What matters in AI today

Weekly volume went from 0.50T (week of Jan 6, 2025) to 172T (week of Oct 5, 2026): up 344× in 89 weeks.

The largest jump between two consecutive periods came in the week of Sep 28, 2026: +20T, from 146T to 166T. The curve climbs in steps, not in a straight line.

caveatThe latest period puts 2.6% of named-model volume on stealth/space-bunny-alpha, an anonymous model in testing. Traffic like this is transient and inflates the top of the ranking.

Comparing absolute volume across distant points calls for care: the source itself grew along with the market, and any share at the start is worth far fewer tokens.

The top 5 models went from 66.1% to 47.6% of volume between the weeks of Jan 6, 2025 and Oct 5, 2026. The model-level HHI went from 1447 to 917, which keeps it “unconcentrated” on the conventional antitrust scale.

Concentration peaked within the window at 66.5%, in the week of Jan 13, 2025.

In the latest period, the largest is DeepSeek V4.1 Flash, with 23.8% of volume on its own.

Routing implication: there are close substitutes at the top, so being locked into one specific model is a choice, not an inevitability.

Chinese labs went from 6.1% to 62.8% of volume; US and Canadian labs, from 82.8% to 25.1%, between the weeks of Jan 6, 2025 and Oct 5, 2026. The gap between the two is now 37.6 points, in favor of China.

Outside the two poles, the largest slice is other origins, with 2.0% (−1.4 pp over the window).

This is the headquarters of whoever trained the model. Where the providers serving it are based is in the providers card, further down.

Open-weights models went from 23.2% to 69.9% of volume; proprietary models, from 73.2% to 21.1%. 9.0% of the latest period has no identified license, and ignoring that band would inflate the other two.

At the end of the window, open weights and Chinese labs are 7.1 points apart (69.9% vs. 62.8%); at the start, the gap was 17.1. The curves converged over the window and now move almost together, because most relevant open-weights models are Chinese: it is the same phenomenon seen from two angles, not two independent trends.

Traffic on free endpoints went from 0.9% to 6.1% of volume over the window, peaking at 30.0% in the week of Mar 30, 2026.

Free volume inflates adoption without signaling willingness to pay. Every share on this page includes that subsidized demand, which has not yet been tested against price.

This is the headquarters of whoever serves the model, the inference provider that receives the request. That is a different thing from the origin chart, which shows the headquarters of whoever trained it: a Chinese model served by a US provider counts as China there and as the United States here.

56 of the 112 providers (50%) declare headquarters in the United States. 37 declare no headquarters at all, and for anyone with jurisdiction constraints this is the group that needs case-by-case checking.

Zero retention is a per-endpoint policy, not a per-model one: 62 of the 112 providers offer it on at least one endpoint, and the same model can have endpoints with and without it.