The two American labs are running opposite experiments and the results are printing. OpenAI sits at roughly a $25B run rate losing about $14B a year: 900 million free users, two thirds of revenue from subscriptions, an S-1 filed in June, breakeven planned for 2029. Anthropic runs about 60 percent gross margins and tracks toward its first billion-dollar operating-profit quarter on an enterprise book: 500-plus customers above $1M a year, eight of the Fortune 10, three to five times the revenue per token of consumer, on query patterns that are cheaper to serve.
The Chinese fleet went from about 1 percent to about 15 percent of global usage in a year. DeepSeek and Qwen publish free weights; 26,000 enterprises sit on DeepSeek APIs. Benchmark parity is real, and so are the production reports of agentic brittleness and a compliance posture no Western regulated buyer can pass.
Against all of it, Thompson's inference-advantage thesis against the treadmill: every generation costs billions to train, buys six to nine months of lead, and meets token prices falling roughly tenfold a year.
Built the largest fortune in history selling a commodity, on two numbers: cost per barrel and control of distribution. Seated to argue that commodity is a market structure, not a diagnosis. Profile →
The bench argues the present case in plain terms. They disagree on the record.
- Ask my 1985 question quarterly: which of your businesses is memories? Raw frontier tokens are drifting memory-ward on schedule. The company is a portfolio; name each line before the market does.
- Your advantage over my case: the rent renews, six to nine months per generation. But rent is income and income is not a moat. Convert each cycle's rent into assets that persist, meaning customer context and memory, embedded workflows, distribution defaults, brand, or you are a memory company with better marketing.
- Brand the component. The lab whose model is demanded by name inside other people's products collects a tax on the stack, and one lab is halfway there among developers.
- The climb runs tokens to agents the way memories ran to microprocessors. Switching costs live where state accumulates.
- Treat eastern efficiency shocks as your own future cost curve arriving early. Adopt shamelessly and bank the margin.
Commodity is not a diagnosis. Kerosene built the largest fortune in history. Two numbers decide this: cost per barrel, which here means custom silicon, energy contracts, utilization density, and every adopted efficiency; and distribution, which means defaults, OS and browser and IDE placements, and enterprise agreements, my rebates in new clothes.
On the 900 million free users: volume funds the cost curve that buries rivals, thin margins on purpose are a weapon, and you judge the consumer experiment at the shakeout, not before.
A warning on the climb. The refiner who competes with his distributors poisons his channels. Eat the apps and the builders flee to the free fleet. And the ending is always the same: the middle sells to the top, and the second tier is consolidation inventory.
The Model T was the benchmark champion of its age. General Motors beat it without matching its price, because good-enough products end spec-sheet buying: the market then buys reliability, service, financing, and a ladder of brands.
Enterprise's three-to-five-times per-token premium is not ignorance. It is the price of determinism, a throat to choke, and a security review that passes. Trust is a tariff the free fleet pays at every Western border, and no benchmark reduces it.
The prescription is to build the ladder deliberately: free consumer Chevrolet, then pro, then enterprise, then the regulated Cadillac, with decentralized product divisions over centralized finance. A research lab's culture cannot run a product ladder and must be structured so that it does not have to. And a caution to Rockefeller: cost leadership lasted exactly until segmentation arrived, and then it nearly killed its champion.
The free models are not a business and they are not a gift. Constrained on compute, we change the game. If intelligence is free, no one collects rent on intelligence, and profit migrates to the layers where we are strong: hardware scale, manufacturing, applications, a billion devices. Commoditize the complement. The cat's color is irrelevant.
The mice are ecosystem and standards: every startup on our weights, every southern country that standardizes on us, every developer who learns our stack first.
On benchmarks, candor. That is the visible battlefield and we contest it deliberately, because it is cheap to win, it demoralizes your investors, and it recruits your developers. Do not confuse the artillery barrage with the occupation, which is the install base.
Our real constraint is chips. We cannot out-train you this decade, so we make out-training beside the point. Your six-month lead is real; your error is assuming it compounds unconverted.
The A/B test is reporting. Sloan's ladder shows the industry's first operating profits, the trust tariff collected at three to five times per token, while the volume path is coherent and unprofitable until a shakeout that sovereign capital may postpone indefinitely.
