wrong. The companies that treat evals as a marketing exercise are writing the trial exhibits that will be used against them. As lawyers, we’re already seeing this happen, and it’s asinine to pretend the problem will get better rather than worse as ill-advised companies push deeper into frontier capabilities with poor leadership at the helm. If you have your head in the sand, it’s time to join the rest of us in the cool air of reality. Welcome back to reality!

For architectural traceability to mean anything, the system has to preserve what it did and why. Provenance frameworks like C2PA Content Credentials and Google’s SynthID provide content with cryptographic markers that survive copying and remixing. Cryptographic agent identity, the kind being worked out at NIST’s NCCoE, gives the actions an autonomous system takes a tractable answer to “who did this”: signed, attested, non-repudiable. Audit logs of agent reasoning, tool calls, and decision points turn an opaque system into one that can be investigated when it fails. Note that this means structured, queryable, chain-of-custody-preserved logs, not whatever the application happened to capture before its log rotation kicked in. The difference matters in litigation. The first time you sit in a deposition where the case turns on whether logs were preserved before they rolled over, audit infrastructure stops looking like overhead. It starts looking like a moat.

For agency doctrine to apply cleanly, the agent’s authority must be implemented architecturally, not merely described in policy. Capability tokens, scoped permissions, sandboxed tool access through Model Context Protocol servers or function-calling interfaces, least-privilege design with no ambient authority: these are how “what is this agent allowed to do” stops being a marketing claim and becomes a technical fact. An agent that can do anything is an agent that can do anything wrong, and a deployer who hasn’t thought through the difference between technical capability and granted authority is going to discover that difference in litigation. I’ve watched this exact conversation play out about a dozen times this year. It begins with “but the agent is just doing what we told it to do” and ends with the realization that nobody actually wrote down what they told it to do.

Underneath all of it: explicit governance artifacts. Model specifications, such as OpenAI’s published Model Spec. The policies underneath Anthropic’s Constitutional AI training. Acceptable use policies. Written statements about what the system should and shouldn’t do, in language that can be audited against. These matter legally because they’re the closest thing AI systems have to a statement of intent. Designer intent that’s documented can be evaluated. The designer’s intent, that’s implicit, is whatever a plaintiff’s lawyer convinces a jury it was.

Kinda an aside here…but also a tremendous point to be made for the future…A study just showed that more law grads are looking to go straight into plaintiff work over BigLaw, so y’all better gird your loins for a generation of lawyers chomping at the bit for a piece of what you’re building in AI. Remember, it just takes one mistake…

These aren’t compliance burdens. They’re the architecture of trust. The builders who treat them as engineering problems get regulatory legibility as a side effect, insurance pricing that reflects their actual risk profile, and enterprise sales conversations that go materially better than those of competitors who can’t answer questions about provenance, authority, and traceability. The builders who treat them as marketing get crushed when the harm arrives. The builders who treat them as cost centers are confessing they don’t understand what they’re building.

This is where I implore you to understand what you’re building or investing in. Just go ahead and pretend I’m ranting about that now.

That’s what the architecture looks like, in practice. The next question, the one most lawyers won’t touch because answering it requires saying things that several powerful constituencies don’t want said, is which parts of the AI economy aren’t building this way, and what that means for the law that will eventually catch up to them.

The Things Lawyers Don’t Say

Three uncomfortable truths about AI clinical decision support, public-company AI disclosures, and AI safety advocacy.

There’s an etiquette in AI law right now that involves not naming things, and the etiquette is doing harm. Some lawyers in this space are reluctant to call out specific failure modes because the parties involved are clients, prospective clients, fellow members of the bar, sources of speaking invitations, and future employers. I understand the etiquette. I’m declining to participate in it. I wore Adidas Sambas under my dress for my debutante ball, so, ya know, now probably isn’t the time to start caring about etiquette. (My nana is officially rolling in her grave right now, so I will stop there.) There are three things specifically that need to be said out loud right now, and this glorious manifesto is dishonest if I stop short of saying them.

The first is what’s happening in clinical decision support. The “learned intermediary” doctrine is the legal framework that allows a manufacturer of a medical product to offload its duty to warn onto the physician, on the theory that physicians have the time, training, and authority to evaluate the product and exercise independent judgment. The doctrine was developed for prescription pharmaceuticals in the 1960s, and it works passably for drugs, where the physician really is making the call. It’s being imported wholesale into AI clinical decision support, where the model is doing the substantive cognitive work and the physician is signing off because today is a Tuesday and the panel is ninety patients deep. The doctrine no longer fits. The physician has stopped functioning as a learned intermediary. They’re functioning as a malpractice insurance policy that the manufacturer purchases by supplying the system. And when the harm arrives, or is arriving, the plaintiff sues the doctor, the carrier pays, and the manufacturer keeps shipping the product that caused the error. This is malpractice as a system design choice. The fix isn’t subtle: capability-conditioned scienter applied to the manufacturer, architectural traceability that makes the model’s recommendation discoverable, and a clear-eyed acknowledgment that the learned-intermediary doctrine has limits the FDA hasn’t yet been willing to draw, even in its own predetermined change control plan guidance for AI/ML-enabled devices.

The second is what’s happening in public-company AI disclosures. Companies are making claims about their models, their capabilities, and their AI-driven revenue that materially overstate what their systems actually do. Some of this is puffery. A great deal isn’t. It’s the kind of forward-looking statement and capability representation that the SEC has been prosecuting as fraud since the Securities Act of 1933, and the Commission’s relative silence so far isn’t because the conduct is permissible. It’s because the bench is still building the technical capacity to read AI claims with the scrutiny it has long applied to financial statements. That capacity is being built. The lawyers helping companies draft AI disclosures right now should be writing them as if a future enforcement action is reading over their shoulder, because one is. The companies that treat “AI” as a marketing adjective in their 10-Ks are writing the materials the Division of Enforcement will cite in 2027 and 2028. AI washing isn’t a new species of fraud. It’s the oldest species of fraud applied to the newest technology, and the legal framework to address it has been on the books for ninety-three years.

The third is what’s happening with “AI safety” advocacy. I’ve gestured at this earlier, but it deserves its own treatment because it’s the move I most often see lawyers in this space refuse to make, even quietly. Some of the most-quoted arguments for heavy AI regulation are advanced by parties who would benefit enormously from a regulatory regime that licenses development to a small number of incumbents. Not all of them. Not most. But some. The arguments aren’t necessarily wrong. The risks they describe aren’t necessarily fake. But the proposed solutions aren’t neutral, and a safety argument that happens to entrench the speaker’s market position is, at minimum, one that warrants a second look from someone whose interests don’t align with the speaker’s. The lawyer who can’t read regulatory positioning when it’s wearing safety advocacy as a costume isn’t being useful to clients who aren’t incumbents, which is to say, most clients. There are doomer arguments worth taking seriously, and there are doomer arguments that are 10-K filings with footnotes. Telling them apart is part of the job.

These three things, the clinical decision support laundry, the disclosure fraud, and the captured safety discourse, aren’t exotic edge cases. They’re happening at scale, right now, in part because the surrounding professional culture has agreed not to name them. The cost of that silence is paid by patients, by retail investors, and by everyone who would prefer that AI develop in a market that isn’t pre-engineered for the benefit of three companies. Naming the things is a small act. It’s the kind of small act that, accumulated across a profession, determines whether the architecture this manifesto has been describing actually gets built.

The Inflection Point

On the two failure modes of AI law and why both cede the architecture to the wrong people.

Whether the architecture gets built is the question of the moment, and the moment isn’t theoretical. The legal architecture for AI is being written right now: by the rules judges hand down in cases nobody is watching, by the regulatory comments nobody bothers to file, by the contract templates that get copied across deals because somebody has to be the first to insist on different language, by the disclosures drafted by exhausted associates at 11 p.m. that become the precedent for the next ten thousand companies that file something similar. The architecture being written this decade is the architecture we’ll live with for the next century. None of that is hypothetical.

There are two failure modes the legal profession is currently exhibiting in response, and they look opposite but are the same. The first is the lawyer who refuses to engage with the technology. They treat AI as somebody else’s vocabulary problem. They decline to learn how the systems actually work because the doctrine they trained on didn’t require it. They cede the architecture to whoever is writing rules with the most certainty. The second is the lawyer who has engaged with the technology but treats accountability as someone else’s job. They serve the builder so completely that the builder’s interests become the limit of legal imagination. They cede the architecture to whoever is most willing to insist that rules shouldn’t apply to them. Both lawyers think they’re doing different work. They’re not. They’re both refusing the same thing, which is the work of holding the technology and the doctrine in mind at the same time, with equal seriousness, when the temptation in either direction is to drop one and call the dropping a virtue.

