About this index
Tokens Per Day tracks how many AI tokens the world processes each day, a leading indicator of AI inference demand, the way barrels/day track oil or terawatt-hours track electricity.
FAQ
Is this an official number?
No. There is no IEA- or USGS-equivalent authority for token consumption. This is an independent, open estimate. The headline is a reported floor (real usage is higher) plus a clearly-marked estimated band.
Why “floor” instead of a single total?
Because a confident single number would be wrong. Summing only verifiable, non-overlapping disclosures gives a defensible lower bound that can’t be inflated by double-counting or guesswork. Estimates sit above it, explicitly labeled.
What counts as a “token”?
It varies by source: input vs output vs cached tokens are counted differently, and multimodal (image/video/audio) tokens inflate counts. We preserve what each source reported rather than forcing a single definition, and flag this as a known limitation.
How often is it updated?
On a manual, dated pass, each observation carries a verified_date. The figures
move fast (Google’s throughput grew ~7× in a year), so treat older quarters as snapshots.
Data last verified 2026-07-16.
Can I use the data?
Yes. It’s all in the open repository as CSVs. If you cite it, please link back and note it’s a reported floor plus estimates.
Sources
Every disclosure in the reported floor, with its archived snapshot.
| Source | Link | Archive |
|---|---|---|
| Google I/O 2025 (restated in Alphabet Q2 2025 CEO remarks) | original | snapshot |
| Sundar Pichai at Google I/O 2026 | original | snapshot |
| Microsoft FY2025 Q3 earnings (50T in March) | original | pending |
| Microsoft FY2025 Q4 earnings call | original | snapshot |
| Sam Altman via Epoch AI usage dataset | original | pending |
| OpenRouter/a16z State of AI 2025 | original | snapshot |
| OpenAI statistics (15B tokens/min) | original | snapshot |
| Fireworks Series C announcement | original | pending |
| Fast Company (Lin Qiao interview) | original | snapshot |
| Fireworks AI homepage | original | pending |
| Tomasz Tunguz (The Token Race) | original | pending |
| National Data Bureau (Liu Liehong) via China News Service | original | snapshot |
| Volcano Engine via Robonomics Token Tracker | original | pending |
| Volcano Engine via China Daily | original | snapshot |
| Volcano Engine (Tan Dai) via KuCoin | original | pending |
| Business Insider (CEO Weinberg) via AOL | original | snapshot |
| Alphabet Q3 2025 CEO remarks | original | pending |
| DeepSeek Open Source Week Day 6: V3/R1 inference system overview | original | pending |
| Microsoft FY2026 Q3 earnings (Nadella) | original | pending |
| OpenRouter on X: Qwen3.6-Plus first model past 1T tokens in a day | original | pending |
| Fireworks Virtual Cloud GA blog | original | pending |
Limitations we’re open about
- Major providers (Anthropic, Meta, xAI, Amazon) don’t disclose token counts, handled as estimates or gaps, not invented.
- The reported series is coverage-limited in earlier quarters and widens over time.
- Token definitions are inconsistent across sources.
- Estimates are order-of-magnitude, meant to be refined as better anchors appear.