Estimates

Implied estimates, above the floor

The reported floor (300.1 T/day) counts only disclosures. These estimates add the big names that don’t publish usable token numbers, bringing the estimated total to roughly 329.8 T/day (315.5–359.3). They are deliberately low-confidence and meant to be refined.

Entity	Method	Low	Mid	High	Confidence	How it’s calculated	Source
Anthropic all (no token disclosure exists)	revenue-implied	10.3	16.4	27.4	low	~$30B annualized run-rate (Apr 2026) / $3-8 per Mtok blended	source
OpenAI consumer ChatGPT only (API 21.6T already in reported floor)	usage-implied	4	10	25	low	2.5B prompts/day (Feb 2026) x 1.5-10k tokens/round incl context/system/output	source · archive
xAI Grok (X consumer + API)	usage-implied	0.1	0.3	0.8	very low	~7-10M DAU x 5-20 msgs x 1.5-3k tokens; revenue ~$0.5B mostly free so revenue-implied fails	source · archive
Meta Meta AI assistant across WhatsApp/IG/FB	usage-implied	1	3	6	very low	~1B MAU low-intensity x 1-3 msgs x 1-2k tokens; excludes recommender/ranking inference	unpinned

Why two different methods?

revenue-implied

For usage-billed providers

tokens = revenue ÷ blended $ per million tokens. Works when revenue is mostly metered token usage (e.g. Anthropic). Fails for flat-rate or free products, where heavy users consume far more than per-token revenue implies.

usage-implied

For free / consumer surfaces

tokens = users × messages/day × tokens/message. Used where the driver is usage, not billing: consumer ChatGPT (anchored to ~2.5B prompts/day), Grok on X, Meta AI.

No double-counting. Estimates are additive to the floor; none duplicate a floor row. The OpenAI estimate, for instance, is consumer only; OpenAI’s API tokens are already counted in the reported floor.

Still a gap, not an estimate. Some names (e.g. Amazon Bedrock) disclose no aggregate tokens and have no clean anchor, so they are omitted rather than guessed. Honesty about what we can’t measure is part of the method.