The lawyer’s tradition has a name for that refusal. It’s malpractice of a different kind. Not the negligence that gets a license suspended. The failure of professional imagination that gets recorded in the law that ends up on the books. The bar admits us to do work that the public can’t do for itself. Refusing to do it because the work is hard, or unfamiliar, or might cost us speaking invitations, isn’t a permissible response.

This is what I mean by the inflection point. Not because AI is the new printing press or the new electricity or whatever overheated metaphor will be on the conference circuit by next year. Because the legal frameworks being chosen right now will set the terms for autonomous systems for decades, and the frameworks are being chosen with or without the lawyers who could make them better. The lawyers who show up matter more than the ones who don’t.

The Work

What good AI counsel looks like, and why the question is whether lawyers will show up.

I started this piece by telling you why I became a lawyer. The story is about someone who took an unpopular case because he believed the legal system depended on defending the case well, regardless of whether anyone thanked him for it. I think we’re in a version of that case now. Not because AI is the defendant in any literal sense (again, say it with me: AI isn’t a person, and the analogy that turns the technology into a victim is the wrong one) but because the work the lawyer was doing in that case is the work that needs doing in this one. Engagement with the unfamiliar. Insistence on procedure. Refusal to let the surrounding culture’s certainty about what the technology is replace the harder question of what it actually is.

That’s the work I took the ridiculously arduous bar to do. I do it this way because I believe the architecture of the law depends on lawyers who can hold both sides of the question, the technology and the doctrine, the builder and the harmed party, the ambition and the accountability, without flinching from either. Refusing to engage with the technology cedes the architecture to people who don’t understand what they’re regulating. Refusing to engage with accountability cedes it to people who don’t believe rules should apply to them. Both refusals are the same refusal. I’m not interested in either one.

The IBM slide says machines can’t make decisions. I disagree, and I’m willing to argue about it. The harder claim, the one I’ll defend, is that lawyers can. We always have been able to. The question is whether we will.

Stevie Cline, Esq. is a legal practitioner in the Artifical Intelligence governance space. She can be found at www.steviecline.com. The views expressed here are her own and not that of her employer.

I buy hay by the ton and cattle by the head, and I can tell you the exact moment a banker’s eyes glaze over.

It happens right around the time you explain that your collateral has hooves. That it appreciates through calving cycles, not quarterly earnings. That your cash flow is seasonal, your revenue is weather-dependent, and your balance sheet includes a line item for veterinary emergencies involving a 1,200-pound animal who decided the fence was more of a suggestion than a rule. [1]

Basically, when I sit across from a loan officer at a traditional bank, trying to finance a cattle operation, I might as well be speaking Elvish.

Which brings me, with genuine enthusiasm, not grievance, to Erebor.

The Lonely Mountain, Relocated

For those who did not spend their adolescence reading J.R.R. Tolkien, or for those who did but have the good sense not to mention it in professional settings, Erebor is the Lonely Mountain, a dwarven kingdom built atop a vast hoard of gold. It is a bank name that signals ambition and a certain kind of imagination, both of which I mean as compliments. [2]

Erebor Bank received its national charter from the Office of the Comptroller of the Currency in February 2026. It’s the first new full-service national bank charter in four years and the first of the current administration. [3] It opened on a Sunday, a deliberate nod to its blockchain settlement capabilities. [4] It launched with $635 million in venture capital backing, a $4 billion valuation, and a client roster drawn from the defense-tech, AI, and cryptocurrency sectors. [5]

The bank’s stated business model is to provide deposit and lending products to businesses and individuals in the technology, payment systems, investment, and defense industries, as well as to virtual currency market participants. [6] In a way, it is a spiritual successor to the late Silicon Valley Bank. (RIP)

The bank is headquartered in Columbus, Ohio.

I want to pause here, because I think this fact deserves more attention than it has gotten, again, not as a criticism, but as an invitation.

Erebor planted its flag in the capital of a state where agriculture generated $15.9 billion in cash receipts in 2022. [7] A state where net farm income has declined nearly 30% over the past two years. [8] A state whose farm families are navigating the worst financial conditions in a generation. And Erebor, a bank founded on the principle that the existing financial system fails innovators with unconventional assets, has not yet built a single product for any of them.

I don’t think this is malice. I think it’s a blind spot. And I think it’s fixable.

First off, I like Palmer Luckey. He has always struck me as someone who is genuinely interested in solving hard problems, not just profitable ones. The entire premise of Anduril was that the defense industrial base was broken and someone needed to rethink it from first principles. That same instinct, the willingness to look at a legacy system and say “this is failing the people it’s supposed to serve”, is exactly what American agriculture needs applied to its banking infrastructure right now.

This essay is a pitch, not a protest. Seriously. I am trying to convince Erebor to do something I believe it is uniquely positioned to do well.

Founders and Farmers

Before I get into the numbers, and there are many, I want to make an argument that I think most people in technology have never considered:

Farmers have more in common with startup founders than with any other category of business owner in the American economy.

This is not a metaphor. It is a structural observation.

A startup founder and a farmer both operate capital-intensive businesses with long time horizons and uncertain returns. They both make large bets on things they cannot fully control. Market timing for one, weather for the other, and regulatory risk for both. They both manage complex supply chains. They both depend on specialized equipment that depreciates in ways the IRS hasn’t fully figured out. They both deal with seasonal or lumpy revenue. They both routinely take on debt to finance operations before revenue materializes. And they both do all of this while being told by people in more conventional industries that they’re crazy.

A first-generation cattle rancher building a herd from scratch is doing essentially the same thing as a first-time founder building a company: deploying capital against a thesis, managing risk across a multi-year time horizon, making irreversible decisions with incomplete information, and operating with the understanding that most of the return, if there is one, will be illiquid and back-loaded. (I wanted to make a manure joke here…but I decided to be *mature* and not take advantage of great comedic timing.)

The farmer’s “product-market fit” is the intersection of genetics, soil, climate, and consumer demand. Her “burn rate” is feed, fuel, and fertilizer. Her “runway” is measured in hay bales and the number of calving seasons she can finance before the herd reaches critical mass. She does not have a pitch deck, but she has a business plan, and it is written in mud and veterinary records and generational knowledge that no accelerator can replicate.

The reason this matters for Erebor is that the bank was built to serve exactly this risk profile. Like Silicon Valley Bank and it’s robust winery lending operations before, Erebor’s entire thesis is that conventional lenders fail innovators who operate outside standard underwriting models, people whose collateral is unconventional, whose revenue is nonlinear, and whose businesses require a banker who actually understands the underlying operations. That description fits a precision agriculture operation or a registered livestock program at least as well as it fits a crypto fund or a defense startup.

The farmer is a founder. She just doesn’t have a Founders Fund partner to introduce her to the right bank.

The Bank That Disappeared

Before we talk about what Erebor could build, we need to talk about what farmers lost. Because the crisis in agricultural finance is not just about falling commodity prices or rising input costs. It is about the systematic disappearance of the institutions that once made farm lending possible.

In the 1980s, approximately 15,000 community banks served local communities across the United States. Today, there are fewer than 5,000. [9] From 2000 to 2020, the total number of FDIC-insured institutions fell from 8,315 to 4,277, and more than 2,000 bank branches closed in rural America during that period. [10] Mergers averaged over 400 per year from 1990 to 2016. [11] The pace has not slowed. More than 46% of community banks recently surveyed were planning a merger within the next twelve months. [12]

This matters enormously for agriculture because farming depends on community banks in a way that almost no other industry does. Farmers disproportionately depend on loans from small community banks, which account for 70% of all agricultural lending from commercial banks. [13] These are relationship lenders. Nearly half of ag lenders surveyed by the American Bankers Association say they also farm themselves. [14] They evaluate borrowers not by running a credit score through an algorithm, but by understanding the operation, the soil quality, the herd genetics, and the operator’s track record through good and bad years. They understand that a farm’s balance sheet in March tells you almost nothing about its balance sheet in November, and that this is normal, not alarming.

When these banks get acquired, that knowledge walks out the door. The acquiring bank is typically three to four times larger than the target and often has no agricultural lending expertise. [15] Among institutions that acquire agricultural lenders, 10% exit ag lending entirely upon completion of the merger. [16] The remaining 90% technically continue making farm loans, but “continuing” and “understanding” are not the same thing. A loan officer at a $20 billion regional bank is not going to drive past your field to evaluate your hay crop. (I meannnn…my hay is looking fine this spring, so I wish they would.)

