The Generative AI Bubble Is really Going to Pop - Part Deux
- Thread starter AndrewZ
- Start date
Generally its when companies start at a valuation of X, it increases by hundreds of percent, and then falls back down to X or below. Do you think any of the companies mentioned before would drop to pre-ai valuations if the ai hype disappeared? I don't think so.
Well, I’m sure you could pick a “pre internet” point such that dotcom crash isn’t a crash, then. I named examples of companies whose stock rose and fell enormously in the dotcom bubble, companies that were profitable all along from before the bubble, through the bubble, and after the bubble.
That’s what makes a bubble happen, though.
OK, so these companies as a whole are profitable but ALL their AI sections are losing Billions a month. There are NO AI sections even close to being profitable. These AI driven companies have highly inflated stock values and when the bubble pops, hundreds of billions of dollars of stock valuation will disappear. That would likely trigger a correction in the stock market, a huge loss of consumer confidence. a hit to everyone's stock portfolio, which would then trigger a recession. God knows what the current administration would do, the adults have left the building.
That is kind of my concern. Microsoft and NVidia etc will of course remain profitable, but there's downstream effects from their stocks correcting.
I don't think it being a bubble is even predicated on usefulness or not of AI any more. Obviously even with something as useful as the internet, you can still have a bubble that crashes quite badly.
It's also a technology that is very easy to mispredict the trajectory of. GPT 5 appears to be closer to a human than a chimpanzee. But if an AI, utilizing just a tiny sliver of a datacenter's compute, was as close to a human as a chimpanzee, then it would be months + ~3x the compute away from true human level intelligence. And each datacenter would be able to run thousands such intelligences.
I work close with some very very large tech resellers and integrators. With the teams that are in place to provide thought leadership to customers and in my role specifically I deliver briefings and workshops to the would-be thought leaders in these VARs. These are orgs that focus on the F500 and web scale shops.
This is definitely a bubble, but the entire industry is invested into it because there's nothing else to generate large productivity gains on the horizon, so unlike past bubbles (block chain) I expect that this won't "pop" in the normal sense. We're just in for more evolutionary gains of productivity vs. revolutionary gains like the AI companies are proposing. Think slow-leak instead of rapid disassembly. The leak might even be slow enough to just mean stagnation for a while, rather than a proper bear market downturn.
Ultimately, I expect the first movers get hammered on their ROIs and write down huge losses, but in 7-9 years I assume every developer will have a CoPilot subscription like they had MSDN subs, and IT will be using these tools in massively valuable ways. But, the supply chain issues and (in the USA at least) gutting of entry level hires mean that there won't be a generation of "AI Natives" to help companies along on this journey, so I expect balkanization to happen.
If you're a big company, you'll have the money to get direct business value from AI models. Insurance, Legal and Development are all going to see a further erosion of entry-level positions from AI. The problem will be the 80/20 rule asserting itself and once you're at the limits of what the AI tools can do, there won't be a fresh crop of humans to help understand the limitations and to help improve things.
I expect t his will look like a lot of other technologies that can deliver incredible value, but will wind up gate-kept due to lack of institutional knowledge that can be applied to developing better tooling, and C-suites won't care because they'll still have gotten massive productivity gains from the tools.
This will also stress test Google's entire business model, and I'm not sure what the knock-on effects of Google becoming irrelevant to the internet. Irrelevant is too strong, but like, what does the internet look like w/o Google being Google? We literally haven't seen that world in a generation.
Returning to the integrators I work with, they're struggling for how they're going to make money off AI. Hardware margins aren't huge, and they can't hire enough people able to deliver professional services at scale, even with paying very high (relative) wages. Also, their customers are, largely, in a pucker-mode, waiting for these tools to become more accessible and value propositions to become more obvious. There's been a sort of back-tracking by these VARs to re-emphasize cybersecurity and governance to prepare for AI solutions to come.
But, the market is definitely being propped up. Once we see the brave-new-economy our President has wrought, I think the corrections will begin. Whatever future gains and earnings are expected, might even hold, but new, politically driven costs and headwinds are going to deflate the bubble and bring people back to reality.
I sort of feel like AI is where the desktop computer and internet access was in 1998. People knew this was going to be huge, and it was, but it evolved over almost a decade, and not in a few quarters like people are behaving. SV companies know how to manage a hype-cycle better now, than ever before.
This is definitely a bubble, but the entire industry is invested into it because there's nothing else to generate large productivity gains on the horizon, so unlike past bubbles (block chain) I expect that this won't "pop" in the normal sense. We're just in for more evolutionary gains of productivity vs. revolutionary gains like the AI companies are proposing. Think slow-leak instead of rapid disassembly. The leak might even be slow enough to just mean stagnation for a while, rather than a proper bear market downturn.
