Fireworks Raises $1.5 Billion at a $17.5 Billion Valuation

Fireworks closes a $1.5 billion round at a $17.5 billion valuation

Fireworks, an AI inference platform based in San Mateo, raised a $1.5 billion Series D at a $17.5 billion valuation, the company confirmed on July 16. The round was led by Atreides Management, Index Ventures, and TCV, with participation from existing backers Evantic, Lightspeed Venture Partners, and NVIDIA.

The valuation is built on unusually hard numbers for a company at this stage. Fireworks says it has crossed $1 billion in annualized revenue run rate, a 5x increase from its last raise, and now serves more than 40 trillion tokens a day, up from 15 trillion over the same period. It plans to spend the money expanding its engineering team and global compute capacity and deepening partnerships with Microsoft and NVIDIA.

Why inference, not model training, is where the revenue is

The frontier of AI attention is model training, the multi-billion-dollar runs that produce systems like Claude and GPT. The frontier of AI revenue is increasingly somewhere quieter, in inference, the business of actually running models in production, fast and cheap, at enterprise scale. Fireworks sits in that second layer. It does not train the largest models. It runs them, and it runs the tuned versions enterprises build on top.

That position is what investors paid $17.5 billion for. A model is a fixed asset that depreciates the moment a better one ships. An inference platform with a billion dollars of run-rate revenue and 40 trillion tokens flowing through it every day is a recurring, sticky, hard-to-replace piece of infrastructure. Once a company's production traffic runs through Fireworks, moving it is expensive and risky, which is precisely the quality that commands a premium.

The 95% number that explains the raise

The single most telling figure is not the valuation. It is that 95% of the tokens Fireworks serves come from specialized models rather than general-purpose frontier systems. Enterprises are not, in the main, sending their production traffic to one giant do-everything model. They are running smaller, tuned, task-specific models, and they need somewhere to run them at scale.

That is a data point about where enterprise AI is actually going, and it explains why the round closed at this size. It is also the same shift showing up elsewhere in the same week. Thinking Machines released an open-weight model built to be customized, and Kimi K3, an open-weight system, topped a human-preference coding board. Fireworks is the infrastructure bet on that world. If the future of enterprise AI is many specialized models rather than one general one, the company that runs them profitably captures value no matter which lab wins the model race.

What a $17.5 billion inference valuation signals for the AI stack

For founders and investors, the signal is that the value in AI is spreading down the stack, away from the model and toward the layers that deploy and serve it. NVIDIA's participation is the tell. The company that sells the compute is also buying equity in the platform that resells that compute as inference, a hedge that pays off in almost any version of the future.

The AI market has spent two years pricing model labs as the prize. Fireworks just raised $1.5 billion on the argument that the durable business is running the models, not building them, and that most of what runs in production will be specialized rather than general. At a billion dollars of revenue and 40 trillion tokens a day, that is no longer a thesis. It is a run rate.