Spending on compute is rising fast. GPU list prices have not moved in 30 days. Frontier AI output prices span 360x, lowest to highest. No tracked compute future has listed yet.
Capital spending by the five largest buyers of compute in the last 12 months, 1.9x a year earlier.
Capital spending, last 12 months
The five largest buyers
Bottom line
What the latest filings and price feeds show. Click any card for the evidence.
How compute works, in five steps
From the chip to the answer. Click a step to see its data.
Who is spending
Every figure from the company's own filings.
League table, last 12 months
Share of the five
What a million tokens costs
The newest of each maker's main models, on each cloud's own public list. A maker's direct price may differ.
Frontier model prices
| Model | Maker | Cloud | Input | Output |
|---|---|---|---|---|
| OpenAI | Azure | $33.00 | $198 | |
| OpenAI | Azure | $11.00 to $12.00 | $55.00 to $60.00 | |
| Anthropic | AWS | $11.00 | $55.00 | |
| Anthropic | AWS | $11.00 | $55.00 | |
| OpenAI | AWS | $11.00 | $55.00 | |
| Anthropic | AWS | $4.40 | $22.00 | |
| Moonshot AI | AWS | $3.30 | $16.50 | |
| Cohere | AWS | $3.00 | $15.00 | |
| Amazon | AWS | $2.50 | $12.50 | |
| $2.00 | $12.00 | |||
| OpenAI | Azure | $2.20 to $2.40 | $11.00 to $12.00 | |
| Anthropic | AWS | $2.20 | $11.00 | |
| $2.00 | $10.00 | |||
| xAI | AWS | $2.20 | $6.60 | |
| OpenAI | Azure | $1.10 to $1.20 | $5.50 to $6.00 | |
| $0.75 | $3.75 | |||
| Z AI | AWS | $1.00 to $1.20 | $3.20 to $3.84 | |
| DeepSeek | AWS | $0.62 to $0.74 | $1.85 to $2.22 | |
| Mistral | AWS | $0.50 to $0.61 | $1.50 to $1.82 | |
| MiniMax | AWS | $0.30 to $0.36 | $1.20 to $1.44 | |
| Meta | AWS | $0.24 | $0.97 | |
| Qwen | AWS | $0.22 to $0.29 | $0.88 to $1.16 | |
| Nvidia | AWS | $0.15 to $0.18 | $0.65 to $0.78 | |
| OpenAI | Azure | $0.11 to $0.12 | $0.55 to $0.60 | |
| Anthropic | AWS | $0.11 | $0.55 |
Output price by maker
Lowest to highest frontier model, per million. Log scale.
What compute costs
List prices per GPU-hour, one provider line at a time, never averaged.
Price against memory
Lowest on-demand price against memory per GPU. Bubble size: live lines.
Price range by GPU
Bar: lowest to highest on-demand. White dot: lowest Spot.
Who sells it, and where
Tiles sized by the number of live prices.
Who is listed
GPU lines by country
Capital spending is each company's cash paid for property and equipment, last 12 months, from its own SEC filings. It covers all capital spending, not only AI. Sales are total company sales. Signed backlog is each filer's reported remaining performance obligations, by its own definition. A filing stops where the latest one does, so periods differ by company. Prices are each provider's own list price, one line at a time. Nothing here is a DAP price, an index or investment advice.