Startups

Fireworks AI hits $17.5B valuation on $1.5B Series D as enterprises chase cheaper, specialized models

The NVIDIA-backed inference startup founded by the PyTorch team crossed $1 billion in annualized revenue and now serves 40 trillion tokens a day, betting the future of enterprise AI is owned, not rented.

By Giulia Donatello · Startups & funding · 2026-07-24 · 3 min read

Photo: Unsplash / Taylor Vick — Rows of networked servers in a data center, illustrating AI inference infrastructure

Fireworks AI closed a $1.505 billion Series D at a $17.5 billion post-money valuation, more than quadrupling the $4 billion mark it carried last October when it raised $250 million. Atreides Management, Index Ventures, and TCV led the round, with Nvidia, Lightspeed Venture Partners, Bessemer Venture Partners, Menlo Ventures, Evantic Capital, and 20VC participating. The size and speed of the markup say less about Fireworks itself than about where enterprise AI budgets are actually going.

The company, founded by the PyTorch team out of Meta and now led by CEO Lin Qiao and president George Hu, formerly of Salesforce, says it has crossed $1 billion in annualized revenue, five times its year-ago run rate. Token volume nearly tripled over the same window, from 15 trillion a day to more than 40 trillion. For scale, that’s above the roughly 27 trillion tokens Google processes through its developer API and the roughly 22 trillion OpenAI moves.

More than 95% of those tokens flow through specialized or proprietary models rather than general-purpose frontier ones. Fireworks offers over 200 models across text, image, and multimodal formats, and pitches customers like Doximity and Geico on inference that’s five to 10 times cheaper than an equivalent-quality closed model.

That’s the structural story. Gartner expects worldwide end-user spending on AI models and platforms to reach $64 billion in 2026, up 63.4% from $39 billion in 2025, with domain-specific and specialized models growing 210%. The 2017 cloud-repatriation debates rhyme here: once workloads mature, the economics of renting a general-purpose product from a hyperscaler stop penciling out against something narrower and owned.

Fireworks is positioning itself as the substrate for that shift. It runs on more than 20 compute suppliers, headcount sits around 200 today and is expected to reach 600 by year-end, and Nvidia is simultaneously an investor and a vendor. The unusual part isn’t the valuation. It’s that the biggest inference customers appear to have already decided the frontier-model API is a transitional product, and are pricing the transition accordingly.

Sources