Fireworks says specialized models now generate 95% of its 40 trillion daily tokens

In its July funding announcement, Fireworks said its platform serves more than 40 trillion tokens a day. Over 95% of that traffic reportedly comes from models specialized for customers' proprietary data and specific jobs. Researcher Nathan Lambert drew fresh attention to the number around Fireworks' November Forge conference.

Forty trillion daily tokens give specialization production scale

Fireworks uses specialized model for systems adapted to customer data and particular tasks. That can include fine tuning, other post training or a model optimized for a narrow product job. The announcement does not break down the share of each technique or model.

Volume rose from 15 trillion tokens at the previous funding round to more than 40 trillion per day. The company also announced a $1.505 billion Series D, a $17.5 billion valuation and an annualized revenue run rate above $1 billion. All of these figures come from Fireworks.

Production teams buy task outcomes, not the winner of a general benchmark

A narrow model can be more useful when it completes one job faster, more cheaply and more predictably. Tool routing, code edits or summaries of company data do not always require the most expensive general model. They require stable performance on the team's own input distribution.

Value consequently moves toward evals, post training and inference infrastructure. The base model remains important, but a product's advantage may come from its dataset, feedback loop and ability to deploy several specialized variants economically.

One platform does not represent the whole market and sets its own definition

Fireworks makes money from serving and adapting open models. Its customers are therefore not a random sample of all AI users. Traffic from ChatGPT, Claude or Gemini is absent from the 40 trillion figure. It does not show that specialized models account for 95% of global inference.

There is also no independent audit of the volume or detailed public methodology. Tokens are not equal units of economic value: short classifications, long reasoning traces and batch jobs have different costs and benefits. Fireworks is reporting its own workload mix, not a universal verdict on frontier models.

Costs and deployed quality will determine the strength of the trend

Useful evidence would include public data from other inference platforms and customer evals before and after adaptation. Teams should track cost per completed task, latency, error rates and the ongoing expense of retraining. Token volume alone can rise because serving became cheaper.

If other providers report a similar mix and specialized models retain quality as input data shifts, this will amount to a change in market architecture. Until then, 95% is persuasive evidence about Fireworks customers, not a referendum of every company.

Lilith's verdict

Fireworks shows a factory where 95 out of 100 tokens head down a custom production line. Before treating it as a map of the whole industry, we need a view inside the other factories too.