The result is a banking desert, and I mean that term precisely. As of the CFPB’s most recent analysis, the Federal Reserve had identified 1,132 banking deserts in the United States, defined as census tracts with no bank branch within a ten-mile radius. Sixty-five percent were in rural areas. An additional 1,055 potential banking deserts were identified, 81% of them rural. [17] Rural census tracts are ten times more likely to be in a banking desert than urban tracts. [18]

FDIC-insured institutions fell from a peak of over 18,000 in 1986 to just over 5,000 by 2020. Institutions with less than $1 billion in assets, the banks that actually make farm loans, fell from 17,514 to 4,116 during that same period. [19]

The United States once had a robust financial system designed to serve farmers. It was built on community banks that understood agriculture, staffed by lenders who often farmed themselves, and rooted in the communities whose economies depended on the success of local operations. That system has been dismantled, not by a single policy decision, but by three decades of consolidation, regulation, and institutional indifference. What remains is a landscape in which the largest agricultural lender is not a bank at all but a government-sponsored cooperative — the Farm Credit System, which holds 46% of all agricultural debt — and the commercial banking sector’s share of farm lending keeps shrinking. [20]

At this point I will mention that I have no idea where my local Farm Credit office is and after farming in some capacity my whole life I have never interacted with them. I don’t know if that’s a bug or a flex or whatever. Just a fact lol.

Into this landscape arrives Erebor: the most heavily capitalized new bank in a generation, headquartered in one of the country’s most important agricultural states, with explicit expertise in lending against unconventional collateral.

The Farm Economy in Freefall

The numbers are bad for farmers. And they are getting worse. Go Us!

Chapter 12 farm bankruptcies, the legal mechanism specifically designed for family farming operations, rose 46% in 2025, reaching 315 filings nationally. The Midwest alone accounted for 121 of them, a 70% year-over-year increase. [21]

Those are only the filings we can count. Chapter 12 requires that the majority of a family’s income come from farming. As off-farm income has become essential to survival, health insurance alone will do that; many families in financial distress no longer qualify for the protection Congress designed for them. They just close. [22] Between 2017 and 2022, more than 140,000 farms shut down, most under 1,000 acres. [23]

Total U.S. farm debt is projected to hit a record $624.7 billion in 2026, driven not by expansion but by operating loans taken out just to cover input costs. [24] Interest expenses are expected to reach a record $33 billion. [25] Fewer than half of farm borrowers are expected to remain profitable this year. [26] Operating loan volumes jumped more than 30% for three consecutive quarters in 2024, with smaller and mid-tier lenders absorbing 75% of the increase, concentrating risk in the community banks least equipped to bear it. [27] In case you are wondering…these numbers are why off-farm income is now so important. My job keeps my animals fed and my crops planted.

Tariffs have added $100 per ton to some fertilizer costs. [28] USDA staffing cuts have left fewer people to help farmers access federal programs. [29] The war in Iran disrupted shipping through the Strait of Hormuz, spiking natural gas prices critical to fertilizer production. [30] Ohio experienced its warmest June and July in 131 years and its driest August on record. [31] National cash corn prices fell for the fourth consecutive year. [32]

And the lenders who remain are tightening. Two out of five ag lenders plan to tighten underwriting for farmland-secured loans. More than half plan to do the same for production loans. [33] Credit quality deterioration has become the number one concern facing agricultural lending institutions. [34]

This is the state and the moment where Erebor chose to build. Pickaway County, where Palmer Luckey’s Anduril just opened Arsenal-1 and began producing autonomous combat drones, is cattle-and-crop country. [35] A reporter who toured the facility described it as “surrounded by cornfields and horse farms.” [36]

The neighbors could use a bank that understands weird situations and sees that they’re not lost causes or sources of immense price-gauging levels of interest that will bankrupt them off-season. (Because…ya know…people need to eat.)

The Collateral Problem: That Erebor Already Solved

Here is what excites me about Erebor’s potential: the bank has already built the exact underwriting infrastructure that American agriculture desperately needs. The engineering is done. The framework exists. It just hasn’t been pointed in this direction yet.

Erebor’s entire value proposition rests on its ability to evaluate and lend against assets that traditional banks don’t understand. Cryptocurrency. Private securities. Precision machine tools. GPU clusters. The bank exists because its founders recognized that conventional lenders were choking off capital to an entire sector of the economy by refusing to underwrite collateral they couldn’t assess with existing models. [37]

That diagnosis is correct. It is also a perfect description of what has been happening to American farmers for forty years.

A registered Polled Hereford cow is a living, appreciating asset with a traceable genetic lineage, verified performance data that includes: birth weight, weaning weight, yearling weight, carcass quality, and a market value that can be precisely benchmarked against national sale averages. My first registered Polled Hereford purchase, a heifer named Nora, came with expected progeny differences for seventeen traits, genomic-enhanced EPDs, bio-economic index scores, breeding soundness data, and health records. [38] She has a BII$ of 624, meaning she is projected to produce more lifetime economic value per dollar invested than roughly seventy percent of registered Herefords in the country. (LFG Nora!!!!)

Nora has better performance data than most crypto tokens. She also produces something. [39] She is the OG utility token y’all.

The blockchain infrastructure Erebor is building: 24/7 settlement, transparent ledger, real-time transaction verification; maps almost perfectly onto the traceability demands the modern beef supply chain is already moving toward. Breed registries are databases. Performance records are data. Cattle sales are transactions that settle in defined cycles. If Erebor can put a crypto portfolio on-chain, it can put a herd inventory on-chain. The American Hereford Association has been maintaining a national herd registry since 1881. They had blockchain before blockchain had a name. They just called it a “herd book.”

The AgTech Opportunity

Set aside the family farm for a moment and consider the startup play in case you haven’t already been convinced that this is a great idea…

Agricultural technology sits at the intersection of every vertical Erebor already serves: AI, advanced manufacturing, defense-adjacent supply chain security, and frontier hardware. The global agrifoodtech market reached $16 billion in funding in 2024. [40] Y Combinator is backing autonomous cattle-herding drones. [41] Venture-funded startups are building robotic weeders, implantable livestock biometric sensors, AI-powered crop monitoring, and blockchain commodity exchanges. [42]

These companies have the same banking problem Erebor was created to solve. They are technically sophisticated, venture-backed, hardware-intensive firms with complex capital needs, seasonal revenue, and collateral that makes traditional lenders nervous. The autonomous tractor startup and the autonomous weapons platform have more in common, structurally, than either has with a SaaS company.

Erebor’s own investor base funds these companies. Andreessen Horowitz has agtech in its portfolio. Lux Capital invests in frontier science. The Founders Fund thesis is that technology should transform the physical world. [43] Feeding people is about as physical-world as it gets.

Palmer Luckey has said, in various forms, that Anduril exists because the defense sector needed someone willing to rethink manufacturing from first principles. The ag sector needs the same thing applied to finance. The person who looked at the defense industrial base and said “we can build this better” is exactly the person who should look at rural banking and reach the same conclusion.

The Strategic Case

Senator Elizabeth Warren’s probe into Erebor’s charter approval raised questions about whether the bank’s rapid approval was influenced by its founders’ political connections. [44] I take no position on that inquiry. It is ongoing. But I will observe that the strongest possible answer to the question “is this bank a favor to Silicon Valley billionaires?” is a client roster that includes Ohio farmers.

Treasury Secretary Scott Bessent called Erebor an important milestone in creating “safe and sound banking choices for all Americans.” [45] OCC Comptroller Gould said it reflected a commitment to a “dynamic and diverse financial system.” [46]

Those are beautiful words. Here is a way to make them true: build products for the largest economic sector in the state where you are chartered.

Columbus is not a mailing address. It is the capital of a state with over 75,000 farms. And again…Pickaway County, where Arsenal-1 is now assembling combat drones, is almost entirely farmland. The big annual event in Pickaway County is a PUMPKIN SHOW. That’s how rural we are talking. [47] The families who own the fields surrounding that factory are the literal neighbors of Erebor’s most prominent founder’s most prominent project.

Banking has always carried civic obligations alongside commercial ones. The Community Reinvestment Act exists because Congress recognized that a bank charter is a public trust, not merely a business license. Serving Ohio agriculture wouldn’t just be good business for Erebor. It would be the most effective legitimacy-building strategy available — and it would happen to help people who need it.

The Product Roadmap

This is not a fantasy. (Well, I mean right now it’s all in my head…but this could get real.) The roadmap writes itself, and most of it is just existing Erebor products with different labels.

Operating lines with ag-specific underwriting: Use the existing alternative-collateral framework to build lending products that accept livestock inventories, equipment, and crop projections as security. If your system can price a basket of altcoins hourly, it can price a herd of registered cattle with verified EPDs weekly. The cattle, for what it’s worth, are less volatile. [48]

AgTech startup banking: Extend the SVB-replacement thesis to agricultural technology companies. These are venture-backed firms with complex capital needs, novel hardware, and founders who are tired of explaining their business model to banks that have never seen a field sensor.

