Liminal - AI TokenomicsUS vs China Model Flows, Capability Gap, and Price Premium
FrameworkWhere AI model spend is flowing, how the capability gap is shifting, and what buyers pay per unit of intelligence.
As of … · last settled week …
Market Share

US vs China share of top-model activity

Share of named top-model tokens (left) and revenue proxy (right) by country of origin, weekly. Revenue is a list-price proxy at the 70 / 30 in-out blend; realized revenue could be lower due to caching, batching, and enterprise discounts. Faint lines show raw values; solid lines are a 3-week centred moving average.

Token share

Revenue-proxy share

Absolute Levels

US vs China absolute tokens and revenue

Weekly totals spent. Tokens on the left; list-price revenue on the right.

Tokens per week

Revenue per week ($)

Capability Gap

How far ahead is the US frontier?

Frontier is defined as the maximum Epoch ECI score released by any lab in each country up to that month. Left panel is the gap in ECI points. Right panel translates the gap into US lead in months (ie how many months earlier the US frontier reached today's Chinese level).

US − CN gap (ECI points)

US lead (months)

Price Premium

Relative value of different models across two regions

For each recent model we regress log price on ECI capability, then look at the residual, the price paid above (or below) what a model of that capability would predict. Left panel is the latest 12-month cross-section; right panel is the quarterly average residual by origin. The slope of the hedonic line is roughly …% per ECI point.

Latest 12-month cross-section

Quarterly residual over time

Sources. Weekly token and revenue data come from OpenRouter public rankings, restricted to weeks with all seven days reported and settled for at least seven days. Capability scores come from the Epoch AI Capabilities Index (ECI), cross-checked against Artificial Analysis. Country of origin is assigned by developer lab. Refreshed monthly.
Caveat on scale. OpenRouter is a routing platform that captures only a fraction of total model spend, not the whole market. First-party APIs (OpenAI, Anthropic, Google) and hyperscaler deals (Azure, AWS Bedrock, GCP Vertex) are much larger and not visible here. Use the OpenRouter numbers as a directional read on relative flows and trends, not as a market-size estimate. Revenue is a list-price proxy at the 70 / 30 in-out blend; true realized revenue is lower because of caching, batching, and enterprise discounts.
Disclaimer. This tool presents output from a research framework published by Liminal Capital. Figures are aggregations of public model-usage data and should not be interpreted as forecasts or guarantees. This is not investment advice, not a recommendation, and not an offer or solicitation of any investment product or service. Past performance, modelled or otherwise, does not guarantee future results.