Ultimately, I expect the first movers get hammered on their ROIs and write down huge losses, but in 7-9 years I assume every developer will have a CoPilot subscription like they had MSDN subs, and IT will be using these tools in massively valuable ways. But, the supply chain issues and (in the USA at least) gutting of entry level hires mean that there won't be a generation of "AI Natives" to help companies along on this journey, so I expect balkanization to happen.
If you're a big company, you'll have the money to get direct business value from AI models. Insurance, Legal and Development are all going to see a further erosion of entry-level positions from AI. The problem will be the 80/20 rule asserting itself and once you're at the limits of what the AI tools can do, there won't be a fresh crop of humans to help understand the limitations and to help improve things.
I expect t his will look like a lot of other technologies that can deliver incredible value, but will wind up gate-kept due to lack of institutional knowledge that can be applied to developing better tooling, and C-suites won't care because they'll still have gotten massive productivity gains from the tools.
This will also stress test Google's entire business model, and I'm not sure what the knock-on effects of Google becoming irrelevant to the internet. Irrelevant is too strong, but like, what does the internet look like w/o Google being Google? We literally haven't seen that world in a generation.
Returning to the integrators I work with, they're struggling for how they're going to make money off AI. Hardware margins aren't huge, and they can't hire enough people able to deliver professional services at scale, even with paying very high (relative) wages. Also, their customers are, largely, in a pucker-mode, waiting for these tools to become more accessible and value propositions to become more obvious. There's been a sort of back-tracking by these VARs to re-emphasize cybersecurity and governance to prepare for AI solutions to come.
But, the market is definitely being propped up. Once we see the brave-new-economy our President has wrought, I think the corrections will begin. Whatever future gains and earnings are expected, might even hold, but new, politically driven costs and headwinds are going to deflate the bubble and bring people back to reality.
I sort of feel like AI is where the desktop computer and internet access was in 1998. People knew this was going to be huge, and it was, but it evolved over almost a decade, and not in a few quarters like people are behaving. SV companies know how to manage a hype-cycle better now, than ever before.
As much as my wife would like to object to the fact that teachers also provide daycare as a side effect... it's true.
As we all know, screentime is famously good for children's brain development. It's a well-known fact that it leads to improvements in concentration, emotional regulation, reasoning abilities, hand-eye coordination, and measurable increases in happiness and measurable decreases in anxiety disorders. It's basically magic.
I appreciate your counterpoint argument that some AI will be useful and stick. I would agree to that. The question is how useful will generative AI actually be. I would submit that the results of generative AI look amazingly human but are in fact mostly crap. When you realize that the models are just learning to associate things based on the training data but don't actually have any semantic substance, you start to understand why it's all just a pretty fascade. Here are some important takeaway's from the MIT report State of AI in Business in 2025
- AI Will Replace Most Jobs in the Next Few Years → Research found limited layoffs from GenAI, and only in industries that are already affected significantly by AI. There is no consensus among executives as to hiring levels over the next 3-5 years.
- Generative AI is Transforming Business→ Adoption is high, but transformation is rare. Only 5% of enterprises have AI tools integrated in workflows at scale and 7 of 9 sectors show no real structural change.
- Enterprises are slow in adopting new tech → Enterprises are extremely eager to adopt AI and 90% have seriously explored buying an AI solution.
- The biggest thing holding back AI is model quality, legal, data, risk → What's really holding it back is that most AI tools don't learn and don’t integrate well into workflows.
- The best enterprises are building their own tools → Internal builds fail twice as often.
I just can't imagine that. What is the thought logic here? Investors will one day think "oh no, AI is terrible, this makes msft/aapl/nvdia/etc. worthless! I have to sell everything!". Most of these companies had a market cap above one trillion dollars even before the ai hype. Most of their funding is from themselves, they aren't relying on outside investors to prop up their companies. The whole "metaverse" and "NFT" market flopped, which Meta invested heavily in. It caused their stock to fall, but they didn't go bankrupt and there was no recession. They quickly recovered just a year later.
Two interesting things in this:
This totally aligns to what my personal experiences with the AI thought leaders in these extremely large integrators/VARs and my companies own R&D. Agents and Co-Pilot are providing real-world business outcome improvements, but the amount of effort to design, build, train and troubleshoot those models has been massive.
A point that gets overlooked, but is manifesting at orgs who are trying to build their own, or at least heavily augment existing, models is that the hardware and energy cost to re-traing models after you train them initially and find they're not doing what you want, isn't trivial.
The companies having the most success are the ones using off-the-shelf models and tools (Primarily MSFT Copilot, from what I'm seeing) in very targeted ways. Even then, this looks more like super well honed indexing of unstructured data than true "AI". Basically just suped up natural language search. The AI isn't making insights, it's just doing very fancy reporting to allow for insights to be made.
