Fireworks, the Nvidia-backed inference cloud startup, raised $1.5 billion at a $17.5 billion valuation this week after crossing $1 billion in annualized revenue, five times where it stood a year ago. The round was led by Atreides Management, Index Ventures and TCV, with Nvidia, Evantic and Lightspeed Venture Partners also participating. It lands as finance chiefs push engineering teams toward cheaper, open-weight alternatives to frontier models. CEO Lin Qiao told CNBC the company is seeing “super-linear demand” and called the current market “a once-in-a-lifetime opportunity.”
The number that matters here is $17.5 billion for a company that hosts other people’s models rather than building its own. Anthropic and OpenAI have each been valued above $800 billion this year, so Fireworks sits two orders of magnitude below the labs whose output it serves. That gap is the business model. Fireworks runs the infrastructure that lets companies plug in DeepSeek, MiniMax, Z.ai and OpenAI’s own open-weight releases instead of paying frontier prices for every query.
Fireworks competes for that workload against Together AI and Baseten, with newer entrants like Fal chasing similar demand in media generation. What sets this round apart is the Nvidia stake alongside a striking usage claim: Qiao says her platform runs the equivalent-quality closed model at “five to 10 times” lower cost, a figure she gave CNBC without independent verification.
Scale is the other pitch. Qiao says Fireworks now moves 40 trillion tokens daily, a volume that dwarfs what the frontier labs have disclosed about their own developer traffic. Google put its processing at roughly 19 billion tokens per minute in a May disclosure, a pace equal to more than 27 trillion a day, while OpenAI’s March update pegged developer traffic near 15 billion tokens per minute, or about 22 trillion daily (each token runs roughly three-quarters of a word). If Qiao’s number holds, a five-year-old infrastructure reseller is now routing more daily token volume than either lab discloses processing for its own users.
The concentration risk behind that revenue deserves scrutiny. As of last year, the AI coding startup Cursor generated roughly half of Fireworks’ revenue, and Cursor is being acquired by Elon Musk’s SpaceX in a $60 billion stock deal expected to close this quarter. Qiao says the client base has since broadened to Elastic, GitLab and MongoDB, but a company whose growth story leaned on one customer, one now changing owners, carries a dependency the new valuation does not fully erase.
Fireworks is also courting the cloud giants rather than fighting all of them at once. It struck a partnership with Microsoft’s Foundry service in March, drawing on compute from more than 20 suppliers, and its “specialized intelligence” pitch echoes arguments this month from Microsoft CEO Satya Nadella and Palantir CEO Alex Karp, both of whom have said companies should be able to run a model without handing proprietary data to a frontier lab. Fireworks expects headcount to nearly triple, from around 200 to 600, by the end of 2026, and it hired former Salesforce executive George Hu as president in April to build a formal sales operation.
For any team currently budgeting frontier-model API spend, Fireworks’ token volume and Coinbase’s public move toward cheaper models are a signal to benchmark an open-weight alternative before the next contract renewal.
CNBC’s Jordan Novet reported the Fireworks funding round and revenue figures on July 16, 2026.