Blockchain-enabled supply chain finance: Use distributed ledger capabilities for transparent, real-time commodity transaction settlement. GrainChain and others are building blockchain commodity exchanges, but no nationally chartered bank offers it as an integrated banking product. [49] Erebor could be the first.

Seasonal cash management: Build deposit and treasury products for businesses whose revenue arrives in concentrated periods, harvest, sale season, calving cycles, rather than assuming monthly recurring revenue. A bank that settles transactions on Sundays can certainly manage a deposit account that peaks in October.

Equipment financing for precision ag: GPS-guided tractors, drone platforms, sensor networks, robotic systems. Capital equipment with quantifiable ROI and strong residual values. Finance it the same way you finance GPU clusters, because functionally, it is the same product. Except the GPU cluster doesn’t also plant your soybeans.

None of this requires Erebor to abandon its tech focus. All of it strengthens it. Agriculture is technology. The farmer running a 2,000-acre row crop operation in western Ohio is managing a more complex hardware-software integration than most Series B startups. The rancher genomically testing herd sires is doing applied biotech. These are founders. They just haven’t been recognized as such by the financial system that is supposed to serve them.

The Mountain Meets the Field

There is a passage in Tolkien, and I kindly invoke it with affection, because naming your bank after a dwarven fortress is the kind of bold, slightly nerdy move I genuinely respect, where the inhabitants of the Lonely Mountain learn that the gold beneath them has value only in relation to the communities around them. The mountain cannot feed itself. It depends on the field. [50]

I believe Palmer Luckey and the Erebor team set out to build something genuinely new: a bank that serves innovators who the existing system has failed. That mission is important. It is also incomplete.

The most innovative, most capital-starved, most technologically underserved operators in Ohio are not all writing code. Some of them are calving heifers at 2 a.m. in February. They are managing multi-million-dollar enterprises on margins thinner than anything a venture-backed startup would tolerate. They are deploying capital against uncertain theses, iterating across seasons instead of sprints, and building businesses they hope to hand to the next generation, which is a longer time horizon than any VC has ever contemplated.

They are founders. And they are right next door.

Erebor chose Ohio. It has the capital. It has the technology. It has the underwriting infrastructure. And it has, in Palmer Luckey, a founder whose career has been built on the conviction that someone should look at a broken system and build something better.

The agricultural banking system is broken. The farmers are right there. The technology already exists.

The mountain has gold. The field has everything else. A bank worthy of its name would serve them both.

Stevie Cline, Esq. is the General Counsel at a Series C stage AI lab. She is also the owner/operator of her family’s farm. She can be found at www.steviecline.com. The views expressed here are her own.

Sources

[1] The author’s senior retired guard dog has also generated veterinary invoices that would make a crypto fund manager wince, but he is not collateral. He is a very good boy and his contribution to the farm is morale.

[2] Tolkien scholars will note that Erebor’s great challenge was not acquiring gold but learning to share it with the communities that made the gold worth having. This is not a criticism. It is, perhaps, a business plan.

[3] Gabrielle Saulsbery, “Erebor Bank Receives National Bank Charter,” Banking Dive, February 9, 2026.

[4] The Wall Street Journal, as reported by Banking Dive, February 9, 2026.

[5] “Silicon Valley’s New Bank Is a Columbus Startup,” Ohio Tech News, February 16, 2026. Investors include Andreessen Horowitz, Lux Capital, 8VC, Peter Thiel’s Founders Fund, and Palantir co-founder Joe Lonsdale.

[6] Federal Deposit Insurance Corporation, “FDIC Approves the Deposit Insurance Application for Erebor Bank, N.A., Columbus, Ohio,” press release, December 16, 2025.

[7] U.S. Department of Agriculture, Economic Impact of Agriculture, Ohio state profile, 2022 data.

[8] Ohio Farm Bureau, “The Ohio Agriculture and Rural Communities 2025 Action Plan,” February 2025.

[9] “What Is Happening to Community Banks and Agricultural Lenders?,” Mesch Clark & Rothschild, P.C., May 1, 2024.

[10] “Trump’s Latest Executive Order Puts Rural Community Banks at Risk,” Newsweek, March 19, 2025.

[11] Consumer Financial Protection Bureau, “Data Spotlight: Challenges in Rural Banking Access,” April 2022.

[12] Mesch Clark & Rothschild, P.C., May 2024.

[13] CFPB, “Data Spotlight: Challenges in Rural Banking Access,” April 2022; CFPB, “New Effort Focused on Financial Issues Facing Rural Communities,” October 2024.

[14] Tyler Mondres, “Agricultural Lenders’ Top Concerns for 2025,” ABA Banking Journal, December 23, 2024.

[15] Mesch Clark & Rothschild, P.C., May 2024.

[16] Id.

[17] CFPB, “Data Spotlight: Challenges in Rural Banking Access,” April 2022, citing Federal Reserve Bank of St. Louis data.

[18] Id.

[19] Id.

[20] Gerald Mashange, “The Lending Activity and Performance of the Farm Credit System and Community Banks,” farmdoc daily, University of Illinois at Urbana-Champaign, January 28, 2026.

[21] American Farm Bureau Federation, “Farm Bankruptcies Continued to Climb in 2025,” Market Intel, 2026.

[22] Id.

[23] Hoosier Environmental Council, “Securing the Future of American Farming,” December 10, 2025.

[24] AFBF Market Intel, 2026.

[25] Id.

[26] American Bankers Association and Federal Agricultural Mortgage Corporation, “Ag Lenders Signal Cautious Outlook for Farm Profitability in 2025,” joint survey, 2025.

[27] “2025 Farm Debt Surge: What Producers Should Know,” Conterra Ag Capital, June 2025.

[28] “U.S. Farmers Are in Crisis. The Iran War Is Not Helping,” Time, March 19, 2026.

[29] Id.

[30] Id.

[31] “Ohio State Experts Warn: Farm Income Growth Masks Deeper Challenges,” Ohio Business Review, September 27, 2025.

[32] “2025 Year in Review: Record Yields, Trade Turmoil and the Future of Ohio Agriculture,” Farm and Dairy, December 23, 2025.

[33] ABA/Farmer Mac survey, 2024, as reported in ABA Banking Journal, December 2024.

[34] Id.

[35] Anduril Industries, press release, January 16, 2025; “Inside Arsenal-1,” Ohio Tech News, March 2026.

[36] “Anduril FURY Combat Drone Production Starts at Ohio’s Arsenal-1,” DroneXL, March 22, 2026.

[37] Ohio Tech News, February 16, 2026.

[38] The author purchased her first registered Hereford at the Ohio Beef Expo in March 2026. Nora’s documentation includes registration papers, five-generation pedigree, expected progeny differences for seventeen traits, genomic-enhanced EPDs, bio-economic index scores (BII$ 624, CHB$ 161), breeding soundness data, and health records. The average Series A pitch deck contains twelve slides, at least three of which are aspirational.

[39] Specifically, she produces calves. Which produce more calves. This is a business model with compounding returns and no dilution, which is more than most token issuances can say.

[40] AgFunder, “Global Agrifoodtech Investment Report,” 2024.

[41] Y Combinator agriculture portfolio. Includes GrazeMate (autonomous cattle-herding drones), Bear Flag Robotics (autonomous tractors, acquired by John Deere), and others.

[42] See, e.g., GrainChain (blockchain commodity exchange), Ceres AI (satellite crop monitoring), Nexa (implantable livestock biometrics), Orchard Robotics (AI precision fruit farming). “Top 100 AgTech Startups to Watch in 2026,” Failory, November 2025.

[43] Peter Thiel’s line: “We wanted flying cars, instead we got 140 characters.” Founders Fund’s investment thesis emphasizes transformative technology in the physical world.

[44] Senator Elizabeth Warren (D-Mass.), letters to OCC Comptroller Gould, FDIC Chair Hill, and Palmer Luckey, February 26, 2026, Senate Committee on Banking, Housing, and Urban Affairs.

[45] Treasury Secretary Scott Bessent, post on X, February 2026.

[46] Office of the Comptroller of the Currency, “OCC Announces Conditional Approval for Chartering Erebor Bank,” news release NR 2025–101, October 15, 2025.

[47] Chris Gordon, “Look Inside Anduril’s New Factory as CCA Production Begins,” Air & Space Forces Magazine, March 2026.

[48] Ethereum’s annualized volatility over the past twelve months: approximately 60–80%. Feeder steer price volatility on the CME: approximately 15–20%. The cattle are also edible, which gives them a price floor that Ethereum lacks.

