The Subsidy Cliff

OpenAI, Anthropic, and Google are not pricing their API products to make money. They are pricing them to acquire users. The gap between what it costs to serve a GPT-4 or Claude Sonnet inference and what they charge for it is funded by investors — not by the economics of the product itself. This is not a secret. It is the stated strategy, legible in every fundraising round and every pricing announcement.

The numbers are stark. Anthropic reportedly loses money on a significant portion of API usage even after aggressive infrastructure optimisation. OpenAI's compute costs dwarf its revenue. Google can cross-subsidise from search for now, but that business is under its own pressure. The companies burning the most cash are the ones setting the price benchmarks the market compares against.

This matters because the discount is not permanent. Investor patience is finite. The current interest rate environment makes 'grow at all costs' a harder story to sell than it was in 2021. At some point — whether through direct price increases, metered rate limits, tier restructuring, or the quiet removal of features from lower plans — the subsidy will be reduced. The cost of running frontier inference will not go down enough fast enough to compensate.

We have already seen the early signals. Anthropic removed code execution from the free tier. OpenAI has progressively tiered features behind higher-spend accounts. The direction of travel is clear even if the pace is uncertain.

Local inference is structurally immune to this dynamic. When you run a model on your own hardware, the cost is denominated in electricity and the amortised cost of the machine you already own. There is no per-token charge. There is no subscription. There is no pricing team at a startup deciding whether to reset your rate limits based on your monthly spend. The cost is flat, predictable, and does not scale with usage.

The models you can run locally today are not compromises. Qwen3 14B, Gemma 4 9B, and Mistral Small 3.1 are competitive with the models that were considered frontier two years ago. The quality of open-weight models improves every quarter. The gap between what you can run locally and what requires a cloud API is narrowing. The economic gap between local and cloud inference is likely to widen.

None of this means cloud AI is going away, or that it is a bad choice for every use case. But the argument that cloud is cheaper is on borrowed time. The price you pay today for a GPT-4 class API call is subsidised by capital that has a finite patience. Plan accordingly.