Do AI margins move from the labs to the infra?
Gavin Baker put a number on the thing everyone has been circling: the frontier labs earn more than 90% margin on inference. His claim is that when cheaper models get good enough, those margin dollars do not disappear — they move down the stack to whoever sells compute. The question for an allocator is whether that move is real, and when.
inference margin at the frontier labs, the number the whole argument hangs on
Baker · Jul 12what it costs a lab to serve a dollar of tokens, on Sisinty's arithmetic
Sisinty · Jul 12unfunded obligations and off-balance-sheet liabilities carried by those labs
zerohedge · Jul 12views on the original post inside a day, with 100+ quote-posts arguing it
X · Jul 13The chain of the argument
Gavin Baker set it out on Sunday: the mega bull case for AI infrastructure is a world where share shifts away from labs charging 90%+ margins on inference toward cheaper models, open-source or closed. Cheaper intelligence raises the buyer's return on AI spend, which pulls more token demand through the system. Lower margin percentage at the model layer, more margin dollars at the infra layer. He thinks this is the real reason Jensen Huang keeps pushing open weights — and he notes Grok 4.5 already beats Fable on some useful tasks at a much lower cost.
Vaibhav Sisinty made the arithmetic explicit: of every dollar spent on tokens today, roughly ninety cents stays with the lab and under ten cents covers the compute. If the buyer switches to a cheaper model, the dollar does not shrink — it gets redistributed toward the people who own the chips, the memory and the data centres. Others piled in with the demand half of the story. Rohan Paul reached for Jevons: the largest AI workloads probably do not exist yet, because they are still too expensive to run at all.
Then the pushback arrived, and it was better than the usual bubble talk. It came in three flavours: the moat is growing, not shrinking; the pricing mechanism does not work the way the thesis assumes; and even if it does work, the thing that breaks is not equity.
Cheap intelligence only lowers the clearing price once cheap capacity is abundant enough to satisfy the marginal buyer.
The two sides
For — the margin moves down the stack
Baker & the infra bulls
Lower margin percentage at the model layer means more margin dollars at the infra layer. The infra winner is whoever has the lowest cost per token; the model winner is whoever is most token-efficient.
@GavinSBakerNinety cents of every token dollar sits with the lab today. Switching to cheaper models does not cut AI budgets — it reroutes them to the GPU makers, the memory, the data centres.
@VaibhavSisintyAt some point you don't need the best model, only one that's good enough. Infra players don't care whose model you run, which is exactly why Nvidia builds Nemotron.
@edge_of_powerHistory says outsized margins rarely persist. Every cut in inference cost lowers the price of intelligence and stimulates demand — and Nvidia earns regardless of which model wins.
@rickyho_1989Enterprises are losing trust in the labs and moving to fine-tuned open models. Bullish infra, bearish the labs — and OpenAI delaying its IPO may prove a costly mistake.
@justmy2cents111Against — the margin stays where it is
The moat, the price mechanism, the credit
The labs are gaining share because they now train on data their own products generate. Faster, better output wins more users, who feed back more data — a moat that widens rather than erodes.
@iamzeroalphaThe low-cost producer only sets the price if it has capacity to clear the market. If cheap supply is capacity-constrained, the marginal producer sets the price and the cheap model simply prices just below the next-best alternative — scarcity rent, not commoditisation.
@DratchCapThe causality is backwards. Heavy open-source users adopted open weights because their bills would have exploded otherwise — they are not spending more because they started cheap.
@ainativefirmIt assumes buyers keep absorbing hardware inflation. They are already pushing back, and the escape route is running models locally on one-time hardware — which is the bear case for infra, not the bull case.
@sameer_singh17Those labs carry roughly $2 trillion of unfunded obligations already monetised through bond issuance that assumes today's revenue and margins hold into the 2030s. Compress the margin and the credit breaks first.
@zerohedgeWhy this is an investment question
Because it is the value-accrual question, and crypto investors have seen this film before. A layer being essential to a stack tells you nothing about whether it captures the economics of that stack. Chains that host billions of dollars of application revenue collect a rounding error of it. The AI argument is the same shape one level up: everyone agrees the compute is essential, and they are still arguing about who gets paid.
What decides it is scarcity, not necessity. Baker's thesis is really a claim that intelligence becomes abundant while compute stays scarce. Dratch's rebuttal is the one that binds: while compute is scarce, a cheap model does not set a commodity price — it prices a shade under the frontier and keeps the difference as rent. That means the labs' margin compresses only in the world where cheap capacity is genuinely abundant. Which is also the world where selling inference is worth almost nothing.
The TT desk thoughts
We think Baker is directionally right and early, and that the trade most people are putting on is the wrong one. "Long infra, short the labs" is the correct direction with the wrong clock: as long as compute is capacity-constrained, cheaper models harvest scarcity rent instead of forcing a commodity price, and lab margins hold. Position for margin compression to show up in 2027 numbers, not in this quarter's tape — and note zerohedge's point, which is the genuinely non-consensus one: if it does arrive, the first thing that repricing hits is the credit written against those margins, not the equity everyone is watching.
For our book, the crypto expression of this trade is decentralised compute — Bittensor, Akash, Render — and we are passing on it. The DePIN compute networks win share precisely in the abundant-capacity world, which is the same world in which a token of inference carries no margin. You are being sold a bet on AI and handed a bet on volume in a zero-margin commodity. The rule that survives both sides of this argument is the one that has always applied to token value accrual: own the bottleneck, not the layer that is merely necessary. In AI equities that means power, memory and packaging — tokens per watt, as one reply put it. In crypto it means the layer where scarcity is real and enforced, not the layer whose only claim is that nothing works without it. That test disqualifies most of the AI-token complex, and it is the one we are running our AI × crypto exposure through.
Keep reading
AI × Crypto Financial Infrastructure — the full framework · Token vs equity: where value actually accrues · All notes