[49] GrainChain, founded 2013, builds blockchain-based commodity exchanges for agricultural transactions. They are building the plumbing. Erebor could be the bank.

[50] J.R.R. Tolkien, The Hobbit (London: George Allen & Unwin, 1937). The dwarves of Erebor learn that the Lonely Mountain thrives only through trade with the communities of Dale and Lake-town. It’s a children’s book. It is also, apparently, a banking policy white paper.

In 1999, two Boston College researchers named Paul Schervish and John Havens published a prediction that set the nonprofit and financial services world on fire. Between 1998 and 2052, they said, at least $41 trillion would pass from one generation to the next, and charities would be swimming in bequests by 2017.

The number was enormous. Financial firms used it to sell products. Nonprofits built planned-giving teams around it. Policymakers cited it. It had the gravitational pull of a number too big to ignore and too optimistic to question.

By 2006, the Chronicle of Philanthropy ran a follow-up. The headline was grim. Charitable bequests had reached only about 7.5% of what Schervish and Havens projected would arrive by 2017. To hit their target, charitable bequests would have needed to average six times the highest annual amount ever recorded, every single year for thirteen years. Fundraisers were, in the Chronicle’s word, “disappointed.” That’s a polite way of saying they’d been had.

Now we’re doing it again. Same story, bigger numbers.

The Number Everyone Loves to Cite

The current version of the Great Wealth Transfer narrative goes like this: Baby Boomers, the wealthiest generation in American history, hold over half the nation’s household wealth. As they age and die, an estimated $84 trillion to $124 trillion will pass to their heirs and to charity over the next two decades. The figure comes primarily from Cerulli Associates, a financial research firm, which projects $124 trillion in total transfers through 2048, $105 trillion to heirs and $18 trillion to charity.

Read those numbers in a Fortune headline or a Merrill Lynch white paper, and it sounds like a rising tide that will lift all boats. A generation crushed by student debt, stagnant wages, and unaffordable housing is about to be rescued by the simple passage of time and the generosity of their parents’ mortality.

This framing is, at best, misleading. At worst, it’s a lie that serves everyone except the people it claims to help.

Follow the Money Because It Doesn’t Go Where You Think

Start with the single most important fact about the Great Wealth Transfer, the one that gets buried in every breathless headline: more than half of the total volume of transfers, $62 trillion of that $124 trillion, is expected to come from high-net-worth and ultra-high-net-worth households. That’s the top 2% of all American households.

Cerulli’s own data confirms that 68% of wealth transferred between 2020 and 2045 will come from households with at least $1 million in investable assets, a category that includes just 6.9% of all households. The wealthiest 1.5% of households alone account for 42% of expected transfers, roughly $35.8 trillion.

So when someone tells you “$124 trillion is about to change hands,” what they are actually describing is the already-wealthy passing their wealth to their already-positioned heirs. They’re describing the perpetuation of concentrated wealth, not its democratization. The “Great Wealth Transfer” is, for the vast majority of Americans, someone else’s money moving into someone else’s account.

For the median American household, the one that actually exists at the center of the distribution, the picture is radically different. Federal Reserve data shows the average inheritance across all households is approximately $46,200. But even that number is wildly misleading, because it’s dragged upward by the enormous bequests at the top. The median inheritance, the one a typical recipient actually receives, is $69,000. For households in the bottom half of the wealth distribution, it’s closer to $9,700. And between 70% and 80% of American households never receive any inheritance at all.

Read that again. The majority of American households will inherit nothing. Zero. And for those who do inherit, half will receive less than $10,000. Someone is lying to someone.

That’s the Great Wealth Transfer. That’s the $124 trillion number stripped of its marketing copy.

Rich on Paper, Broke in Practice

But wait, don’t Boomers have all this wealth? They do, on paper. The problem is where it sits.

Boomers control an astounding $19 trillion in housing wealth, nearly 40% of all residential real estate in the United States. That’s the number people see on Zillow or Redfin and mentally convert into an inheritance. But home equity is not a liquid asset. You can’t write a check against it. You can’t split it between three siblings without selling the house. And selling the house, after agent commissions, closing costs, deferred maintenance on a home that may not have been updated since 1987, and months on a softening market, yields considerably less than the number on the screen.

Meanwhile, the liquid wealth picture is grim. The median Boomer household reports about $289,000 in retirement savings which is waaaaaaay less than what financial planners say a single person needs to fund their own retirement. Only 20% of Boomers feel they’re on track with retirement savings. Two-thirds say they’re behind. With numbers like that, no shit.

And here’s the part no one talks about: Boomers are drowning in debt. According to Experian, the total debt carried by all Baby Boomers at the end of 2024 was $4.5 trillion. The average debt per Boomer is roughly $94,880. That includes an average mortgage balance of around $188,000, average credit card debt of roughly $6,800, and average personal loan debt over $22,000. The share of households headed by someone 75 and older carrying debt rose from 32% in 1992 to 53% in 2022. The share of households 65 and over with mortgage debt has nearly doubled over the past two decades.

So when a Millennial looks at their parents’ $400,000 house on Zillow and mentally earmarks half of that as their future inheritance, here’s what they’re not seeing: the $188,000 still owed on the mortgage, the $25,000 in credit card and personal loan debt, the property taxes and insurance that have been quietly devouring the monthly budget. That $400,000 house has maybe $190,000 in equity, before selling costs, before any reverse mortgage, before Medicaid estate recovery if their parents needed long-term care. The Zillow number is a fantasy. The net number is the reality. And the net number, for a startling number of Boomer households, is close to zero.

In Summary…Your parents’ house is not your retirement plan.

The Healthcare Black Hole

Even the modest wealth that middle-class Boomers have managed to accumulate faces a voracious adversary: the American healthcare system.

Roughly half of Americans who reach age 65 will require some form of long-term care. The monthly cost of full-time care ranges from $4,600 to $9,000, depending on type and intensity. Medicare, the program most people assume will cover them, provides minimal coverage for long-term care and no coverage at all for custodial care like assistance with daily activities or dementia supervision. Medicaid covers long-term care only for people who have already spent down nearly all of their assets.

The math is ruthless. A Boomer with $289,000 in savings who needs three years of assisted living at $6,000 a month will burn through $216,000, roughly 75% of their entire nest egg, on care alone. The inheritance their children were counting on doesn’t vanish in some dramatic financial crisis. It just gets eaten, slowly, by the cost of keeping their parents alive and dignified. As finance writer Annie Logue put it on Marketplace, her husband inherited about $2,000 from his grandmother after her care was paid for. It went toward new windows.

Healthcare cost inflation has consistently outpaced both overall inflation and wage growth for decades. Over a long retirement, this becomes a compounding wealth destroyer. The hoped-for transfer of accumulated wealth from Boomer parents to their Millennial and Gen X children may, for a large swath of America, ultimately end up in the medical system.

The Bank of Mom and Dad Is Already Overdrawn

Here’s the bitter irony that makes the inheritance expectation even more delusional: Boomers aren’t hoarding their wealth waiting to die and hand it over. They’re already spending it…on their adult children.

Half of all parents now provide some form of regular financial assistance to their adult children. The average monthly support per adult child is $1,474, roughly $18,000 a year. Working parents who financially support their grown children are spending 2.3 times more on that support each month than they’re putting into their own retirement accounts: $1,589 to the kid, $673 into the 401(k).

Forty-three percent of parents say helping their adult children has negatively impacted their retirement savings. Again, no shit. Nearly half have put off paying down their own debt. Sixty-one percent of parents with adult children have made some financial sacrifice to help them, and Gen X parents, the sandwich generation now simultaneously supporting aging Boomers and struggling Millennials, are the most likely to be doing it.

So the wealth transfer is happening, it’s just happening now, in $1,474 monthly installments, instead of later in a lump-sum inheritance. And every dollar that goes to rent, groceries, or a grandchild’s car insurance is a dollar subtracted from the estate that was supposed to arrive someday as a windfall. If your parents send you money every month, you are already receiving your inheritance. It just doesn’t come with a death certificate.

The Expectation Gap. And Why It’s Worse Than You Think

Northwestern Mutual’s 2024 Planning & Progress Study found that one-third of Millennials expect to receive an inheritance. But only 22% of Boomers expect to leave one. Of those Millennials who do expect an inheritance, half consider it “critical” or “highly critical” to their financial security.

The gap gets worse the closer you look. Alliant Credit Union found that 52% of Millennials expecting an inheritance believe they’ll receive at least $350,000. But 55% of Boomers who do plan to leave something say it will be less than $250,000. A USA Today Blueprint survey found that 68% of Millennials and Gen Z have received or expect to receive an inheritance, with the average expected amount around $320,000. Among those expecting assets, 62% specifically expect to inherit real estate.

