Fireworks, the Nvidia-backed inference cloud startup, has raised $1.5 billion at a $17.5 billion valuation after hitting $1 billion in annualized revenue, five times what it booked a year ago. The round, announced Thursday, makes plain what finance chiefs have been whispering: the sticker price of frontier models is forcing a rethink.
The revenue milestone is the hard evidence. Fireworks, founded in 2022 by former Meta director Lin Qiao and six colleagues, now runs at a $1 billion run-rate against a $17.5 billion price tag, a 17.5x revenue multiple that sits comfortably below the $800 billion-plus valuations assigned to Anthropic and OpenAI, but well above the trivial. The company employs roughly 200 people and expects 600 by end of 2026. Qiao calls it "super-linear demand." The market calls it a once-in-a-lifetime opportunity. Both can be true.
What Fireworks sells is not intelligence but access. It hosts open-weight models, DeepSeek, MiniMax, Z.ai, and OpenAI's own open releases, on compute drawn from more than 20 suppliers, Microsoft included. A March partnership with Microsoft lets Azure customers pull those models through Fireworks. "Through Microsoft we can get much bigger reach," Qiao said. The arrangement is notable because Microsoft fields its own Foundry service for the same job. The truce suggests the hyperscalers know they cannot hoard every workload.
The pitch is arithmetic. Qiao says Fireworks runs five to ten times cheaper than an equivalent closed model. Coinbase chief Brian Armstrong said in June the exchange is adopting cheaper models where it makes sense. Satya Nadella blogged Sunday that a company should use a model without surrendering the knowledge that makes it unique. Palantir's Alex Karp put it blunter: technical customers want to own the means of production. The rhetoric aligns with the invoice.
Fireworks is not alone. Baseten and Together AI hunt the same inference dollar. CoreWeave, Lambda and Nebius, the neoclouds, rent GPUs for training and now trade at a $42 billion valuation after CoreWeave's $1.5 billion IPO last year. DigitalOcean shares have risen 149 percent this year as growth accelerated. The inference layer is crowded, and the hyperscalers still own the metal. Fireworks' edge is the orchestration layer and the sales motion it is only now building: former Salesforce executive George Hu joined as president in April to turn self-serve sign-ups into enterprise contracts.
The new capital buys GPUs and headcount. Qiao says this is the year they hit the gas. The test is whether open-model performance keeps closing the gap on the frontier labs fast enough to keep the cost advantage decisive. If the best closed models pull away again, the five-to-ten-times-cheaper argument frays. If they don't, Fireworks' $17.5 billion valuation starts to look like a down payment on the next cloud layer.