Sitemap
David Schaerf, North Star Labs

David Schaerf, North Star Labs

Press enter or click to view image in full size

The Intelligence Stack

7 min readJun 10, 2026

--

AI tokens right now are priced like Facebook ads in 2015.

If you were spending on Facebook acquisition in 2015, you remember the economics. A click cost thirty cents. A qualified lead cost a few dollars. The ROI was so obvious that the only question was how much budget to move in. Then the window closed. Costs tripled. Then tripled again. By 2020 the same lead cost ten times what it cost five years earlier. The companies that built organic distribution early survived. The ones that stayed dependent on paid acquisition got squeezed.

The same pattern is playing out in AI right now. Frontier model pricing is artificially low. Not because inference is cheap. Because the hyperscalers are buying market share and eating the loss.

And just like Facebook ads in 2016, the prices are already starting to move.

The subsidy window is closing

On June 1, GitHub switched every Copilot user from flat-rate to usage-based billing. Developers started burning through a month of credits in two hours. GitHub’s own explanation was plain: the pricing needed to reflect actual compute costs.

This was not one company making a strange decision. Enterprise AI software prices are up 20 to 37% on renewal across the board, according to Tropic’s procurement data from March 2026. Anthropic, OpenAI, and Google have all raised prices. GitHub is killing annual plans entirely.

Frontier AI models have been priced at a loss since the beginning of the current cycle. The hyperscalers financing them were buying market share and training moats, not running a sustainable business. It worked. Enterprises built workflows on those prices, hired around them, and assumed costs would keep falling.

Inference costs for standard tasks did fall over the past two years. But frontier model pricing was never on a real downward trajectory. It was suppressed while the market was being captured.

That suppression is ending. The floor is becoming the ceiling.

And here is the thing that makes this moment particularly sharp. The expensive end of the stack is getting more expensive at the exact time the cheap end is getting dramatically cheaper.

I think of this as the cost inversion. It is the central dynamic of the next three years. And most organizations have not noticed because they are still looking at last year’s API pricing page.

Four tiers, four cost structures

There are four tiers of AI deployment available right now. They have different economics, different capability ceilings, and different ROI profiles. Understanding the difference is the most important strategic decision in enterprise AI in 2026. I call the framework The Intelligence Stack.

On-device is where the volume goes. Apple committed to this at scale during WWDC this week, rebuilding its entire AI architecture around a System Orchestrator that routes simple tasks to on-device models running on Apple Silicon and sends only complex requests to the cloud. When iOS 27 ships in September, that architecture reaches 2.5 billion active devices. On the open hardware side, NVIDIA’s RTX Spark arrives this autumn with a Blackwell-class GPU and 128GB of unified memory inside a laptop. The on-device tier is no longer a compromise for lightweight tasks. It is serious compute at zero marginal cost per query.

On-premise is the quiet winner for regulated industries. Llama 3.1 70B, Mistral Large, fine-tuned open source variants running on your own infrastructure. No per-token billing. Full data sovereignty. The cost per useful output keeps dropping as the infrastructure investment spreads across more workloads. If your organization has compliance constraints, this tier gets more attractive every quarter.

Cloud mid-tier is the workhorse nobody talks about. Claude Haiku, GPT-4o mini, Gemini Flash. Scalable, predictable cost, strong enough for most enterprise tasks. Proposal drafting. Report generation. Code completion. Structured data extraction. This is where the ROI is clearest right now, and most organizations are barely using it because they jumped straight to the top.

Frontier cloud is where most of the money is going and most of the waste is happening. Claude Opus 4, GPT o3, Gemini Ultra. Extraordinary reasoning capability. Extraordinary cost. The right tier for complex legal analysis, novel engineering problems, multi-step research that a cheaper model genuinely cannot do. But the gap between frontier and mid-tier has to be material to the outcome. If it is not, you are paying for horsepower you are burning in traffic.

Press enter or click to view image in full size

The cost inversion

Look at where the four tiers are heading.

Get David Schaerf, North Star Labs’s stories in your inbox

Join Medium for free to get updates from this writer.

Frontier pricing is at its lowest relative point right now. The subsidy window is still open, barely. GitHub’s shift, the 20 to 37% renewal increases, the Anthropic and OpenAI price adjustments: these are early signals of a normalization that accelerates from here.

At the bottom of the stack, costs are moving in the opposite direction. Apple is routing routine AI work on-device by default. NVIDIA is putting Blackwell-class inference into laptops. On-premise infrastructure amortizes steadily downward. The lower tiers get structurally cheaper every month.

The lines on the chart are crossing. Organizations that have already built tiered routing are on the right side of the inversion. The ones running everything through Tier 4 are on the wrong side and the gap is widening.

Uber burned through its full 2026 AI budget in four months. Microsoft cancelled Claude Code licenses across the company before the fiscal year closed. Every headline called these cost blowups. They were not. They were routing failures. Both deployed frontier intelligence on standard coding workflows and documentation. Tasks that belonged two or three tiers lower.

They were not outliers. They were the first organizations large enough and visible enough that the budget failure became public. The same misconfiguration is happening at hundreds of companies that are not making headlines yet.

Where human judgment still compounds

There is a second dimension to the routing decision that gets less attention. It is not just about which model tier. It is about where human judgment is still the thing that matters.

Sorting email, classifying documents, scheduling, compliance checks. High-volume, low-judgment tasks. They belong on Tier 1 or Tier 2. No human in the loop. Near-zero cost.

Board decisions, novel product strategy, a difficult conversation with a key client. Low volume, high stakes, high judgment. A model can draft the brief. The human makes the call.

The zone I find most interesting is where volume and judgment are both high. Portfolio analysis across hundreds of positions. Legal review across a thousand contracts. Governance at scale inside a hospital system. This is where human insight scales through AI reach, and where the compounding gets real. The human sets the direction. The agent executes. Neither gets there alone.

The mistake runs both ways. Deploying agents where you need human judgment gets you fast, wrong answers. Deploying humans on work that should have been automated years ago is the most expensive routing mistake in the stack. We just could not see it before agents made it obvious.

Press enter or click to view image in full size

The third cost pressure

Two cost corrections are already underway. Subsidies are normalizing. Routing misallocation is burning budgets faster than anyone planned.

A third one is coming.

Governments will tax agent deployment. The logic is not complicated. Income tax and payroll contributions assume that economic output flows through people. When agents absorb that work, the revenue base erodes. The fiscal response is an agent levy of some kind. My hunch is that the timing is closer than most people expect.

When it arrives, cloud-tier costs shift upward again. On-device and on-premise sit below the threshold.

One more thing worth flagging. Hyperscaler AI infrastructure spend is approaching $650 billion in 2026 and the full supply chain number trends toward a trillion. All of that investment assumes a world where most AI work runs through centralized cloud compute. But Apple just committed 2.5 billion devices to on-device-first architecture. NVIDIA put Blackwell-class inference in a laptop. If the volume actually migrates to local tiers, the demand assumptions behind the biggest capex cycle in technology history are overstated.

I want to be clear that this is a hypothesis. But the economics and the architecture are both pointing the same way.

The routing decision

The subsidy window is closing. The cost lines are crossing. And the organizations that figured out the routing decision early are pulling away from the ones that did not.

This is not a technical question for the infrastructure team. It is the most consequential financial decision of the AI era.

The organizations getting this right are building a cost advantage that compounds as pricing normalizes. The ones running everything through Tier 4 are building on a foundation that is already cracking.

--

--