BUT, there is huge value in a super up natural language search engine. Like, really huge value, and that alone might be enough to keep these companies afloat and to keep the bubble from popping. Anyone who has used ChatGPT to dig into what are fundamentally quantitative questions knows it's actually pretty good at it, and the hurdle to massage data into even just a spreadsheet, and then do queries/sorts/filters against that data can tell you that you can save a lot of time with ChatGPT doing a lot of that work for you.
Also, I feel like we're wildly glossing over the real politik issues that are surrounding AI, namely that chip fabs are a huge constraint and even without the extra AI demand induction, it's not like we, collectively, had tons of fab capacity to spare. It takes one dust up over Taiwan to throw this whole market into chaos. But that doesn't mean the market itself is a bubble. The second real politik issue, imo, will be governance of these tools both in the actual government/jurisdictional laws, but just corporate internal governance to ensure these things don't become massive liability engines.
I think the right way to think about this "bubble" is less "when will this pop?" and more "where else should/could that capital be flowing that carries a better potential return?" I'd argue that green energy should be where that's happening, but with the largest economy in the world being cajoled into rejecting any investment into green energy, that's not really going to work in the North American stock markets.
Also, I do think there is a fundamental expectation that some of these mega-caps that are dumping money into AI are going to fail. The issue will be if they all fail in the same way, and/or within close timing to each other.
This totally aligns to what my personal experiences with the AI thought leaders in these extremely large integrators/VARs and my companies own R&D. Agents and Co-Pilot are providing real-world business outcome improvements, but the amount of effort to design, build, train and troubleshoot those models has been massive.
A point that gets overlooked, but is manifesting at orgs who are trying to build their own, or at least heavily augment existing, models is that the hardware and energy cost to re-traing models after you train them initially and find they're not doing what you want, isn't trivial.
The companies having the most success are the ones using off-the-shelf models and tools (Primarily MSFT Copilot, from what I'm seeing) in very targeted ways. Even then, this looks more like super well honed indexing of unstructured data than true "AI". Basically just suped up natural language search. The AI isn't making insights, it's just doing very fancy reporting to allow for insights to be made.
BUT, there is huge value in a super up natural language search engine. Like, really huge value, and that alone might be enough to keep these companies afloat and to keep the bubble from popping. Anyone who has used ChatGPT to dig into what are fundamentally quantitative questions knows it's actually pretty good at it, and the hurdle to massage data into even just a spreadsheet, and then do queries/sorts/filters against that data can tell you that you can save a lot of time with ChatGPT doing a lot of that work for you.
Also, I feel like we're wildly glossing over the real politik issues that are surrounding AI, namely that chip fabs are a huge constraint and even without the extra AI demand induction, it's not like we, collectively, had tons of fab capacity to spare. It takes one dust up over Taiwan to throw this whole market into chaos. But that doesn't mean the market itself is a bubble. The second real politik issue, imo, will be governance of these tools both in the actual government/jurisdictional laws, but just corporate internal governance to ensure these things don't become massive liability engines.
I think the right way to think about this "bubble" is less "when will this pop?" and more "where else should/could that capital be flowing that carries a better potential return?" I'd argue that green energy should be where that's happening, but with the largest economy in the world being cajoled into rejecting any investment into green energy, that's not really going to work in the North American stock markets.
Also, I do think there is a fundamental expectation that some of these mega-caps that are dumping money into AI are going to fail. The issue will be if they all fail in the same way, and/or within close timing to each other.
Yeah it's weird that "doing AI" has been the equivalent of Oprah giving everyone chatbots or whatever.
I think what Google is trying is interesting: embedding Gemini everywhere. I'm not sure how that helps you unemploy people or crank the productivity of the existing workforce. It's helpful, but I don't exactly trust it. But what they're doing with Gemini generally is useful and helpful, which is a nearly-unlimited context window. That's genuinely useful to people like me.
The real work of "doing AI" is incredible fractal. Doing AI should be the work of embedding AI in contexts everywhere in useful ways. Large enterprises aren't typically set up to foster in-house creativity, so it'll be up to the startup ecosystem to bring those embedding capabilities that actually do yield productivity gains. (Though I think those gains won't be as giant as a player like Softbank might imagine.)
As someone "doing AI" the right way at our company, I've been a little surprised by how much training there actually is to use LLMs effectively. Even developers using Claude Code put the robot into a box mentally, and that really hampers how they use the tool. We did a bunch of mob programming sessions, and the whole team learned a ton about how to use these tools much more productively, and now I'm spending hours pairing with non-dev members of our product team. Mostly to help them do data experiments faster. Instead of spending time writing ivory tower requirements docs, do the experiment and see if it's worth going further. Early days, but there's a lot of enthusiasm. There are more ideas than time + space to try them, and some of these tools REALLY lower the bar/barrier for doing these experiments. Writing little Python scripts to tie together APIs and try things out is very quick to do.
