The Three-Company Town

A company town is a place where the mine, the store, the housing, and the church all have the same owner. The worker earns scrip. The scrip is only good at the store. The store is owned by the mine. The mine owns the rent on the shack the worker sleeps in. The church tells the worker to be grateful.

The generative-AI economy is not a market. It is a company town. The town has three mines and one landlord, and you are living in it whether you noticed or not.

The landlord is NVIDIA

Every frontier model trained in the last four years ran on NVIDIA silicon. Every serious inference deployment in 2026 is running on NVIDIA silicon. NVIDIA’s gross margin on its data-centre GPUs is somewhere above 75 per cent, which is the kind of margin you only get when your customers have no second source.

A monopolist extracts rent until someone builds a competing factory. NVIDIA has had ten years of notice that a competing factory was needed, and there is still no competing factory. Google’s TPU is captive to Google. Amazon’s Trainium is captive to AWS. AMD’s MI300 ships in numbers you can count on your fingers. Intel’s Gaudi is a rounding error. The startups are underfunded.

This is not a market failure. This is a market working exactly as designed, by a player with the capital, the foundry relationships, the software lock-in, and the regulatory lobbyists to keep it that way.

The CUDA software moat is the real lock. Every piece of AI research in the last decade was written against CUDA. Every grad student’s dissertation has CUDA in the dependency graph. Every production pipeline targets CUDA. Rewriting the stack to run on a competitor is a multi-year engineering investment at every single firm in the industry. Nobody has done it because nobody has to. NVIDIA ships, so you ship NVIDIA.

The landlord sets the rent. The rent goes up.

The three mines

Sitting on top of the landlord’s real estate are three mines. OpenAI, Anthropic, and Google DeepMind. These are the firms with the capital, the talent, and the tolerance for loss required to train frontier models at the current frontier. Everyone else is a tenant on one of their platforms, a fine-tuner of one of their base models, or a downstream reseller of one of their APIs.

A market has entry. A company town does not. Consider what it would take to be the fourth firm.

• Eight to ten billion dollars of capital for a training run at 2026 frontier scale.
• Sustained access to tens of thousands of top-bin NVIDIA GPUs, which NVIDIA allocates preferentially to the existing three.
• Two to three years of runway with no product revenue, which requires investors willing to compete with Microsoft’s balance sheet, Amazon’s balance sheet, and Google’s balance sheet.
• A research team at the frontier, which requires compensation packages the three incumbents are currently using to acqui-hire any serious competitor before the competitor can close a Series B.

This is not a short list of challenges. This is a list of moats arranged in a star pattern around the incumbents, each moat reinforced by the others. Mistral had the research. Mistral did not have the capital. So Mistral is now partially owned by Microsoft. Inflection had the capital. Inflection did not have the compute allocation. So Inflection’s team is now inside Microsoft. Stability had the brand. Stability did not have any of the above. So Stability is a distressed asset.

A market that eats every new entrant before it grows is not a market. It is a moat, dressed in a pitch deck.

The company store

Tenancy in the town is mediated by the hyperscalers. Microsoft Azure resells OpenAI. Amazon Bedrock resells Anthropic. Google Cloud resells its own DeepMind work. If you are a Canadian bank or hospital or government, and you want to use a frontier model, you are signing a contract with one of three American companies, hosted in one of two American legal jurisdictions, under the cloud provider’s acceptable-use terms.

There is no fourth door. There is no domestic door. Cohere’s enterprise business is a rounding error against Azure’s share of the same customers. The scrip is the API key. The scrip is only good at the store. The store is owned by the mine.

The acceptable-use policy is the company-town church. It tells you what you are not allowed to build, which is increasingly everything the incumbent would rather build itself. You will read the policy. You will nod. You will sign. Because there is no other store in town.

The Canadian antitrust problem

Canada’s Competition Bureau has, as of this column, not issued a serious public statement about concentration in the AI foundation-model market. The Bureau has published general thought pieces on “digital markets.” It has not opened a market study, has not demanded data access from the incumbents, has not coordinated with the US FTC or the UK CMA on the three investigations those two agencies are currently running.

The reason is not intellectual. The reason is structural. Canadian competition law has a long tradition of treating efficiency as a defence, which means a monopolist can argue that its scale reduces costs to consumers, which means a monopolist usually wins. The efficiency defence was written for a 1980s manufacturing economy. It is incoherent in an economy where the dominant input is data scraped for free.

The second reason is that Canada, unlike the EU, does not currently have an ex-ante regulatory regime for digital markets. The EU has the Digital Markets Act. The UK has its Digital Markets, Competition and Consumers Act. The US has ongoing FTC and DOJ cases. Canada has a consultation paper from 2023 that has not become an act.

Meanwhile every Canadian small business, every Canadian school board, every Canadian hospital is signing long-term contracts with the three-company town, at terms set by the landlord.

What a break-up of the stack looks like

Antitrust, properly applied, would look at the AI stack and ask where the chokepoints are. There are four.

1. Silicon. Force foundry-level disclosure of GPU allocation. Force CUDA to be licensed under FRAND terms. Publicly fund a competing compiler target for at least one non-NVIDIA accelerator.
2. Cloud. Structural separation between the cloud and the model. If Microsoft sells Azure, Microsoft does not get to also own an exclusive commercial relationship with a frontier model lab. Pick a lane.
3. Model. Mandate interoperability of fine-tuning artifacts across vendors. A fine-tune trained on one model should be portable to another. Today it is not. The incumbents want it that way.
4. Distribution. Require that any firm with more than a threshold share of the foundation-model market publish the list of ways it preferences its own model over a competitor’s inside its own consumer products.

None of these four is radical. Each of them has a precedent in 20th-century antitrust. None of them will happen in Canada within the term of the current Parliament, because the political will does not exist and the lobbyists have already shown up.

The argument

The AI stack is not a market. It is a company town with one landlord and three mines, and every firm and every worker in the rest of the economy is a tenant in the shacks.

A tenant’s relationship to the landlord is not negotiation. It is rent. The rent has gone up every year since 2022. The rent will continue to go up until somebody regulates it.

The word for the regulation is antitrust. The word is not popular in a pitch meeting. Say it anyway.