AI serving price models: GPU-hour, per-token, per-seat, outcome

An AI serving quote is a meter, and the meter trades a predictable bill for margin risk. GPU-hour pricing follows compute use, per-token pricing follows model output, per-seat pricing follows users, outcome pricing follows business results, and reserved capacity follows committed compute. Say a hypothetical workload uses 100 GPU-hours at $2 each and emits 20 million tokens at $0.50 per million; the GPU-hour bill is $200, the token bill is $10, and the difference is the pricing decision.

Capacity pricing keeps the cost curve honest

GPU-hour pricing is the easiest meter to audit because the invoice follows hardware time, the cost curve that determines margin. Fireworks AI offers developers managed GPU clusters billed on a usage-based model. Its annualized revenue recently exceeded $1 billion. The company reports processing more than 40 trillion tokens per day for customers. Fireworks AI closed a $1.5 billion Series D led by Atreides Management, Index Ventures, and TCV.

Capacity meters matter because the vendor's cost moves with the same hardware time that the buyer is charged. The buyer still needs a utilization floor before accepting a rate card.

A usage-based cluster bill is easier to model than a blended platform fee because the input is visible. The team can forecast GPU-hours from request rate, model size, and batch size, then compare that forecast against the vendor's rate card. The forecast should include idle time, not just peak demand.

The meters need the same test

Use the rows as a pre-sales checklist. The meter should be visible in the invoice, not hidden in a blended rate.

• GPU-hour: cost driver: hardware time; margin test: utilization; scale risk: idle capacity; works: spiky inference; breaks: low utilization.
• Per-token: cost driver: output volume; margin test: tokens per request; scale risk: longer prompts; works: variable chat; breaks: verbose models.
• Per-seat: cost driver: active users; margin test: cost per user; scale risk: power users; works: predictable teams; breaks: uneven usage.
• Outcome: cost driver: paid results; margin test: conversion rate; scale risk: model gains; works: measurable ROI; breaks: fuzzy attribution.
• Reserved capacity: cost driver: committed compute; margin test: committed utilization; scale risk: demand misses; works: steady load; breaks: unpredictable growth.

Hardware changes move the cost curve

General Compute obtained a $400 million loan from Upper90. The company states that its new chips will run inference 16 times faster than GPU-based clouds. Crusoe raised a $3 billion Series F that lifted its valuation to $30 billion in under a year.

When inference hardware gets faster, the cost per token can fall even if the GPU-hour price stays flat. A token-priced vendor may pass that gain to the buyer; a GPU-hour vendor may keep it. The buyer should ask which meter the discount follows, because a token price that protected margin at one utilization level can become a discount at another.

The right meter matches the load

For an engineering leader buying serving capacity, start with GPU-hour or reserved capacity when the load is steady and the model is fixed. Per-token is the right meter when output length is the main variable. Per-seat works for flat user behavior. Outcome pricing works when the buyer can verify the result and the vendor can absorb the inference cost.

For a technical founder pricing your own stack, anchor the public rate to the cost meter you can measure. If you sell tokens, track tokens per dollar of GPU time. Seat pricing needs cost per active user. Outcome pricing needs inference cost per paid outcome. The price that survives is the one that leaves margin when utilization falls.

Before signing, put the meter and the utilization floor in the same contract clause: the unit, the rounding rule, the minimum, and the floor. If the vendor can't state the unit in plain language, the invoice becomes a negotiation. The rounding rule matters too: billing whole tokens, whole seconds, or whole hours changes the effective price.