Fireworks AI disclosed that the custom models it runs for customers process more than 40 trillion tokens per day—the basic unit models use to process text and code—more than doubling since this summer. We judge that the open-model inference market is expanding rapidly, but the available data still falls short of establishing how large a business it has become.
What this is
Nathan Lambert has long tracked open-model adoption and performance and built the freely accessible Interconnects Dashboard, which aggregates data on open-model adoption and performance. He now tracks the open-model inference market primarily through the token usage that Baseten, Fireworks, Together AI, and OpenRouter report for open models.
Public data points to a trajectory close to exponential growth: Fireworks says it now processes more than 40 trillion tokens per day and estimates that the autumn total could be well above 100 trillion. Yet cloud platforms such as Google, Amazon, and Microsoft do not disclose usage by model category, so this still does not represent a complete market size.
Industry view
We believe the value of open models is shifting from “cheap access to an off-the-shelf model” toward enterprise custom deployments and agentic workflows—AI systems that can independently plan and call tools to complete tasks. Growth in custom-model usage shows that companies are placing greater importance on their data, cost control, and adaptation to specific tasks.
The risks are equally clear: platforms use inconsistent reporting methods, and token volume does not equal revenue or profit. The coverage of major cloud providers remains unknown, while startups’ growth may partly reflect subsidies and early-adopter demand. More importantly, the question posed in the title—what market share open models actually hold against OpenAI and Anthropic—still has no reliable answer.
Impact on regular people
For enterprise IT: Open models provide more deployment options and greater negotiating leverage, but companies cannot look only at unit prices. They must also assess stability, data governance, maintenance costs, and the difficulty of switching models.
For individual professionals: Custom models may gradually enter spreadsheet processing, coding assistance, and workflow automation. Career competitiveness will increasingly depend on whether professionals can connect AI to real business operations, rather than on whether they can write prompts.
For the consumer market: As inference costs fall and custom capabilities improve, AI features in products will become more widespread and affordable. But the long-term coexistence of closed and open models will continue to shape both the final user experience and pricing models.