Fireworks serves 40 trillion tokens a day on custom models — double the level of last summer

The conventional wisdom is that open models are a niche pursuit for enthusiasts, while the real money and traffic live at OpenAI and Anthropic. Nathan Lambert, author of Interconnects, decided to test that against the numbers. The first post in his new paid series on how large the open-model economy really is asks a simple question: is the open-model inference market growing, or is it a statistical rounding error next to the closed labs?

Where the numbers come from

There are no clean measurements here, and Lambert admits as much. What exists are noisy platform metrics, startups' marketing figures and imperfect public benchmarks. So he uses as reference points three companies whose fortunes depend directly on open models: Baseten, Fireworks and Together AI. To these he adds the share of open models in OpenRouter's tokens, which the service exposes through its API.

Hyperscalers such as Amazon Bedrock and Azure most likely occupy a large place here too. But they do not disclose how much of their traffic goes to open models specifically, so they were left out of the sample.

The key new figure comes from Fireworks. At its first conference, the company said it processes more than 40 trillion tokens a day for customers' custom models. That is just one segment of its inference business, and it has more than doubled since summer. Lambert expects Fireworks' total volume to come in well above 100 trillion tokens a day by the next update.

How to compare this with closed models

Lambert's first instinct is to anchor these numbers to something known. Google is the only major player that publishes tokens across its entire platform and API in investor reports. So he overlays Google's data on the open-platform charts, on a logarithmic scale. The result of the comparison itself sits behind the paywall, so we don't know how close the open segment has come to Google.

But some things are visible even without it. On a log scale, the startups' public data looks almost like a straight line going up, which means exponential growth. According to Lambert, this accelerated even further after the release of GLM 5.2, because open models have markedly caught up in agentic tasks. And agents consume orders of magnitude more tokens than ordinary chat: a single request turns into a long chain of tool calls and repeated calls to the model.

What to do with this

For practitioners, the takeaway is not that open models have overtaken closed ones. It is something else: the infrastructure for them has become a full-fledged market, where volumes are measured in tens of trillions of tokens per day. That means several things.

• If you are building a token-hungry agentic product, you have a real alternative to closed APIs. Calculate the cost per million tokens on Fireworks, Together or Baseten for your workload and compare it with your current bill.
• Customization is becoming the main argument. The growth of Fireworks' custom-model segment specifically suggests that customers are fine-tuning and adapting open models to their needs rather than simply taking them off the shelf.
• Lock-in to a single vendor costs more than it used to. If your code is tied to one API format, keep an abstraction layer over providers.

Limitations

It is worth keeping the boundaries of this data in mind. Startup figures are marketing numbers that nobody audits independently, and "tokens per day" is not the same as revenue. Tokens from different models are priced very differently, and agentic workloads inflate volume without a proportional rise in revenue. The comparison with Google is also rough: Google counts all models on its platform, while the startups count only open ones.

So it is worth looking at direction and order of magnitude, not exact shares. The direction, for now, is unambiguous: open inference is growing fast, and ignoring it when choosing a stack is already risky. Lambert's full conclusions are behind a subscription, while the free Interconnects dashboard with data on adoption and quality of open models can serve as a starting point.