Anyway, we're just following the traditional (if slightly elongated) hype cycle with LLMs.
I think what Google is trying is interesting: embedding Gemini everywhere. I'm not sure how that helps you unemploy people or crank the productivity of the existing workforce. It's helpful, but I don't exactly trust it. But what they're doing with Gemini generally is useful and helpful, which is a nearly-unlimited context window. That's genuinely useful to people like me.
The real work of "doing AI" is incredible fractal. Doing AI should be the work of embedding AI in contexts everywhere in useful ways. Large enterprises aren't typically set up to foster in-house creativity, so it'll be up to the startup ecosystem to bring those embedding capabilities that actually do yield productivity gains. (Though I think those gains won't be as giant as a player like Softbank might imagine.)
As someone "doing AI" the right way at our company, I've been a little surprised by how much training there actually is to use LLMs effectively. Even developers using Claude Code put the robot into a box mentally, and that really hampers how they use the tool. We did a bunch of mob programming sessions, and the whole team learned a ton about how to use these tools much more productively, and now I'm spending hours pairing with non-dev members of our product team. Mostly to help them do data experiments faster. Instead of spending time writing ivory tower requirements docs, do the experiment and see if it's worth going further. Early days, but there's a lot of enthusiasm. There are more ideas than time + space to try them, and some of these tools REALLY lower the bar/barrier for doing these experiments. Writing little Python scripts to tie together APIs and try things out is very quick to do.
Anyway, we're just following the traditional (if slightly elongated) hype cycle with LLMs.
I don't think you understand what a stock market bubble is so let me break it down a little. Firstly the metaverse was a mostly Facebook/Meta phenomena. Secondly, NFT was mostly individual investors, and didn't involve as much of the stock market. MSFT/AAPL/NVDIA/TSLA have had their stock prices driven up substantially by investors expecting AI to make those companies much more profitable. This in turn has driven up the NASDAQ by 60% this year alone. If investor sentiment turns sour, based on the expectation of big AI profits not being met in a timely manner, then money will be withdrawn, and the stock prices will drop. If this happens quickly, as it sometimes does, this would be called a bubble pop.
This was a really good article on how despite cost-per-tokens dropping really quickly, demand is so high that even at $200/mo, companies like Anthropic are still losing money:
https://ethanding.substack.com/p/ai-subscriptions-get-short-squeezed
The article is considerably more in-depth than my 1 sentence lede, lots of great charts. Worth the few minutes.
(Anecdotally, I have spent $100 on Anthropic, and have used $432 worth of tokens in about a month, and this tracks with how I've been using AI agents in my daily work. Every other dev on my team has also used multiples of what we've actually paid.)
https://ethanding.substack.com/p/ai-subscriptions-get-short-squeezed
The article is considerably more in-depth than my 1 sentence lede, lots of great charts. Worth the few minutes.
(Anecdotally, I have spent $100 on Anthropic, and have used $432 worth of tokens in about a month, and this tracks with how I've been using AI agents in my daily work. Every other dev on my team has also used multiples of what we've actually paid.)
Interesting article. I didn’t realize they were all running on totally flat subscriptions. My usage is not as advanced as yours and is hidden behind the corporate copilot agreement.
(Also, are the cool kids giving up on capitalization?)
(Also, are the cool kids giving up on capitalization?)
Hmm the math is slightly wonky on this.
(Note I use price per input tokens. Most tasks are VERY read heavy, you spend $10+ on input vs $1 on output even with output being several times more expensive and cached input severely discounted.)
1. Anthropic sets Sonnet and Opus list price quite high and companies, enterprises, actually pay that, using an API key for Claude Code instead of the "personal" Pro or Max subscription. They made a nice chunk of money on them, and it was the only option for CC + centralized billing. Plus it was covered under a different enterprise license agreement vs consumer level one, some companies require that for compliance.
2. GPT 5 throws a wrench in the calculations. It's cheaper per token than 4. GPT 5's list price is $1.25/1M token, and it's equal or better than the old ones. 5 mini is crazy cheap at 0.25/1M and very capable compared to older light models.
3. Sonnet 4 is capable enough that you can get a lot done on it, that throws another wrench in the frontier model only line of reasoning.
4. Viberank leaderboard is people running nonsensical tasks 24/7, this has ended after weekly limits are introduced and the 5 hour block limits seems to have been tightened. It always was a "fair use" style policy, giving away thousands (okay, at "list price") for $200 forever doesn't make much sense. Anthropic was using CC as a loss leader, now they ration tokens. Want unlimited usage? API key at $15 or $3/1M token is still available.
But yeah, as the article said, they will be a day when money runs out and VCs will stop giving away tokens for free and for OpenAI (GPT-5 is a cost cutting release), Anthropic (new usage limits on AYCE plans) and Cursor (severe quota reduction, price increases), this has happened already.
