The Speed Limit Is a Pricing Strategy

In ten days, the AI industry produced three things that don't obviously fit together.

An Anthropic researcher, Jacob Coxon, resigned publicly, warning that frontier labs are racing toward self-improving superintelligence and "gambling with our lives." Three days later, Dario Amodei published We Must Pace the Frontier, calling for embedded third-party evaluators and coordination among frontier labs in democratic countries on safety standards and the pace of development — noting that government may need to mediate the antitrust problems such coordination creates. And in the same window, reporting put Anthropic's compute commitments at roughly $517 billion, ahead of a Nasdaq listing targeting a valuation around $2 trillion.

A viral strand of commentary stitched these into one story: the labs are faking a safety panic to win regulatory capture, protect a duopoly, and choke off cheap open-weight competition before their IPOs expose the books. Some of the All-In hosts went further and alleged the resignation itself was a coordinated rollout — contested, and not something I can verify.

I'm not interested in adjudicating motives. I'm interested in the thing I was trained to do in due diligence: ignore the narrative, open the unit economics, and ask what the numbers would have to be for the story to be true.

One caveat before the numbers. Every figure below comes from journalism sourced to private investor communications — The Information, Bloomberg, the FT. Anthropic's S-1 is still confidential. None of this is audited, and the company has not confirmed the totals on the record. I'd treat all of it as directionally useful and none of it as settled.

The regulatory-capture argument usually rests on one claim: the labs lose money on every subscriber. You pay $200 a month; the lab burns thousands of dollars of compute serving you; open-weight models on a cheap box threaten to expose the whole thing.

That was a plausible story in 2024. The 2026 reporting points the other way. Ahead of the IPO, the FT reported inference gross margins above 80%, a revenue run rate around $65 billion as of the end of July, and an expectation of a second consecutive quarter of adjusted operating profit.

Read the asterisks, though, because they matter. That 80% is measured before revenue shared with distribution partners and before model training costs — the second exclusion usually disappears in the shorthand version. "Adjusted" operating profit has no published reconciliation. And skeptics, Ed Zitron most pointedly, argue the first profitable quarter owed more to temporarily discounted compute than to structural improvement.

Even discounting all of it, the direction is clear enough: serving tokens is not a charity. It's a high-margin business sitting on top of an enormous fixed-cost base. That changes what the incumbents actually have to protect.

A gross margin that high on inference means the price of a token sits far above its marginal cost. That spread survives only while buyers have no credible substitute at comparable quality.

Open-weight models are the substitute, and the economics of the substitute are structurally different:

Past a certain volume, the fixed-cost curve crosses below the variable one. Where it crosses depends on utilization, model quality for the specific task, and the operational overhead most comparisons quietly omit — it is not "a cheap box replaces your API bill." But for high-volume, well-defined workloads — classification, extraction, routing, first drafts — that crossover is arriving fast.

So the pressure on the closed labs isn't that they're losing money. It's that the market price of "good enough" intelligence is increasingly set by someone else's fixed-cost curve, and every point of price compression lands on a very rich margin.

Here's where the M&A lens kicks in. A compute commitment is not capex you can pause; it's a contract. In the framework I use in The Intelligent Economy, the contract is the asset — and the liability.

The $517 billion figure deserves precision. It is The Information's aggregate of deals announced or reported over eleven months — AWS, Google, Microsoft, SpaceX, Nscale, Fluidstack, Lambda, AMD and others — covering roughly 14.8 gigawatts, with spending spread across the next decade. Some are non-binding. It is not one contract, and it is not a company disclosure. What it does show is a step change: the comparable prior number Anthropic gave investors was around $180 billion through 2029.

An obligation of that shape needs two things for the IPO story to hold: sustained demand, and a sustained price per unit of intelligence.

Read through that lens, "pacing" has an industrial-economics shape regardless of the sincerity behind it:

To be fair to the text: the essay does not mention open-weight models at all, and it explicitly says pacing does not mean halting training or technical progress. Anthropic has also, separately, stated it has never advocated banning open-weight models. The regulatory-capture reading is an inference about second-order effects, not something anyone proposed. But investors price second-order effects.

If the closed labs need the price of intelligence to hold, the company selling the shovels needs the opposite — cheaper tokens mean more tokens, and more tokens mean more GPUs sold to more buyers.

Jensen Huang's first-ever post on X, on 24 July, was a letter titled Open Weights and American AI Leadership, arguing the US needs both frontier closed and frontier open models and warning against premature restrictions on open weights. It gathered roughly 235 signatories inside a week, including OpenAI, Google, Microsoft, Meta and AMD. Anthropic did not sign. It published its own position three days later: no ban on open weights, but export controls, a distillation crackdown, and mandatory pre-release safety testing for sufficiently capable models, open or closed.

Six weeks later, on 3 September, NVIDIA announced it had agreed to acquire Hugging Face for $12.93 billion — the home of the open-model developer community, with 18 million developers, 3 million models and 200,000 companies on the platform. Huang committed publicly that the platform stays open and that NVIDIA compute will not be required to build or deploy there. The deal is signed, not closed: completion is expected in the first half of 2027, subject to regulatory review, and this is a real merger filing rather than the licensing-shaped structures NVIDIA has used elsewhere.

That is the cleanest signal in this whole episode. The most important supplier in the stack has moved to own the distribution layer for open weights. Whatever happens in Washington — and the President dismissed the slowdown call on 13 September, framing AI as a race the US has to win — the open-weight ecosystem now has a sponsor with 12.9 billion reasons to keep it healthy.

The fault line in AI is no longer "safety versus acceleration." It's who captures the falling cost of intelligence.

For founders buying intelligence: treat the model bill like any other input cost with concentration risk. The practical moves are unglamorous:

For investors: separate the layers.

A company with high inference margins and enormous long-term compute commitments is unusually exposed to falling intelligence prices. A company selling the infrastructure behind those tokens has the opposite exposure: it can benefit when falling prices drive much higher volumes.

That is a statement about balance sheets, not motives. Safety concerns can be entirely sincere and the pricing consequences can still be real. Our job, as builders and investors, is to price the second thing.

Safety concerns can be sincere and the pricing consequences real at the same time. One is argued in public. The other is where the market reprices it.

Carol Chen

Founder, Compute Notes

Builder of AI-native businesses and investor in AI infrastructure

LinkedIn: linkedin.com/in/carol-c-76461498

Website: computenotes.co

Network: network.computenotes.co

Get a free AI spend teardown → watt.computenotes.co

This is analysis, not investment advice.

#AIInfrastructure #FinOps #EnterpriseAI #AIEconomics #Compute #PhysicalAI #Token