The Rent Is Too Damn Inference

A landlord does not produce housing. A landlord owns housing and charges rent for access to it. The rent is set by the landlord. The rent goes up. If you do not like the rent, you can move, but all the other housing is owned by three other landlords who talk to each other at the same conferences.

This is the structure of the generative AI economy in 2026. Every prompt is a rent cheque. Every API call is a tenancy. The landlords are three American firms, and the rent is going up.

Per-token pricing is a rent, not a price

A price is what you pay for a unit of a thing that has a cost. A rent is what you pay for access to a thing you do not own.

Per-token pricing looks like a price. It is presented as a price. Tenth of a cent per thousand input tokens, quarter of a cent per thousand output tokens, all very tidy, all very engineering-blog-friendly. Do the math and the unit cost is low. So low you do not haggle. So low you do not even notice you are spending it.

Per-token pricing is a rent. Consider the actual cost structure.

The marginal cost of one more inference on an already-loaded model, on a GPU that is already provisioned, in a data centre whose electricity is already billed, is a small fraction of a cent. The price is set by what the landlord thinks the tenant will pay, not by what the computation costs. The gross margin on frontier-model inference is somewhere between 40 and 80 per cent depending on the vendor and the tier. This is a landlord’s margin.

The landlord charges rent because the landlord can. The computation is cheap. The access is rationed. You pay for the access.

The enshittification curve applies

Cory Doctorow’s law of enshittification holds that a platform is good to its users until it has captured them, then good to its business customers until it has captured them, then extracts value from both in favour of the shareholders until the platform dies. Write the law on a whiteboard. It applies without modification to foundation-model APIs.

Phase one, roughly 2020 to 2023. The APIs were cheap, generous, lightly rate-limited, documented with earnest care. OpenAI ran a free tier that small developers built entire products on. Anthropic shipped long-context windows at prices that did not cover the compute. The labs were building an audience. The audience was being captured.

Phase two, roughly 2023 to 2025. The APIs became the substrate for a billion-dollar app layer. Every start-up, every enterprise, every integration built against the same four or five endpoints. Switching vendors became a porting project, not a config change. The audience was now a tenant base. The vendors raised prices quietly, changed tiering aggressively, and added acceptable-use restrictions that removed whole categories of previously-allowed use.

Phase three, already underway in 2026. The labs introduce opaque “fair use” throttles on high-volume accounts. The labs deprecate cheap model tiers on short notice. The labs roll out “enterprise agreements” that replace the transparent per-token menu with negotiated pricing you cannot compare. The labs launch first-party consumer products that compete directly with the start-ups that built on their APIs, at terms the start-ups cannot match, because the labs pay no API fee to themselves.

This is the enshittification curve. The AI APIs are approximately where Facebook’s developer platform was in 2014. The direction is known. The floor is not yet in sight.

”But switching cost is low”

A standard defence offered by the vendors is that switching is easy. The API is an OpenAI-compatible JSON schema. The same schema is now offered by Anthropic, Google, Mistral, Cohere, and a dozen open-weight hosts. Point your client at a different URL; keep your code.

This is true for a demo. It is a lie for a product.

The prompts are different. A prompt tuned for Claude does not perform identically on GPT-4o. Fine-tunes are not portable across vendors. Tool-use schemas are vendor-specific in the details that matter. Rate-limit behaviour is different. Safety-filter behaviour is different. Latency profiles are different. The edge cases that surface in production on one vendor do not surface on another.

A serious production system built on a foundation-model API is, in practice, calibrated to that vendor. Moving vendors is a multi-month engineering effort with real regression risk. The apparent portability of the JSON schema conceals the real immobility of the prompts, the fine-tunes, the evaluations, and the institutional knowledge the team has built around a specific model’s failure modes.

This is the exact same trick that made “cloud portability” a lie for the last decade. The APIs are standard-ish. The behaviour is not. The switching cost is high. The landlords know it.

The Canadian consumer has no leverage

If you are a Canadian consumer paying $20 a month for a chatbot subscription, you are a tenant three times over. You are renting the model from OpenAI. OpenAI is renting the compute from Microsoft. Microsoft is renting the silicon from NVIDIA. Each layer extracts a margin. The margin flows south, out of the Canadian economy, to the same zip codes it has been flowing to since the 1990s.

Canada’s total consumer spend on AI subscriptions in 2026 is on track to exceed a billion dollars. The total domestic capture of that spend — Canadian firms, Canadian wages, Canadian taxes — is an order of magnitude smaller. We are paying for a utility we do not produce and do not regulate, with the proceeds flowing to a foreign oligopoly.

The 20th-century analogue is long-distance telephony pre-1997, when Bell Canada was a utility with a domestic monopoly but the infrastructure was overwhelmingly foreign-owned at the upstream and the tariffs were set abroad. The fix in telephony was structural. CRTC regulation. Rate-setting. A Canadian-content calculation. A universal-service fund.

None of that regulatory apparatus exists for AI. The CRTC does not regulate chatbots. Nobody regulates chatbots. The price is the price.

The price is going up

Every public statement from the major labs in 2025 and 2026 telegraphs that inference prices are at the bottom of the curve. The vendors have been explicit. The current generation of frontier models is being offered below cost to capture enterprise workloads. The next generation will be priced higher. “Premium tiers” have already launched at multiple of the current cost. “Agentic” workloads, which consume tens of thousands of tokens per task, will be priced at premium tiers.

The per-token rate is a loss leader. The target is the enterprise contract. The enterprise contract is a lock-in. Once the lock-in is set, the rent goes up.

A Canadian firm signing a five-year AI contract today is signing a rent agreement with price review clauses that favour the landlord. A Canadian firm declining to sign is simply choosing to lose the workflow to a competitor who signs. The threat is credible because the labs have successfully made AI adjacent to table stakes in every knowledge-work industry.

This is how rent economies work. There is no escape at the individual firm level. The escape, if there is one, is regulatory.

The argument

The generative AI economy, as currently structured, is a rent economy. The landlords are three firms. The rent is per-token pricing and enterprise lock-in. The rent is at its floor. The rent is going up.

Every other economic story about AI — the productivity gains, the consumer surplus, the “democratization of intelligence” — is a narrative layer on top of a rent-extraction layer. The productivity gains are real. The rent extraction is also real. The net depends on which side of the ledger you sit on, and in Canada’s case the ledger is clearly negative.

A consumer-protection politics of AI would start by treating inference pricing the way we treat electricity pricing in a regulated province. Transparent tariffs. Rate caps during transition periods. Mandatory disclosure of cost structure. A consumer advocate at the rate hearing.

Nobody in Ottawa has proposed any of this. The rent continues. The cheques go south.