Why the inflated expectations? Three reasons converge.

First, the media narrative itself.

A decade of headlines about “$84 trillion” and “$124 trillion” in wealth transfers has created a cultural assumption that large inheritances are normal and imminent. Financial firms, real estate consultancies, and news outlets have all amplified the story, Coldwell Banker, Morgan Stanley, Fortune, Forbes, creating a pervasive sense that if you have Boomer parents, a windfall is on the way. The number is so big and so ubiquitous that it feels like a promise.

Second…and again, the Zillow problem.

Again, home values have increased nearly 500% since 1983 when Boomers first started buying. Adult children can pull up their parents’ address online and see a six-figure number. What they can’t see is the remaining mortgage, the reverse mortgage lien, the deferred maintenance, or the fact that the house may need to be sold to fund long-term care. As Teresa Ghilarducci noted in the CFA Institute’s analysis, people look at the value of their parents’ houses on property websites and assume that’s the equity — not knowing that older Americans are entering retirement with more mortgage debt than ever.

Third, families don’t talk about it.

Ninety percent of parents who intend to leave an inheritance don’t have a specific plan in place. Forty-seven percent of Boomers who expect to leave an inheritance have never discussed their financial plans with their family. A Wells Fargo report found that 26% of adult children would rather deal with their parents’ estate after they die than have the conversation while they’re living. Into that silence, expectation fills the vacuum. And the expectation, fed by headlines and Zillow screenshots, is almost always too high.

The behavioral consequences are real and measurable. One in twelve younger adults expecting an inheritance say they feel less pressure to earn money because of it. One in ten say they carry more debt than they otherwise would because they plan to use their inheritance to pay it off. The Great Wealth Transfer isn’t just a bad forecast — it’s actively distorting financial behavior among the people who can least afford bad decisions.

The Boomer Asset Mirage

In March 2026, a financial advisory firm called ROLE Advisory published a white paper that introduced a concept worth paying attention to: the “Boomer Asset Mirage.”

The argument is elegant and unsettling. A substantial portion of Boomer wealth is not truly transferable at its stated value because the value of that wealth is itself sustained by the generation that holds it. Baby Boomers are simultaneously the largest owners of American assets and the largest source of consumer demand propping up the value of those assets. When they exit the economy, through death, downsizing, and retirement drawdown, they flood asset markets with supply while withdrawing the demand that supported prices.

Housing is the clearest example. Boomers hold an extraordinary share of American residential real estate. As they die or move to assisted living, those homes enter the market. But the younger generation that’s supposed to buy them has been priced out for years. Newsweek reports that significant home turnover is expected as properties are sold to pay for care, transferred to children, or liquidated to settle estates, but that the resulting supply surge, combined with weakened demand, could reshape regional housing markets and erode the value of the very asset most middle-class families are counting on.

The Great Wealth Transfer, in other words, may partially evaporate in the act of being transferred. The wealth was never as solid as it looked.

Who Benefits From This Story?

If the Great Wealth Transfer is largely a story about the top 2% of households, with a supporting cast of middle-class families whose modest wealth will be consumed by debt, healthcare, and ongoing support of their adult children…why does the narrative persist?

Because it is extraordinarily useful to a lot of powerful interests. The Great Wealth Transfer is trickle-down economics wearing a black armband.

The financial services industry gets to sell wealth management, estate planning, and advisory products to a massive anxious market. Cerulli Associates, the firm whose $124 trillion estimate anchors most media coverage, is literally in the business of selling research reports to wealth managers. The subtitle of their report is “Capturing Money in Motion.” The audience is not you. I’m sorry to burst your bubble, but it was never you.

Media outlets get an irresistible headline. “$124 Trillion Wealth Transfer” is catnip for clicks. The fact that 50% of it comes from 2% of households is less shareable and less palatable for folks.

Policymakers get to avoid hard conversations. If intergenerational wealth transfer will solve the affordability crisis naturally, there’s no urgency to expand Medicare coverage for long-term care, reform the student loan system, or build public housing. The market, through the blessed mechanism of parental death, will take care of everything. This is fiscal policy by magical thinking.

And the nonprofit sector, which deserves its own note here, has been building fundraising infrastructure around projected bequest windfalls for thirty years, with decidedly mixed results. This is the exact same sector that was promised a golden age of bequests in the 1990s and got burned. Now the projected number flowing to nonprofits is $18 trillion through 2048, but much of that will route through private foundations and donor-advised funds, vehicles controlled by the ultra-wealthy, with minimal or no payout requirements. Forty-one cents of every dollar of individual giving in 2022 went to one of these intermediaries, where funds can sit for years or decades before reaching a working charity. Meanwhile, the number of Americans participating in charitable giving at all has been declining for two decades, even as the largest gifts from the wealthiest donors have gotten bigger, masking the erosion of the donor base beneath a few enormous checks.

Small and mid-sized nonprofits, community land trusts, heritage conservation organizations, food banks, rural health clinics, are being told to build planned-giving programs to “capture” wealth transfer dollars. But the dollars that actually materialize in their accounts will be modest, because their donors are not in the top 2%. Their donors are the people whose parents will spend down their savings on healthcare. The Great Wealth Transfer, for these organizations, is a strategy built on someone else’s donors.

The Real Transfer

There is a massive transfer underway, but it’s not the one described in the headlines.

The real transfer is from middle-class Boomer savings accounts to the long-term care industry, to hospitals, to pharmaceutical companies, and to the Medicaid system (which will claw back from estates after the fact). The real transfer is from home equity to the carrying costs of aging, deferred maintenance, property taxes on fixed incomes, reverse mortgage interest. The real transfer is $1,474 a month from parents to adult children, draining the very retirement accounts that were supposed to fund the inheritance.

For the already wealthy, the Great Wealth Transfer is real and it’s lucrative. It will concentrate dynastic wealth further. It will fund family foundations and donor-advised funds and trust vehicles. It will create a generation of inheritors in the top decile who are wealthier than their parents were at the same age. That is genuinely happening, and it will reshape markets and politics in ways we should be paying attention to.

But for everyone else, for the median American family, for the Millennial carrying extra debt because they think a $350,000 check is on the way, for the nonprofit building a strategic plan around projected bequests, the Great Wealth Transfer is a story about other people’s money. And banking your future on other people’s money is not a plan. It’s a prayer.

We’ve heard this prayer before. In the 1990s, it didn’t come true. The only thing that’s changed is the number of zeros.

Stevie Cline, Esq. is the General Counsel at a Series C stage AI lab. She can be found at www.steviecline.com. The views expressed here are her own.

Why the startup failure rate, the founder hiring penalty, and the venture capital funding gap are all part of the same story.

I want to tell you about a door.

It’s a very appealing door. It has inspirational quotes on it. Rocket ships!!!! It has a podcast. It has a Slack community with 40,000 members and an application that asks you to describe your “superpower.” On the other side, you’re told, is a world where your ideas matter more than your résumé, where the rules of corporate America don’t apply, where you can build something that changes the world and gets rich doing it. Doesn’t matter where you came from or who your parents were. MERITOCRACY.

Sounds amazing right!?!

What nobody tells you, what the data proves, and the ecosystem has a financial incentive to suppress, is that for most people who walk through that door, there is no door back.

I’ve spent most of the second act of my career inside the startup ecosystem. I’ve been general counsel for companies that got acquired and companies that cratered. I’ve sat across the table from founders who were brilliant and broke and terrified. I’ve watched people walk through the door. And I’ve watched what happens to the ones who try to come back. I also walked through the door. So those people are not alone. It happens to the best, most optimistic, most naive of us.

The startup ecosystem has a recruitment problem. Not the kind it talks about. The kind it doesn’t.

Part One: Who Actually Succeeds (It’s Not Who You Think)

Before we can talk about the trap, we need to talk about the profile of the founder who actually makes it. Because the data is not ambiguous. It’s just inconvenient for the people who sell the dream.

The Age Nobody Wants to Hear

You know the archetype. The hoodie. The dorm room. Zuckerberg. Gates. The twenty-something dropout who sees what nobody else can see. It’s the founding myth of Silicon Valley, and it is, empirically, a lie.

Now. I will say that there is a VERY select few who do make it. I currently work for a few of those people. But they are the exceptions to this rule because, in my humble opinion, they are pretty awesome.

Mark Zuckerberg said it plainly: “Young people are just smarter.” Peter Thiel put money on it, offering $100,000 fellowships to entrepreneurs, on the condition that they were under 23. The ecosystem didn’t just believe the myth. It underwrote it.

