Former Meta PyTorch Leader Raises $1.5B to Bet Against the Big AI Labs She Helped Power
For seven years, Lin Qiao led the engineering team at Meta that built and shipped PyTorch, the open‑source framework that now quietly underpins AI development at OpenAI, Google, and Meta itself. In 2022, she left one of the most influential engineering roles in the entire AI industry to build a company premised on a very specific idea, that the intelligence being built on top of frameworks like PyTorch should not end up concentrated in the hands of a small number of large labs.
That company, Fireworks AI, has now raised 1.505 billion dollars in a Series D round at a 17.5 billion dollar valuation. The round was led by Atreides Management, Index Ventures, and TCV, with Nvidia, Lightspeed Venture Partners, and Evantic also participating, alongside a broader investor list that includes 20VC, Bessemer Venture Partners, Insight Partners, Lone Pine Capital, Menlo Ventures, and Ontario Teachers' Pension Plan.
The raise arrives as Fireworks crosses 1 billion dollars in annualized revenue, a fivefold increase year over year, while daily token volume processed on its platform has climbed from 15 trillion to more than 40 trillion. Qiao has framed the company's mission as a direct alternative to a future where a handful of frontier labs control access to advanced AI, describing a choice between one path where intelligence belongs to a few big labs and everyone else simply rents it, and another where every company builds specialized intelligence shaped around the domain only it understands.
In practice, Fireworks lets enterprises fine‑tune and run open‑source AI models on their own proprietary data rather than depending entirely on a general‑purpose system from a frontier lab. A bank, for example, could fine‑tune an open model on years of internal compliance documentation so it understands regulatory language a generic assistant would not, then run that customized model on Fireworks' infrastructure with guaranteed latency and full control over the data. Of all tokens processed on the platform today, roughly 95 percent come from models customers have customized on their own data rather than off‑the‑shelf frontier models, a figure that underscores just how central customization is to the company's actual business rather than serving as a side feature. Its enterprise customer base includes Uber, Shopify, Revolut, and Doximity.
Qiao co‑founded the company alongside five other former Meta engineers, including Dmytro Dzhulgakov and Dmytro Ivchenko, both of whom had served as PyTorch core maintainers at Meta before relocating to the United States to join the founding team. The company's name traces directly back to that shared origin story, a nod to PyTorch's torch carrying forward into fire.
The round comes with a notable backstory around customer concentration that the company has had to work through. As of last year, roughly half of Fireworks' revenue came from a single customer, AI coding startup Cursor. That concentration became a live risk in June, when SpaceX announced a 60 billion dollar all‑stock deal to acquire Cursor, a transaction still awaiting regulatory approval expected in the third quarter of 2026. As part of that shift, Cursor began scaling its Composer model on SpaceXAI's own Colossus compute infrastructure, reducing its reliance on third‑party inference providers including Fireworks. Qiao has said the company is now considerably more diversified, pointing to its broader enterprise roster as evidence, though the scale of Fireworks' fivefold revenue growth looks different depending on how much of it was originally driven by Cursor specifically.
Fireworks is competing in one of the most heavily capitalized corners of AI infrastructure. Together AI raised 800 million dollars at an 8.3 billion dollar valuation just two weeks before Fireworks closed its own round. Baseten raised 300 million dollars at a 5 billion dollar valuation from IVP and CapitalG, with Nvidia contributing 150 million dollars of that round directly. Groq raised 650 million dollars in June after licensing its chip technology to Nvidia. Meanwhile, Microsoft Azure and Amazon Web Services are increasingly folding managed open‑model hosting directly into their own cloud platforms, effectively offering a bundled version of what Fireworks sells as a standalone product.
That dynamic points to the sharpest long‑term question hanging over the company. Nvidia is simultaneously one of Fireworks' investors and a company actively expanding into inference infrastructure through its own acquisitions, while Microsoft functions as both a partner and a platform‑level competitor. Fireworks is staking its position on the depth of its customization capabilities, arguing that most enterprises need models substantially reshaped around their own proprietary data rather than simply served faster off the shelf. Whether that differentiation holds once major cloud providers begin bundling similar tools into existing contracts at little additional cost is likely to be one of the defining questions for enterprise AI spending over the next year.
The global generative AI market was valued at roughly 22.2 billion dollars in 2025 and is projected to reach 324.7 billion dollars by 2033, according to research firm Grand View Research. Fireworks plans to use its new capital to fund global compute expansion, engineering hires, and deeper partnerships with Microsoft and Nvidia, with headcount expected to grow from around 200 employees to 600 by the end of 2026. This marks the company's fourth funding round, climbing from a 552 million dollar valuation at its Series B in July 2024, to 4 billion dollars at its Series C in October 2025, and now 17.5 billion dollars at Series D, a trajectory that either confirms Qiao's original bet against Big Tech‑controlled AI or sets up a far more complicated story once hyperscalers decide to compete on the same ground in earnest.
