What AI compute costs, and who sets the price:from watts to tokens
Independent data and research on chips, data centers, cloud and model serving: the cost, price and pricing power at each layer. Every number carries a date and a source.
Green: data-center regions · Black: upstream chips and HBM · Arcs are illustrative
Research
Cost breakdowns, pricing mechanics, industry structure上游每个供应商的售价，就是 NVIDIA 的成本。以 B200 为例，零件和良率损失约 $6,400，摊上研发和管理费约 $1 万，售价却在 $3–4 万。
- 01B200 的制造成本约 $6,400，售价约 $3–4 万：制造成本只占售价约两成。
- 02制造成本里 HBM 最大，约占 45%，是跟踪 NVIDIA 成本变化最有效的指标。
- 03NVIDIA 定价靠供需而不是成本加成，零件涨价对售价的直接传导有限。
Sells for about $30–40k (industry reports, to be verified): manufacturing is about a fifth of the price.
2025-12 · Epoch AI · Estimate · Details →Coverage
Seven layers of the chain, from chips to developer tools- L01Chips & hardware1 reports1 datasets
- L02Servers & systemsComing
- L03Data centersComing
- L04Cloud & compute1 datasets
- L05Compute financeComing
- L06Model serving1 datasets
- L07Developer toolsComing
Data
Traceable and downloadable| Dataset | Region | Latest | Unit | Updated | Status | ||
|---|---|---|---|---|---|---|---|
| Open-model price basket | Global | 27% | $ / 1M tokens | Weekly | Partly unverified | CSV | |
| B300 cost per GPU-hour | United States | $2.80 | $ / GPU·h | As needed | Partly unverified | CSV | |
| GPU manufacturing cost breakdown | Global | $6,400 | $ / GPU | As needed | Estimate | CSV | |
| GPU rental prices (on-demand / reserved / term) | United States · Mainland China | — | $ / GPU·h | Weekly | Coming soon |
Markets
Equities, read through the supply chainEvery number is traceable
Every number has a date and a source; estimates show their working. Speculation, judgement and unverified items are flagged separately, so you can see what we know, what we don’t, and what we think.
The weekly compute brief
New research and price moves, once a week. Subscriptions open soon — follow via RSS for now.
Contact us
Research partnerships, data citations, press, or a different read on our analysis.