Creek Drive ResearchPublic AI usage telemetry
AI infrastructure dashboard

Token demand is observable at the edges—not at the center.

The database directly meters one important router, then triangulates the missing market with disclosures and external models. Use the trend and relative-strength views; treat the global and U.S. capture rates as ranges, not audited market share.

Exact, estimated, and proxy data kept separate

What share do we actually see?

Primary definition: continuous, directly metered tokens in the database divided by a modeled worldwide or U.S. denominator. The disclosure ledger is shown separately because it is broad but stale and asynchronous.

Modeled denominator
World direct capture
U.S. direct capture
Router 7-day average
OpenRouter public-platform tokens/day
Disclosure representation

Current world denominator

Tokens Per Day's July 16 six-channel band, nowcast to the latest OpenRouter date at its published 2.2× annual growth assumption.

Sources: Tokens Per Day estimates for the modeled denominator; OpenRouter Data API for the metered router numerator. Through .

Coverage arithmetic

The base estimate uses a smoothed seven-day router average, not a volatile single day.

Worldwide
United States
Sources: Tokens Per Day methodology and OpenRouter's reported ~47% U.S. billing-location share. Geography remains modeled, not directly metered.
Direct capture ≠ represented volumeOpenRouter is a continuously observed router. The 300.1T/day disclosure floor describes a larger set of entities, but its dates and scopes do not line up.
U.S. is the weakest estimateNo source reports U.S. tokens directly. The base uses the site's ~47% OpenRouter billing-location share; the range allows 35–55% market and 40–55% router shares.
Provider domicile is not usage geographyGoogle, OpenAI, Anthropic, and Microsoft are U.S. companies serving worldwide demand. Their global output cannot be counted as U.S. consumption.

Relative strength by model lab

Volume share and momentum within OpenRouter—not global lab market share. Free routes, app mix, and new-model launches can move this ranking quickly.

Platform-relative
Leading lab
Top-three labs
Combined latest 7-day share
Labs tracked
Named providers in latest window
Other / long tail
API aggregate, not attributable by lab

Monthly average daily tokens by leading lab

Top eight labs by current volume; all others bucketed as Rest. Monthly averages keep partial August comparable with full months.

Source: OpenRouter Data API — daily model rankings. Labs parsed from model slugs; public-platform scope only; through .

Current lab leaderboard

Sorted by latest seven-day token volume.

#Lab7d tokensShareWoW30d tokens
Source: OpenRouter Data API — daily model rankings. Seven-day windows end .

Model strength and indicative list-price economics

Token volume provides adoption momentum; current list price provides a rough revenue-intensity axis. The cost math is not realized billing because prompt/output mix, caching, discounts, and historical prices are unavailable.

Cost is indicative

Volume versus blended list price

Top 30 models; price assumes 75% input / 25% output and current OpenRouter list prices. Free models appear at zero.

Sources: OpenRouter Data API for seven-day token volume and OpenRouter Models API for current list prices. Through .

Frontier versus budget cost index

MyTokenTracker fixed baskets, $/million tokens, 3:1 input:output. The public history is short and currently flat.

Source: MyTokenTracker AI Cost Index, CC BY 4.0. Fixed frontier and budget baskets; 3:1 input:output blend.

Current model leaderboard

Indicative cost applies today's blended list price to all seven-day tokens. It is a comparable intensity measure, not actual provider revenue.

#ModelLab7d tokensShareWoWBlended $/MIndicative 7d list cost
Sources: OpenRouter Data API and OpenRouter Models API. Indicative cost uses current list price, 75% input / 25% output, and excludes caching or discounts.

Where the tokens are going

Publicly attributed applications and sampled weekly task filters provide a directional demand map. Neither is a complete application census.

Biased sample

Top public apps—trailing 30 days

OpenRouter public attribution; trillions of tokens.

Source: OpenRouter Data API — top apps. Publicly attributed app traffic only; trailing 30 days through .

Task-filter token estimates

Latest sampled/upsampled week. Filters should not be summed into a market total because labeling may overlap.

Source: OpenRouter Data API — category-filtered rankings. Sampled/upsampled estimates for the latest completed week.
Public attribution onlyPrivate, hidden, and zero-data-retention applications are excluded from the app ranking.
Tasks are directionalUse category levels and changes as signals. Do not treat category totals as mutually exclusive without a classifier audit.
Router customer mix mattersOpenRouter overweights developers, agents, experimentation, and open-model discovery relative to consumer chat and enterprise contracts.

Definitions, assumptions, and sources

The dashboard separates metered router traffic, modeled global demand, disclosure representation, sampled classification, and cost proxies.

Capture-rate definitions

  1. Worldwide direct capture: latest seven-day average OpenRouter public-platform tokens divided by the nowcast Tokens Per Day global low/mid/high band.
  2. U.S. direct capture: a modeled OpenRouter U.S. numerator divided by modeled U.S. demand. Base case applies 47% to both; the range allows router geography and market geography to differ.
  3. Disclosure representation: the curated 300.1T/day supply/national floor divided by the global modeled denominator. It is broad but asynchronous and should not be called live capture.

Key assumptions

  • The July 16 global band is carried to the OpenRouter cutoff using Tokens Per Day's published 2.2× annual growth assumption.
  • China remains 140T/day in the U.S. bridge; the base U.S. share is 47% of non-China usage. This is a low-confidence OpenRouter billing-location proxy.
  • OpenRouter token definitions are accepted as reported by upstream providers; tokenizer and hidden-reasoning treatment are not harmonized.
  • Current model prices are blended 75% input / 25% output. Cache discounts, free-route subsidies, enterprise discounts, and historical price changes are excluded.
  • Router, supply, and demand ledgers are never summed. Missing dates remain missing. No world total is inferred from row counts.

Source ledger

What is missing

Continuous first-party ChatGPT, Claude, Gemini, Meta AI, and Grok app usage; direct enterprise lab APIs; AWS Bedrock, Azure, and Vertex totals; private/ZDR routing; self-hosted open-weight inference; and reliable end-user geography. These omissions are why the direct worldwide estimate is only a few percent even though point disclosures represent much more of the modeled volume.