Fireworks AI is in talks to raise a new funding round at a $15 billion valuation, Bloomberg reported Tuesday. Index Ventures, which co-led the company's $250 million Series C last October, is set to co-lead the new round. Terms have not closed and the lead investor on the priced portion has not been disclosed. Update: the round has since closed — see Uristocrat's coverage of Fireworks' $1.505 billion Series D at a $17.5 billion valuation.
The valuation curve is the story
The mark would put Fireworks at roughly 27x its $552 million Series B valuation from July 2024 and 3.75x its $4 billion Series C valuation from October 2025. The company has raised more than $327 million across rounds led by Lightspeed Venture Partners, Index Ventures, Evantic Capital, and Sequoia Capital. The seven-month gap between the $4B and $15B marks is the kind of cadence that defines this generation of AI infrastructure deals: capital is not pricing tomorrow's revenue, it is pricing the runway to own a layer of the stack.
What Fireworks actually does
Fireworks runs an inference platform: enterprises bring open-source or fine-tuned models, Fireworks serves them with sub-second latency at a per-token cost the company says runs up to 8x cheaper than alternatives. The pitch is "product-model co-design" — letting customers customize models on proprietary data without standing up their own GPU clusters. Per the company's Series C disclosure, the platform processes more than 10 trillion tokens per day, serves 10,000+ companies including Samsung, Uber, DoorDash, Notion, Shopify, and Upwork, and crossed $280 million in annualized revenue. The newer Bloomberg-cited figures put daily volume at 15 trillion tokens — a 50% growth in roughly six months.
Lin Qiao's bet
CEO and co-founder Lin Qiao previously led the PyTorch team at Meta — the open-source machine learning framework that effectively powers most of the AI model training and inference happening outside Google. Her thesis is that the inference layer, not the model itself, is where enterprise margin will sit: model quality is converging across vendors, but serving cost and reliability are not. The $15 billion mark suggests the market is paying for that thesis, not for any individual model.
How this lands in the broader compute race
The Fireworks round sits inside a fast-narrowing field of inference-and-deployment companies — Sierra closed $950 million at a $15 billion valuation earlier this month, Together AI and Groq are mid-raise, and the hyperscalers are spinning up purpose-built inference silicon of their own. Even Microsoft is in talks to sell Anthropic its Maia 200 inference chips, which would have been a foundational-customer relationship a year ago. The Fireworks bid is for the layer above silicon: it does not matter whose chip is underneath if a single API can route to the cheapest, fastest one. Whether that abstraction holds margin once Nvidia, AMD, and the hyperscalers start bundling inference directly is the bear case the next round has to outrun.
Comments