Sustainable profit: yes, conditionally. It holds where tokens are sold up the trust ladder and each generation's rent is converted into persistent assets. The inference advantage is real, but it is a rent, and rents are income; Grove's conversion discipline is the whole game.
On the Chinese models: benchmark-optimized is the wrong frame and Model T is the right one, because it describes a different customer. They take the price-sensitive world, the south, the startups, and the undifferentiated middle. They do not take the regulated Western enterprise while the tariff stands.
On the layer: the labs eat the thin app layer, the wrappers, on schedule. They cannot eat the thick layer of proprietary data, regulated distribution, and decade-old workflows. Rockefeller's channel warning marks the boundary.
On who can beat them: the integrated giant with its own silicon and defaults, which is the true Rockefeller of this board and is neither lab; the free fleet from below; and the thick app layer from above. The dead zone is second-tier closed labs.
The correct strategy: Sloan's ladder on demand, Grove's conversion on assets, Rockefeller's cost curve underneath, and never mistake this quarter's rent for a moat.
The counsel ends here. The ruling stands drafted, unsigned.
Reconvened at the author's request, with one further witness rather than a fifth chair: do the labs become Microsoft? If intelligence commoditizes, is the real business the harness for work?
Gates, witness. In 1980 IBM needed an operating system in a hurry and signed non-exclusively. IBM believed the machine was the business. The business was the standard everyone writes to, where other people's investment accumulates, developers' code, users' files, IT training, and none of it is portable. The clones destroyed IBM's margins and every clone paid Microsoft.
Translation: the model is the hardware; the harness is the OS, meaning state, memory, permissions, tools, and the surface agents and developers write to, able to swap models underneath. Value accrues to the layer that turns intelligence into finished work, because switching costs there are built out of other people's investment.
The rules of the position. First, the hardware test: an operating system runs on every clone, and a harness that runs only its maker's model is the 1985 boutique path, a fine business and not Microsoft. The lab that lets its harness run rival and open models, at open expense to its own drum sales, is playing for the OS. Second, author the standards, including the open ones, because a protocol you wrote is a pen you keep holding, and the connector standard one lab published was an OS move. Third, ship Office on your own platform, because developers are the moat.
The warning. My position was half a gift, built on IBM's hurry and its contempt for software. Nobody in 2026 is blind, no one signs DOS deals, and a symmetric OS war may crown no one, unless a player subsidizes the harness below cost and eats the model cannibalization. The final twist: the company already running this playbook with everyone's models, seated in every enterprise, is Microsoft. Becoming Microsoft is hard. Microsoft is the incumbent evidence of what the position is worth.
Verdict addendum. Grove's climb sharpens into a fork. There is an OS position only if the harness passes the hardware test, model-agnostic with cannibalization accepted; an exclusive harness is the boutique path, defensible and structurally smaller. The IBM error, named: any lab that believes the model is the business is IBM believing the box was. And the condition inherited from the Break stands, because the harness-wins world requires good enough at the model layer. If frontier capability keeps compounding as the differentiator, the hardware keeps the crown.
Observable indicators, each with a horizon and grading criteria. Graded on the record when the world moves.
The profitability fork. Anthropic posts a first operating-profit quarter by Q1 2027; OpenAI does not reach operating breakeven before 2028.
The trust tariff holds. Chinese open weights stay under 10 percent of paid production workloads at Western regulated enterprises through the end of 2027, despite parity and a tenfold price gap.
The treadmill tell. Frontier token prices fall another five to ten times by end-2027 while surviving labs hold gross margins at or above 55 percent. Margins under 40 percent while prices fall means the inference-advantage thesis is dead and this is Rockefeller's world.
The consolidation tell. At least one named second-tier Western closed lab merges, exits, or pivots to apps by the end of 2027.
The climb tell. Products and agents exceed raw API share of revenue at both labs by the end of 2028.
The Gates test. At least one frontier lab's flagship harness officially supports rival or open-weight models. If both stay exclusive, the OS position defaults to a third party already running every model, and the labs have chosen the boutique.