Startups

Fireworks AI hits $17.5B valuation on $1.5B Series D as enterprises pivot to open models

The Nvidia-backed inference platform crossed $1 billion in annualized revenue, up 5× year-over-year, as CFOs push teams toward cheaper, customizable alternatives to frontier lab APIs.

By Giulia Donatello · Startups & funding · 2026-07-25 · 3 min read

Photo: Unsplash / Taylor Vick — Rows of illuminated server racks in a data center

Fireworks AI closed a $1.505 billion Series D on July 16 at a $17.5 billion valuation, roughly seven weeks after Bloomberg reported the company was in talks at $15 billion. A $2.5 billion mark-up between the leak and the close is the kind of pricing motion that tells you the term sheet was oversubscribed before it hit the wire.

Atreides Management, Index Ventures, and TCV led the round, with Nvidia, Lightspeed Venture Partners, and Evantic returning. Gavin Baker of Atreides is quoted in the announcement framing this as a bet on specialized intelligence rather than frontier scale.

The numbers underneath the price are the actual story. Fireworks says it’s crossed $1 billion in annualized revenue, up 5× year-over-year, and that daily token volume has climbed from 15 trillion at its prior round to more than 40 trillion. For context, CNBC notes Google disclosed in May that its models were processing about 19 billion tokens per minute for developers (implying north of 27 trillion daily), and OpenAI’s developer tools run roughly 22 trillion per day. An independent inference platform is now moving traffic on the same order as the frontier labs it undercuts.

That’s what CEO Lin Qiao, formerly a director at Meta, calls “super-linear demand” and a “once-in-a-lifetime opportunity.” The customer list, Uber, Shopify, Doximity, Geico, Revolut, reads as procurement rather than experimentation. CNBC’s read is that rising frontier-model prices have finance chiefs directing engineering teams toward open-source alternatives, and Fireworks has positioned itself as the neutral pipe: it draws on more than 20 compute suppliers, and 95% of the tokens it serves come from fine-tuned specialized models rather than generic APIs.

The concentration risk is real but shrinking. Cursor once accounted for over 50% of revenue; the enterprise book now dilutes that. April’s hire of former Salesforce executive George Hu as president, and a March partnership routing Microsoft’s Fireworks Foundry customers through Fireworks’ compute, both point at the same motion: enterprise distribution. Qiao expects headcount to triple from around 200 to 600 by year-end.

The 2000s cloud analogy writes itself, but the more precise parallel is Snowflake’s 2018–2020 run, when CFO-led cost pressure on incumbent data warehouses created room for a neutral, multi-cloud alternative to compound past the point anyone had priced in.

Sources