Open models are growing faster than the market can measure them

Nathan Lambert of Interconnects has tried to estimate the size of the open model inference market. He relies mainly on public figures from Baseten, Fireworks and Together AI, plus the open model share of OpenRouter traffic. Hyperscale clouds such as Amazon Bedrock and Azure remain partly invisible because they do not publish usage by model category.

One fresh signal comes from Fireworks. At its conference, the company said customer customized models serve more than 40 trillion tokens per day and that this slice has grown by more than 2 times since the summer. Interconnects assembles public disclosures into a time series that Lambert describes as close to exponential growth.

Most of the primary article sits behind a paywall. Its public section explains the method and those figures, but does not expose the full chart or Lambert's final comparison with Google. Any conclusion must therefore remain narrower than the source headline's enticing question.

Four public signals stand in for a market with no common ledger

Interconnects uses inference providers as proxies for the segment. Lambert argues that their fortunes are more tightly linked to open models than those of large clouds, which sell many other services. OpenRouter adds a view of token volume actually routed to open models.

The method provides a trend, not a set of audited accounts. Each platform discloses a different slice and may count tokens, API traffic, caching or customized models differently. Even the Fireworks figure covers one part of its inference rather than the company's entire load.

Buyers now have a meaningful alternative to a handful of closed APIs

More than 40 trillion daily tokens suggests customized models have moved beyond laboratory experiments. Companies are paying to run models adapted to their tasks, creating demand for hosting, fine-tuning, routing and performance management.

Product teams gain the possibility of changing a model or provider without rebuilding the whole product. Investors and regulators should also note that substantial AI activity may flow through layers missing from the largest labs' statistics.

Token volume measures activity, not value or reliable market share

Trillions of tokens sound more precise than they are. A long agent run can consume far more tokens than a short chat without generating more revenue or utility. Comparisons require prices, cache behavior, context lengths and workload mix.

The disclosures also come from companies motivated to demonstrate growth. Without a shared method and data from Amazon, Microsoft and other clouds, the available figures cannot establish open models' share against OpenAI or Anthropic. They show only that the measured portion is expanding quickly.

Common reporting and cloud breakdowns would turn the trend into a market measure

The next useful step is not a larger number from one vendor but comparable reporting. Providers would need to separate open and closed models, customized and base variants, and state whether input, output and cached tokens are counted consistently.

Hyperscalers provide the other missing signal. If Amazon, Microsoft or Google begin reporting model usage by category, public estimates can be calibrated. Until then, the chart is a useful directional radar rather than a map of market size.

Lilith's verdict

Trillions of tokens flash around the stadium, but every provider runs its own scoreboard. Growth is visible. Those numbers still cannot honestly rank the runners.