Who Actually Pays for LLM Inference?

01 The Layer Everyone Uses and Nobody Prices

Every chat message, coding completion, and agent action in 2026 ends at the same place: a GPU executing a forward pass, billed in fractions of a cent. Inference is the only part of the AI stack consumers touch daily, and it is the part with the least honest price discovery. Training costs make headlines; inference costs are absorbed, cross-subsidized, and increasingly, lost on purpose. The question that decides which AI companies survive is not who can train a frontier model — it is who can keep paying for every token the world runs through it.

The structure of the service layer explains why the pricing looks irrational. A model lab sells tokens; a cloud rents the GPUs the tokens run on; a startup rents the cloud; an application burns the tokens and charges a subscription that assumes usage patterns which agentic products then shatter. Each layer marks up or marks down depending on strategy, and the losses land somewhere specific.

02 Why Free Tiers Exist: The Loss-Leader Ledger

Free and flat-rate consumer tiers are the most visible subsidy in tech. A heavy ChatGPT Plus user on a reasoning model can plausibly consume several dollars of raw compute in a day of agentic use against a $20 monthly fee — the subscription survives only because the median user consumes far less. Flat-rate plans are actuarial products: the price is set by the silent majority subsidizing the power users, and every capability upgrade (longer context, more tool calls, deeper reasoning) shifts the actuarial table toward the red.

Labs run these tiers deliberately. Inference at a loss buys the two assets that matter: distribution and preference. Every habit formed on a free tier is a future enterprise contract's proof-of-concept, and the cost is booked, in effect, as customer acquisition — at some of the highest acquisition costs in software history.

03 Who Gets Squeezed First

The service layer's casualty list is predictable from its balance sheets. Standalone inference providers — the companies renting GPUs wholesale and selling tokens retail with no consumer brand — compete in a market where the labs themselves price tokens as marketing. When the model owner can sell at marginal cost to win developers, the reseller in the middle has no pricing power at all. The 2025-26 wave of inference-provider consolidation followed directly.

Agentic application startups occupy the next-worst position: their product design multiplies token consumption (an agent loop can burn 50-100x the tokens of a single chat exchange) while their subscription pricing assumes chat-era usage. They pay the labs, absorb the volatility, and face users who notice only the monthly fee. When token prices fall 10x, their costs improve — but usage multiplies faster, a dynamic closer to Jevons paradox than to a windfall.

04 The GPU Financing Beneath It All

Beneath every token price sits a capital structure. Accelerators are financed over three-to-five-year depreciation schedules against revenue streams that reprice downward every year. Token prices for equivalent capability have fallen roughly an order of magnitude per year for several years — a deflation that would be catastrophic in any capital-intensive industry whose assets held fixed cost. The equilibrium holds only while demand grows faster than prices fall: utilization, not price, is the whole game. GPU cloud economics in 2026 are a race to keep accelerators busy enough that the deflation in token prices is offset by volume.

This is why the biggest labs build their own silicon and sign multi-year compute commitments. Vertical integration converts the largest variable cost (rented compute) into a capital structure they control, and scale commitments convert deflation into a bargaining chip with foundries and fabs. The players without that leverage — the ones renting at list price — sit at the top of the casualty list above.

05 Limits of the Who-Pays Story

Honesty requires the countercase: deflation in inference has real beneficiaries, and the squeeze is partly a choice. Application businesses that price in usage rather than flat subscriptions are thriving; open-weight models deployed on owned hardware give large enterprises a ceiling on what they will ever pay per token; and the labs' losses are strategic investments in a winner-take-most market, not structural necessity. There is also a technology escape valve — speculative decoding, batching, KV-cache optimization, and distillation have delivered much of the price collapse, and each improvement is genuine productivity, not subsidy.

The limit of the analysis is that it treats "the AI industry" as one balance sheet. It is not: the losses are concentrated in the consumer-facing service layer and the leveraged GPU renters, while the picks-and-shovels layer — fabs, power, networking — has been paid in full throughout.

06 What to Watch

Three markers will show who is really paying. First, whether consumer subscription prices move for the first time in the ChatGPT era — the clearest signal that actuarial subsidies have hit their limit. Second, whether agentic products migrate to usage-based pricing, which would be an admission that flat-rate and agents cannot coexist. Third, the next round of inference-provider failures or acquisitions: each one redraws the map of who bears compute risk. The price of intelligence keeps falling. The number of companies that can afford to sell it at that price is falling with it.