Now that caveats are done, back to the story…

A landmark study from MIT, Harvard, and the U.S. Census Bureau analyzed the founders of 2.7 million companies in the United States. The average age of a founder who builds a top-performing startup is 45. Not 25. Not 30. Forty-five. A 50-year-old founder is 1.8 times more likely to build a high-growth company than a 30-year-old. Founders in their early 20s have the lowest likelihood of success in the entire dataset.

The reason is almost boringly intuitive: industry experience. Founders with at least three years of prior work in their startup’s industry were 85% more likely to succeed. The 45-year-old isn’t winning despite their age. They’re winning because they spent twenty years learning the terrain — the customers, the distribution channels, the regulatory landscape, the things that can kill you that don’t show up in a pitch deck.

Hold that number in your head. Forty-five. We’re going to need it later.

The Education Nobody Wants to Admit

The dropout founder is an outlier, not a template.

Stanford’s Venture Capital Initiative studied over 1,100 U.S. unicorn companies and found that unicorn founders are six times more likely than the general population to hold a doctoral degree, three times more likely to have a master’s, and twice as likely to have completed an undergraduate degree. Stanford, MIT, and Harvard are the top three producer schools. Computer science, engineering, and economics account for 53% of all unicorn founders.

The dropout stories dominate the cultural imagination because they’re interesting, not because they’re representative. Nobody writes a profile in Wired about the 48-year-old PhD who spent 15 years in energy supply-chain logistics before starting a company. But statistically? That’s your unicorn founder.

Timing Trumps Everything

Bill Gross analyzed over 200 companies and ranked the factors that separated success from failure. Timing, launching when the market is ready, when the enabling technology is mature, when customer behavior has shifted accounted for 42% of the difference. More than the team. More than the idea. More than funding.

And here’s the thing about market timing: the ability to read it isn’t luck. It’s a product of deep industry experience and pattern recognition, which means it correlates directly with age and tenure. The people best positioned to time a market are the people who have spent years inside it. The 24-year-old fresh out of an accelerator is, by definition, the person least equipped to judge whether the market is ready.

The Baseline Brutality

Half of all businesses fail within five years. Two-thirds are gone by year ten. That’s not a statistic about bad businesses, it’s all businesses, including the ones with funding and experienced teams and good ideas.

So the empirical profile of the founder most likely to survive this gauntlet is: mid-40s, deep industry experience, elite education, a strong professional network, and good market-timing instincts built on decades of pattern recognition.

Now let’s talk about who the ecosystem is actually recruiting.

Part Two: The Pipeline

The Wrong People for the Wrong Reasons

If you were designing a system to produce successful startups, you would recruit experienced mid-career professionals with deep domain expertise, strong networks, and enough financial cushion to survive the early lean years. You would look for people who match the profile that the data says actually wins.

That is not what the startup ecosystem does. At all.

The ecosystem recruits young. It recruits undercapitalized. It recruits first-time. It recruits people whose primary qualification is enthusiasm and who have been told by accelerator marketing, by startup media, by LinkedIn influencers with “serial entrepreneur” in their bio, that enthusiasm is the qualification that matters.

“Move fast and break things” was a Facebook company motto for years. Gary Vaynerchuk built an empire telling people to “eat dirt for 36 months” and that sleep is for people who lack ambition. Paul Graham, co-founder of Y Combinator, wrote that the “cutoff in investors’ heads is 32,” as though the data on founder success doesn’t exist. Reid Hoffman told the world to “jump off a cliff and assemble the plane on the way down.” These aren’t fringe voices. They’re the architects of the recruitment pipeline for the whole startup ecosystem.

Why? Because that’s who applies. A 45-year-old with a senior engineering role and a mortgage is not going to quit their job to attend a three-month accelerator on a $20,000 stipend. A 25-year-old with nothing to lose will. And accelerators need cohorts. VCs need deal flow. The startup media machine needs stories. The entire ecosystem runs on a steady supply of optimistic, undercapitalized people who haven’t yet learned what the data says about their odds.

This isn’t a conspiracy. (Sorry to all the conspiracy theorists reading…there is no cabal.) It’s a business model. And the business model requires that nobody look too closely at the math.

The Capital Trap

The math gets worse when you look at who actually has access to the resources that predict success.

Economists Ross Levine and Yona Rubinstein found that entrepreneurs disproportionately come from wealthier, more highly educated families. There is no “entrepreneurship gene.” There is a safety net. One in five businesses is funded by inherited money. A similar number are funded directly by the founder’s parents.

This wealth prerequisite is heavily racialized. White families hold approximately 13 times the median wealth of Black families. The ability to self-fund, absorb early losses, or raise a “friends and family” round is vastly unequal. And the formal funding system doesn’t correct for this, it significantly amplifies it. Women, who are 51% of the population, receive roughly 2% of venture capital. Black and Latino founders each receive approximately 1.5%. White male founders capture 85%.

The people the ecosystem most aggressively recruits, young, first-time, undercapitalized, are also the people least likely to have the wealth buffer that makes failure survivable. The system is selecting for fragility.

And when Stanford’s Latino Entrepreneurship Initiative found that Latino entrepreneurs receive full funding only 21% of the time (versus 40% for white entrepreneurs) and get feedback on denials only 51% of the time (versus 87%), what you’re looking at isn’t just a funding gap. It’s an information gap layered on top of a capital gap layered on top of a wealth gap. The people with the least margin for error are also getting the least information about how to improve.

Meanwhile, the children of entrepreneurs are 60% more likely to become entrepreneurs themselves, and the transmission is environmental, not genetic. Role modeling. Network access. Knowledge transfer. The advantages compound across generations, while the ecosystem continues to recruit from the same pool of people who don’t have them.

What the Pipeline Actually Produces

So let’s be precise about what’s happening. The startup ecosystem recruits people whose demographic profile (young, inexperienced), the data says, is least likely to succeed. It funnels them into an endeavor where half will fail within five years, regardless. It systematically undercapitalizes the ones who face the most structural barriers. And calls all of this “opportunity.”

That’s the setup. Now let me show you what happens when they try to come back.

Part Three: The One-Way Door

(Insert a big ole sigh here.) This is the part of the essay I didn’t want to write. Not because it’s controversial, but because it’s personal. I’ve watched it happen to people I care about. It’s happened to me. And the popular narrative, that failure is a “learning experience,” that you should “fail fast and fail forward” is, for a significant percentage of founders, a dangerous fiction told by people who have never experienced the consequences.

The Career Graveyard

Researchers at Yale sent 2,400 fictitious résumés to software engineering jobs, identical in every way except whether the candidate had been a startup founder or a traditional employee. Founders were 43% less likely to receive an interview callback.

And here’s the twist that should make your stomach drop: successful founders were penalized even more heavily than failed ones. The callback rate for non-founders was 24%. For failed founders, 16.2%. For successful founders, 10.9%.

Recruiters explained why: founders are perceived as autonomous, uncommitted, culturally alien, and likely to leave and poach talent. The exact traits the startup ecosystem celebrates, independence, breadth, adaptability, are serious liabilities on a corporate résumé. One recruiter told the researchers that founders “look like aliens to people.”

This is the one-way door. Walk through it, and the path back to stable employment narrows dramatically. And the data says the damage is permanent. A UCLA Anderson study found that startup workers earn approximately $27,000 less over a decade than peers at established firms. (Eepa! ) Eighty-four percent follow a career trajectory that keeps them in the lower-paying small firm ecosystem. The path back isn’t just hard. It’s structurally blocked.

Think about what this means at scale. Every year, accelerators mint thousands of new founders. Most of their companies will fail. When they do, these people enter a job market that penalizes them for the experience the ecosystem told them would be their greatest asset. They’re overqualified for entry-level roles. They’re “aliens” to corporate recruiters. They can’t go back to the career track they left because the track has moved on without them.

Where do they go? Le sigh.

They start another startup. Not because they’ve found a better idea or because the timing is right, but because it’s the only door that’s still open. And the cycle repeats, with less capital, less energy, and less runway than before.

Sooooo…this is what I mean when I say the ecosystem is manufacturing unemployable people. Not intentionally. Not maliciously. But predictably, documentably, at scale.

The Mental Health Epidemic

Quite frankly, the psychological toll of this cycle is staggering.

Dr. Michael Freeman at UCSF studied entrepreneurs and found they are twice as likely to suffer from depression as the general population, six times more likely to have ADHD, three times more likely to struggle with substance abuse, and ten times more likely to have bipolar disorder. Seventy-two percent report that their entrepreneurial journey has negatively affected their mental health. These are NOT tiddly wink numbers y’all.

These numbers are often cited as evidence that entrepreneurs are “wired differently”, as though the mental health crisis is a feature, not a bug. What gets less attention is the causal arrow. Are people with certain psychological profiles drawn to entrepreneurship? Almost certainly. But is the experience of founding a company, the chronic stress, the isolation, the financial terror, the identity fusion where you become the company and its failure becomes yours, making people worse? The data says yes. And are those things good for those entrepreneurs with a mental illness. Definitely not. And the ecosystem has no safety net for them, because acknowledging the mental health toll would undermine the recruitment pitch.

