Fireworks Raises $1.5 Billion at a $17.5 Billion Valuation as Inference Demand Explodes

The pitch used to be simple: rent your intelligence from a big lab. On Thursday, a San Mateo startup founded by the engineers who built PyTorch raised $1.505 billion to argue that the entire industry has it backwards — and investors handed over a $17.5 billion valuation to find out.

Fireworks, which trains and serves customized open models for companies including Doximity, Revolut, Shopify, and Uber, announced its Series D on July 16. Atreides Management, Index Ventures, and TCV led the round, with participation from existing backers Evantic, Lightspeed Venture Partners, and NVIDIA. The full cap table runs deep: 20VC, Bessemer Venture Partners, Insight Partners, Lone Pine Capital, Menlo Ventures, Operator Collective, Ontario Teachers' Pension Plan, Original Capital, Prysm Capital, Quantum Capital, and TIME Ventures all wrote checks.

The timing is almost too neat. The round landed in the same 48-hour window that Moonshot AI unveiled Kimi K3, a 2.8-trillion-parameter open-weight model that sent shares of Chinese AI rivals sharply lower on Friday, according to Quartz. Open-weight models are precisely what Fireworks exists to serve.

From $4 billion to $17.5 billion in nine months

The numbers behind the valuation are the story. Fireworks has surpassed $1 billion in annualized revenue run rate, a 5x increase year-over-year from its last funding round. Daily token volume on the platform nearly tripled over the same period, from 15 trillion to more than 40 trillion tokens a day.

For context on the velocity: in October 2025, Fireworks raised a $250 million Series C at a $4 billion valuation, bringing its total funding at the time to $327 million. Nine months later, the company is worth more than four times that, and the new round alone is six times the size of everything it had raised before.

The most revealing figure isn't the valuation — it's the 95%. That's the share of tokens flowing through Fireworks that come from models specialized on a customer's proprietary data rather than pulled off the shelf. It suggests that enterprises reaching production scale aren't just price-shopping for cheaper inference. They're building something they intend to keep.

"Two paths forward"

Lin Qiao, Fireworks' co-founder and CEO, who previously led PyTorch at Meta, framed the round as a bet on which way the industry breaks.

> "There are two paths forward for AI. In one, intelligence belongs to a few big labs, and everyone else rents it. In the other, every company in the world builds specialized intelligence of its own, shaped by the domain only it understands. We are building towards the second," Qiao said. "Every company holds knowledge no one else has: its data, its workflows, its customers, its definition of quality. Fireworks turns that knowledge into specialized intelligence they own and can keep improving."

Gavin Baker, chief investment officer and managing partner at Atreides Management, made the investor case in less binary terms — notably, he does not think open models win outright.

> "Fireworks has assembled one of the most elite and technical teams in AI, paired with technology that consistently sets the pace for the industry and commercial momentum that very few companies have ever achieved at this scale," Baker said. "We believe both frontier and open models will increasingly be used together. With a differentiated platform used and trusted by enterprises to train, customize, and serve open models in production, Fireworks is exceptionally well-positioned to unlock the value of specialized intelligence built on proprietary data and workflows."

The platform now carries more than 200 models across text, image, and multimodal formats, with support for major open releases typically live within hours of launch. Proceeds go toward engineering headcount, global compute capacity, and deeper partnerships with Microsoft and NVIDIA. Qiao told CNBC she plans to grow the team from 200 to 600 by year-end, per The Next Web: "This is the year when we'll really hit the gas."

Analysis: the Kimi K3 thread

Here is why the Moonshot news and the Fireworks news are the same story.

Fireworks' entire thesis depends on open models being good enough that specializing one beats renting a frontier one. For years that was a stretch. It is no longer. Stanford's 2026 AI Index puts the best closed model just 3.3% ahead of the best open model. Kimi K3 took the top spot in the Frontend Code Arena with 1,679 points, ahead of Claude Fable 5 and GPT-5.6. Every time an open-weight release closes the gap, the value of a platform that turns those weights into production systems goes up — and the pricing power of the labs charging premium per-token rates goes down.

That's the arbitrage Fireworks is monetizing, and the token figure proves the mechanism works at scale. The Next Web, citing CNBC's numbers, notes that Fireworks moves more daily traffic than Google or OpenAI report serving developers — on a fraction of the revenue. Fireworks says its option runs at a fifth to a tenth of frontier cost. Volume without margin capture is exactly what you'd expect from a business selling the cheap substitute.

The risk is concentration. Cursor once accounted for roughly half of Fireworks' revenue, though Qiao says the base is broader now. The company also competes with Together AI and Baseten on inference, and increasingly with neoclouds on training — a market where the product is fast, cheap, and structurally hard to differentiate.

What to watch

Moonshot releases K3's weights on July 27. That's the test: Fireworks promises support for major open releases within hours, and K3 is the largest open-weight model ever built. How fast it goes live — and how much traffic it pulls — is the cleanest read yet on whether "specialized intelligence" is a durable category or a well-funded bet on the labs staying expensive.