The output side of the AI-compute economy — the companion to the Compute Tightness Index (which tracks the GPU input). The viral charts show the falling price of a token and conclude the AI trade is deflating. The number that actually settles it is demand — and demand is exploding: +93% since 2026-05-06, with open-source already carrying roughly half of all tracked inference tokens at a fraction of frontier price. A market getting bigger, not a margin getting crushed.
As of 2026-06-25, tracked AI inference runs at 45.3T tokens/week — up +84% in 30 days and +93% since 2026-05-06. Open-source carries 46.5% of those tokens, holding the demand-weighted price near $2.18/1M — a fraction of frontier list prices even as frontier models themselves got pricier.
⤓ Download full daily series (CSV)| Group | Blended $/1M | Tokens/week | Share | 30d demand |
|---|---|---|---|---|
| Open-source | $0.40 | 21.1T | 46.5% | ▲ +125% |
| Other | $0.93 | 10.2T | 22.5% | ▲ +266% |
| Claude | $10.00 | 6.6T | 14.5% | ▲ +19% |
| OpenAI | $3.47 | 3.4T | 7.4% | ▲ +22% |
| $0.85 | 4.0T | 8.8% | ▼ -3% | |
| xAI | $1.56 | 124.1B | 0.3% | ▲ +13% |
Open-source = aggregated weights-available labs (DeepSeek, Qwen, Llama, Mistral, etc.). "Other" = unclassified OpenRouter providers. Share = % of tracked tokens. *Implied spend = list price × tokens, not realized revenue.
Built from OpenRouter's model rankings + pricing, grouped into Claude / OpenAI / Google / xAI / Open-source / Other. Token counts are OpenRouter's trailing-7-day figures (the rankings "week" view) — i.e. each point is tokens processed over the prior week, not a single day. Per group: blended median $/1M (75/25 input/output) and weekly total_tokens. Demand index = latest weekly total ÷ the first week's total × 100. Effective $/1M = demand-weighted blended price (as-of prices carried forward across snapshots where only one side updated). Series since 2026-05-06; it grows as new snapshots land. Full history as CSV. Coverage is OpenRouter traffic — a large but partial slice of the market.
Pair it with the Compute Tightness Index (GPU input) for both sides of the AI-compute economy. Query the raw data from an agent via the SwiftAlerts MCP endpoint (get_inference_economics).