The Relational Fallout

A University of California study of 3,900 married business owners found that nearly one in three entrepreneurs divorced. Which is roughly double the rate for non-founders in the same age bracket. The average founder workweek of 63 hours doesn’t leave much room for presence and connection.

But this isn’t just about time management. It’s about the identity trap. When you’ve staked your professional identity, your financial future, and your sense of self on a venture, every hour spent away from it feels like a betrayal of the mission. The “hustle culture” that glorifies 120-hour workweeks isn’t just bad advice about productivity. It’s an ideology that frames family time as an inefficiency to be optimized away. The data is unambiguous: continuing to view extreme work hours as a badge of honor is statistically volunteering your family as collateral damage.

The Stolen Milestones

For young founders, the ones the ecosystem most aggressively recruits, the years dedicated to a startup overlap directly with the critical window for marriage, parenthood, and homeownership.

In 1975, 45% of adults aged 25–34 had achieved the traditional markers of adulthood — living independently, working, married, with children. By 2024, that number had collapsed to 24%.

Entrepreneurship is an accelerant for this trend. The founder who dedicates their 20s and early 30s to a venture that fails emerges with a decimated balance sheet, having foregone years of salary, retirement contributions, and career progression. While their peers built savings and families, the failed founder starts from scratch, in a job market that punishes them for trying.

Sidenote: One of my biggest pet peeves is folks in the startup ecosystem preaching about marrying young and having all the babies, “DO IT NOW”, they say. As if it’s just that easy, bing bong! All while preaching the great benefits of grind culture. Y’all have got to be kidding me. LOLZZZZZZ.

Elon Musk has bragged about sleeping on the factory floor. Sam Altman once tweeted that “it’s possible to be both a great founder and have a family” but that “if you want to be great, you need to be somewhat obsessed.” The quiet part is always the same: if you failed, you didn’t want it enough. Sure Jan.

The Financial Trap

For founders without a wealth buffer, business failure doesn’t stay in the business. The legal fiction of limited liability dissolves the moment you sign a personal guarantee for a lease, a loan, or a credit line, which most early-stage founders (especially those founders who would be labeled by the industry as “diverse”) do because they have no other option.

A Federal Reserve Bank of Minneapolis study found that at age 25, the average entrepreneur earns less than a paid employee. The popular image of high entrepreneurial earnings is driven by a tiny fraction of outliers. Eighty percent of all self-employment income goes to the top quintile. For the other 80%, entrepreneurship is a path to lower earnings and greater financial exposure.

The financial math for the median failed founder looks like this: years of below-market salary, depleted personal savings, possible personal liability for business debts, a $27,000/decade earnings penalty in subsequent employment, and a job market that doesn’t want them. That’s not a risk/reward calculation. That’s a trap.

The Generation Nobody Is Talking About

Here’s the thing that I can’t stop thinking about, and the reason I wrote this essay.

Somewhere out there right now is a growing population of people in their late 20s and early 30s who were told by Demo Day audiences, clout-maxxers on X, startup podcasts, and LinkedIn posts with rocket ship emojis that founding a company was the highest and best use of their talent. They believed it. They walked through the door. Their companies failed because most companies fail.

And now they’re stuck.

They can’t get corporate jobs because recruiters see them as flight risks and cultural misfits. They can’t go back to their old career tracks because those tracks have moved on. They can’t start another company because they’re out of capital and runway. They’re carrying student debt from the degrees the ecosystem told them didn’t matter. They haven’t bought houses. Some of them haven’t started families. Many of them are struggling with depression or anxiety that the ecosystem told them was just “part of the journey.”

“Failure is just part of the process,” they were told. “Fail fast. Fail forward. The best founders fail three times before they succeed.” Nobody mentioned that each failure makes you 43% less hirable than you were before you started.

They are professionally homeless. And the ecosystem that made them that way has already moved on to the next cohort.

I don’t know what to call this generation. But I know they exist, because I’ve watched the ecosystem produce them for fifteen years. And I know the number is growing, because the accelerator pipeline keeps expanding, the failure rate hasn’t changed, and the one-way door hasn’t gotten any wider.

The startup ecosystem talks constantly about disruption. It doesn’t talk about the people it discards.

What Would Honest Look Like

I’m not writing this to tell people not to start companies. Entrepreneurship is real. Innovation matters. Some people walk through the door and build extraordinary things. I work for a group of smart, mentally-well twenty-somethings building something really cool. I am betting on them succeeding. THEY ARE OUTLIERS.

But I am writing this to say that the current system is not honest about the odds, the costs, or the exit options. And that dishonesty is causing measurable, documented harm to a growing number of people.

Honest would look like accelerators publishing five-year outcome data for their alumni, not just the ones who raised a Series A, but the ones who went back to the job market and couldn’t get callbacks. Honest would look like startup media covering the founder penalty with the same enthusiasm it covers funding rounds. Honesty would look like VCs acknowledging that the demographic profile they recruit for is exactly the one the data says is least likely to succeed, and asking themselves why.

Honesty would mean that before anyone walks through the door, they get the same briefing a surgeon gives before an operation: here is what we’re going to do, here is the probability that it works, here is what happens to you if it doesn’t, and here is what recovery looks like. Informed consent. Not inspiration porn.

The question that should be on every accelerator application, the one that never is, isn’t “What’s your vision for changing the world?” It’s “What is your plan for re-entering the workforce if this doesn’t work, and do you understand the data on how hard that will be?”

Because right now, there are people walking through a one-way door who don’t know it’s one-way. And the people holding it open are getting paid regardless of what happens on the other side.

If you’re on the other side of the door right now, wondering what went wrong, nothing went wrong with you. The math was never in your favor, and nobody who profited from your optimism was ever going to tell you that. This essay is for you. And for next year’s cohort, who still has time to read the fine print.

Stevie Cline, Esq. is the General Counsel at a Series C stage AI lab. She can be found at www.steviecline.com. The views expressed here are her own.

Sources

1. Azoulay, P., Jones, B. F., Kim, J. D., & Miranda, J. (2020). Age and high-growth entrepreneurship. American Economic Review: Insights, 2(1), 65–82.

2. Strebulaev, I. A., & Gornall, W. (2021). The Unicorn Founder Myth: Education Matters. Stanford Graduate School of Business.

3. Gross, B. (2015). The single biggest reason why startups succeed. TED.

4. U.S. Bureau of Labor Statistics, Business Employment Dynamics.

5. Botelho, T. L., & Chang, M. (2023). The Evaluation of Founder Failure and Success by Hiring Firms: A Field Experiment. Organization Science, 34(1), 484–508.

6. Sorenson, O., & Dahl, M. S. (2021). Do Startup Employees Earn More in the Long Run? UCLA Anderson Review.

7. Freeman, M. A., Staudenmaier, P. J., Zisser, M. R., & Andresen, L. A. (2018). The prevalence and co-occurrence of psychiatric conditions among entrepreneurs. Small Business Economics, 53(2), 323–342.

8. Female Founders Fund. (2025). Women in VC & Startup Funding: The 2025 Report.

9. NBER. (2022). Funding Black High-Growth Startups.

10. Stanford Latino Entrepreneurship Initiative. (2023). State of Latino Entrepreneurship.

11. Levine, R., & Rubinstein, Y. (2017). Smart and illicit: Who becomes an entrepreneur and do they earn more? The Quarterly Journal of Economics, 132(2), 963–1018.

12. Pantin, L. P. (2021). The Wealth Gap and the Racial Disparities in the Startup Ecosystem. Saint Louis University Law Journal, 66(1), 59–92.

13. Lindquist, M. J., Sol, J., & Van Praag, M. (2015). Why do entrepreneurial parents have entrepreneurial children? Journal of Labor Economics, 33(2), 269–296.

14. University of California study on entrepreneurial divorce, as reported in Entrepreneur.com (2024).

15. U.S. Census Bureau. (2024). Young Adults and Traditional Adulthood Milestones.

16. Federal Reserve Bank of Minneapolis. (2025). Tax data reveal rewards and risks of self-employment.

17. Fan, J. S. (2021). Startup Biases. UC Davis Law Review, 56(4), 1835–1896.

18. Morris, M. H., Santos, S. C., & Neumeyer, X. (2020). The liability of poorness. Journal of Business Venturing, 35(4), 105991.

19. Horioka, C. Y. (2024). The impact of intergenerational transfers on the distribution of wealth. Pacific Economic Review, 29(5), 567–585.