So good job predicting this. But from now going forward, with a positive gross margin, they are fine with you letting it rip - either paying by the token or with the now nerfed subscriptions where max "loss" is limited.
I used to be able to use search better when there were intelligible directives for excluding results I didn't want.From one of those links:
and
Even before AI I would scream and pull my hair out watching otherwise intelligent people using a search engine. That hasn't improved even with the advancements of AI and natural language interfaces.
First are the keywords, many people just don't get them and even at the natural language level their questions are so nebulously phrased as to be nearly semantically meaningless. That's why there will likely always be a place for good doctors, to translate the vague conversations with the patients to meaningful diagnostic inputs for an expert system.
The second problem that bothers me even more, got a good set of keywords, near perfect results - then the user not recognizing the answer to their question in the top result.
It's a great tool but humanity may be too stupid to use it. Ultimately fully functional humaniform robots with infinite patience and gentle interrogation skills to elicit the desires of the human without needing ESP may be in order. A corollary question might be what's the point of the humans.
I'm seeing a lot of that. GenAI replacing purpose built tools, for whatever the reason is. The thing about the purpose built tools is that produced accurate, consistent results. GenAI isn't well suited to that.
Markets run on greed and fear, so that sounds like whistling past the graveyard. I think the more likely thing is the market loses confidence that AI can deliver the massive profits that were promised, investors fear that their stocks aren't worth what they paid for them, and then we get a panic.
On the plus side, it's also replacing a lot of regex based tools, which is good, because those things suck.
GenAI can automate writing that specific tool.
LLMs aren't exact. They can remember short stretches in their context, but longer chunks are JPEGy, so feeding it a big CSV will mess up some fields. But instead of asking it to calculate over the data, ask it to write a python script that does it instead - voila, repeatable, reusable little script that does exactly one thing well.
For now general purpose LLMs don't often offer to write code but something like Claude writes me .py files on the Web UI all the time. And Claude Code of course, you can tell it in plain text that there are huge csv, jsonl, xml files and it should analyze them for structure, then chat about what you want to do with them. Things that used to be a tangle of Excel formulas is now readable code.
Anthropic charges companies by token via API key, so they've solved the usage problem for corporations. Individual users can use way more tokens via personal subscriptions, though. Here's my usage, for example.
The Jevons paradox comes to mind.
| Date | Models | Input | Output | Cache Create | Cache Read | Total Tokens | Cost (USD) | |
| 2025-07-21 | - sonnet-4 | 397 | 20,677 | 215,479 | 2,118,581 | 2,355,134 | $1.75 | |
| 2025-07-23 | - sonnet-4 | 239 | 16,780 | 288,849 | 8,644,789 | 8,950,657 | $3.93 | |
| 2025-07-28 | - sonnet-4 | 1,697 | 2,578 | 63,798 | 209,859 | 277,932 | $0.35 | |
| 2025-07-30 | - sonnet-4 | 133 | 6,614 | 68,099 | 746,910 | 821,756 | $0.58 | |
| 2025-07-31 | - sonnet-4 | 9,719 | 51,869 | 636,206 | 11,500,932 | 12,198,726 | $6.64 | |
| 2025-08-01 | - sonnet-4 | 9,494 | 7,365 | 267,054 | 1,522,542 | 1,806,455 | $1.60 | |
| 2025-08-02 | - sonnet-4 | 7,048 | 7,756 | 133,787 | 2,802,284 | 2,950,875 | $1.48 | |
| 2025-08-03 | - sonnet-4 | 623 | 12,950 | 228,250 | 2,122,959 | 2,364,782 | $1.69 | |
| 2025-08-04 | - sonnet-4 | 7,703 | 37,598 | 551,409 | 5,391,063 | 5,987,773 | $4.27 | |
| 2025-08-05 | - sonnet-4 | 5,560 | 19,873 | 539,558 | 5,605,960 | 6,170,951 | $4.02 | |
| 2025-08-06 | - sonnet-4 | 186 | 5,813 | 90,113 | 1,274,652 | 1,370,764 | $0.81 | |
| 2025-08-07 | - sonnet-4 | 1,112 | 9,592 | 335,615 | 3,654,748 | 4,001,067 | $2.50 | |
| 2025-08-08 | - sonnet-4 | 2,636 | 19,290 | 467,336 | 5,487,679 | 5,976,941 | $3.70 | |
| 2025-08-09 | - sonnet-4 | 50 | 2,362 | 56,277 | 420,074 | 478,763 | $0.37 | |
| 2025-08-10 | - sonnet-4 | 1,351 | 44,257 | 598,425 | 10,566,438 | 11,210,471 | $6.08 | |
| 2025-08-11 | - sonnet-4 | 7,003 | 8,267 | 331,164 | 3,490,660 | 3,837,094 | $2.43 | |
| 2025-08-13 | - sonnet-4 | 2,339 | 68,675 | 761,611 | 15,560,920 | 16,393,545 | $8.56 | |
| 2025-08-16 | - sonnet-4 | 2,352 | 15,271 | 416,925 | 7,037,084 | 7,471,632 | $3.91 | |
| 2025-08-17 | - sonnet-4 | 1,014 | 13,013 | 640,477 | 13,957,982 | 14,612,486 | $6.79 | |
| 2025-08-18 | - sonnet-4 | 550 | 11,498 | 523,200 | 8,786,120 | 9,321,368 | $4.77 | |
| 2025-08-19 | - opus-4 / - sonnet 4 | 6,344 | 138,894 | 2,994,844 | 37,120,546 | 40,260,628 | $29.51 | $95.74 |
| 2025-08-20 | - opus-4 / - sonnet 4 | 4,067 | 19,619 | 556,244 | 10,946,071 | 11,526,001 | $6.58 | Upgraded to Max |
| 2025-08-21 | - opus-4 / - sonnet 4 | 2,184 | 21,714 | 1,047,933 | 14,532,902 | 15,607,733 | $17.22 | |
| 2025-08-22 | - opus-4 / - sonnet 4 | 358 | 12,660 | 384,396 | 7,866,710 | 8,264,714 | $14.83 | |
| 2025-08-23 | - opus-4 / - sonnet 4 | 2,349 | 58,719 | 3,176,033 | 65,484,344 | 68,721,445 | $58.11 | |
| 2025-08-24 | - opus-4 / - sonnet 4 | 1,405 | 21,855 | 2,400,103 | 30,659,609 | 33,082,972 | $39.25 | |
| 2025-08-25 | - opus-4 / - sonnet 4 | 1,097 | 19,794 | 1,545,784 | 26,541,221 | 28,107,896 | $30.10 | |
| 2025-08-26 | - opus-4 / - sonnet 4 | 4,708 | 35,736 | 1,328,338 | 22,802,831 | 24,171,613 | $25.47 | |
| 2025-08-27 | - opus-4 / - sonnet 4 | 4,177 | 70,990 | 2,735,863 | 49,560,333 | 52,379,382 | $55.07 | |
| 2025-08-28 | - opus-4 / - sonnet 4 | 8,389 | 72,768 | 2,489,933 | 42,529,280 | 45,100,370 | $56.70 | |
| 2025-08-29 | - opus-4 / - sonnet 4 | 1,660 | 49,119 | 1,767,127 | 18,557,542 | 20,375,448 | $32.85 |
They HAVE become more profitable. This isn't speculation that the companies "might" generate more revenue, they actually are, right now, actively generating more revenue.
See also: the mag 6 beat estimates this quarter, and a bunch of them beat it last quarter also:
https://arstechnica.com/civis/threa...-going-to-pop-part-deux.1508924/post-43938736
Btw, it can still be a bubble. If revenue increases don't keep up with expectations, then the bubble will pop. It's a bit early to tell, IMO. If someone were to have a lot of conviction, they could short any of the mag 6. But lots of things affect revenue; causality would be difficult to establish -- like maybe it would be tariffs depressing consumer spending which doesn't necessarily point the finger at LLMs.
Investment bubble are kinda interesting, though. Historically investment bubbles have kinda led to "the next thing": overinvestment in infrastructure allows new industries to form on top of that overbuilt infra. The social benefits over the long term usually exceed the pain of a short-term bubble burst.
Debt bubbles are different.
https://arstechnica.com/civis/threa...-going-to-pop-part-deux.1508924/post-43938736
Btw, it can still be a bubble. If revenue increases don't keep up with expectations, then the bubble will pop. It's a bit early to tell, IMO. If someone were to have a lot of conviction, they could short any of the mag 6. But lots of things affect revenue; causality would be difficult to establish -- like maybe it would be tariffs depressing consumer spending which doesn't necessarily point the finger at LLMs.
Investment bubble are kinda interesting, though. Historically investment bubbles have kinda led to "the next thing": overinvestment in infrastructure allows new industries to form on top of that overbuilt infra. The social benefits over the long term usually exceed the pain of a short-term bubble burst.
Debt bubbles are different.
Last edited:
We are still benefitting from the crazy amount of dark fiber in the ground. A GPU / datacenter bubble is less durable, silicon is much less valueable in 20 years, but buildings won't disintegrate or rot.
And even silicon seems to live longer nowadays. A100 from 5 years ago is still very powerful for inference, B100/200 had a much longer learning/optimizing phase and still isn't used in training runs (needs to be stable for months long jobs). Rubin (rumored sep 2025) R100 is also about to be released, but will face maybe the same chunky phase of getting the bugs out and restricted to light inference duty until then. Once it has been written off it's not that expensive to keep running, new GPUs aren't that much ahead in perf/watt, especially in memory heavy workload (inference) - all they care about is GB/s, not FLOPS.
And even silicon seems to live longer nowadays. A100 from 5 years ago is still very powerful for inference, B100/200 had a much longer learning/optimizing phase and still isn't used in training runs (needs to be stable for months long jobs). Rubin (rumored sep 2025) R100 is also about to be released, but will face maybe the same chunky phase of getting the bugs out and restricted to light inference duty until then. Once it has been written off it's not that expensive to keep running, new GPUs aren't that much ahead in perf/watt, especially in memory heavy workload (inference) - all they care about is GB/s, not FLOPS.
I was thinking more of "LLMs everywhere" like you've got an LLM on your phone, your computer, etc. You upgrade it with the OS, or add your own LLM to a library of LLMs. Specialty LLMs, general purpose LLMs, etc. That could be a downstream side effect.
WRT LLM factories: there's a lot of infrastructure being built that may also be useful over the long term. I'm specifically thinking about power, but I also kinda wonder if weird stuff like better desalination might be a side benefit of these investments.
(I don't know about desalination specifically, but more like "what stuff exists in the supply chain that might be stressed by LLM factories, and how might it improve as a consequence of this private sector investment bonanza". I think it's hard to foresee side effects of investment bubbles until they're behind you. Netflix and YouTube seem super obvious now; they weren't in 2002.)
WRT LLM factories: there's a lot of infrastructure being built that may also be useful over the long term. I'm specifically thinking about power, but I also kinda wonder if weird stuff like better desalination might be a side benefit of these investments.
(I don't know about desalination specifically, but more like "what stuff exists in the supply chain that might be stressed by LLM factories, and how might it improve as a consequence of this private sector investment bonanza". I think it's hard to foresee side effects of investment bubbles until they're behind you. Netflix and YouTube seem super obvious now; they weren't in 2002.)
Last edited:
Clusters and clouds with them are also the most concentrated engines for FP64 calculation, since pretty much everything that's come after them has traded that off for increased throughput of (sometimes much) lower precision. As someone whose work is deep in computational science, I'm really looking forward to chip and infrastructure providers being willing to take the merely "very good" margins these applications offered them once the huge margin LLM bubble pops.
I'm not sure which companies, exactly, are at issue. But google is very plausibly one of those heavily invested in AI across its business, and is certainly exposed (and investing huge amounts of capex) via its status as a hyperscaler. Their revenue is up ~13% YoY for the last two years. Source: https://abc.xyz/assets/cc/27/3ada14014efbadd7a58472f1f3f4/2025q2-alphabet-earnings-release.pdf
Maybe you think this is despite, not due to, their massive investment in AI. Maybe! But they are in fact generating more revenue.
https://www.theregister.com/2025/09/03/anthropic_funding/
Investors throw another $13B on the Anthropic cash bonfire
Or 37x Price/Sales. It's not that outrageous, some more bubbly outfits get an even higher score. It certainly is higher than 5x for midcap SaaS, or 10x for rapid growing startups, but I expected an absurd 10x or more multiplier for AI, not barely 4x.
Plus when you can 5x your revenue in 8 months, what seems like a high ratio will shrink to a middling one in just 2 more years. There's a whole bunch of other companies with P/S over 15.
Investors throw another $13B on the Anthropic cash bonfire
Or 37x Price/Sales. It's not that outrageous, some more bubbly outfits get an even higher score. It certainly is higher than 5x for midcap SaaS, or 10x for rapid growing startups, but I expected an absurd 10x or more multiplier for AI, not barely 4x.
Plus when you can 5x your revenue in 8 months, what seems like a high ratio will shrink to a middling one in just 2 more years. There's a whole bunch of other companies with P/S over 15.
Last edited:
That summary indicates that AI is part of Google Cloud Services' division, which accounted for $13.6B out of $96B for that quarter. The lion's share of that is the cloud infrastructure, like mail and other office services. Their investment alone dwarfs any revenue it produces in that division.
They are increasing investment by $10B this year, primarily for datacenters for AI.
https://www.theguardian.com/us-news/2025/jul/23/google-expected-to-report-earnings
Even if the entire growth in that division is due to AU, they are spending over $2 to earn $1. That's a significant drain on overall margins. So, try again.
Last edited:
Still a profitable company. The original claim was that the company is profitable. It is. All of them - Nvidia, Microsoft, Google, Amazon, are profitable companies, even if heavily invested in AI.
The claim wasn't that their AI divisions are profitable and not losing money.
ETA: the only one among them that's not so profitable is Tesla, but its CEO decided to do a public hara-kiri so... Is it relevant to the discussion?
The claim was that the AI companies, you know, Anthropic and OpenAI, are unprofitable. Trying to dodge that argument by bringing up Alphabet is a misdirection because they can burn a mountain of money before AIO is shown to be a dog.
You guys know this.
Microsoft, Google, and Amazon have diversified operations and will be fine, despite AI.
Nvidia will suffer from the bursting of the bubble as its revenue and valuations are dependent on fairy tale AI-driven growth curves.
Meta.. well, they have Zuckerberg burning money on one dumb marquee project after another due to their main product being mostly used by oldsters, and no amount of AI will help that.
X-Twitter will continue to decline, as it has always been a second-class AI product engineered to be biased.
The original claim heavily implied that AI was driving revenue growth when it wasn't.
TSLA is mismanaged, and the captive board is indifferent. They just voted to pay out evey dollar of profit the company ever made to their mostgly-absent and drug-addled CEO.
You guys know this.
Microsoft, Google, and Amazon have diversified operations and will be fine, despite AI.
Nvidia will suffer from the bursting of the bubble as its revenue and valuations are dependent on fairy tale AI-driven growth curves.
Meta.. well, they have Zuckerberg burning money on one dumb marquee project after another due to their main product being mostly used by oldsters, and no amount of AI will help that.
X-Twitter will continue to decline, as it has always been a second-class AI product engineered to be biased.
The original claim heavily implied that AI was driving revenue growth when it wasn't.
TSLA is mismanaged, and the captive board is indifferent. They just voted to pay out evey dollar of profit the company ever made to their mostgly-absent and drug-addled CEO.
Right! That's why I said:
To repeat: Google is increasing revenue. The original claim was about the companies, not the divisions within them. Maybe "Google AI" (Whatever this is supposed to be) is losing money. Fair enough!
But Google isn't losing money! They're making more than ever, again maybe despite but definitely while, investing buckets of cash into AI.
That's an irrelevant argument. AI is still a drag on revenue and profit. It has not made anyone any money yet, except individual officers and founders.
Like I said in my original post, I wasn't sure which companies were at issue. But I think claims about "AI companies" should obviously include the largest AI companies on planet earth. So, Google (alphabet), Meta, Microsoft, Amazon seem like fair picks. It'd be one thing if these companies were slowly backing away from the investment, but they aren't. Maybe you think it's all hype! Fair enough, but that hype is now building a lot of very profitable infrastructure. And since it's only (say) three years old, it seems a bit early to call it a dud. I'm not sure I know of any really big new technology requiring massive infrastructure buildout that was revenue positive in less than a decade?
As an aside: I'm not sure if you meant to be focused on equity markets, but if so, then who cares about anthropic and openai burning VC $?
There's a huge misconception here. Revenue (which recurs every year) and capex (which can be cancelled next year, and stuff you bought stays with you) can't be cancelled out with A - B = profit.
When you calculate profitability, CapEx doesn't matter. Especially in real estate and infrastructure, you can write that off in 10 to 20 years so if you have
$1000 revenue, $5000 CapEx
$1000 revenue goes onto profit and loss sheet
$5000 CapEx is just trading cash for a fixed asset, no effect on P&L.
$500 of that CapEx can be written off to reduce tax on year 1
$500 profit!
So Anthropic has $5B revenue, ??? spending on other stuff, and possibly they will dump $10B of $13B raised into assets that write off on an average of 8 years (GPU's in 5, buildings much longer). How does this show up on the balance sheet?
$5B revenue
-$10/8=1.25B depreciation
-??? cost of goods sold etc etc
What you can't do is $5B revenue, minus $10B spending on GPU's, voila $5B loss. No tax guy will agree to this and the government wherever you are want their cut of your revenue.
When you calculate profitability, CapEx doesn't matter. Especially in real estate and infrastructure, you can write that off in 10 to 20 years so if you have
$1000 revenue, $5000 CapEx
$1000 revenue goes onto profit and loss sheet
$5000 CapEx is just trading cash for a fixed asset, no effect on P&L.
$500 of that CapEx can be written off to reduce tax on year 1
$500 profit!
So Anthropic has $5B revenue, ??? spending on other stuff, and possibly they will dump $10B of $13B raised into assets that write off on an average of 8 years (GPU's in 5, buildings much longer). How does this show up on the balance sheet?
$5B revenue
-$10/8=1.25B depreciation
-??? cost of goods sold etc etc
What you can't do is $5B revenue, minus $10B spending on GPU's, voila $5B loss. No tax guy will agree to this and the government wherever you are want their cut of your revenue.
OK, but that's just not true, is it? Google has been making money by using "AI" (machine learning) for roughly two decades. In that frame, the recent investment in large foundation models can just be thought of as an extension of their existing commitment to leverage data to organize the world's information into profitable products. I agree: it's a new product, yet to find product market fit! But the AI company behind it has a proven track record of leveraging the exact same tech to make trillions of dollars in revenue.