The inference startup founded by ex-Meta PyTorch lead Lin Qiao has hit $1 billion in annualized revenue and now processes 40 trillion tokens per day, cementing a bet that enterprise AI runs on cheaper, specialized models, not rented frontier APIs.
Nine months is a long time in AI infrastructure. In October 2025, Fireworks AI closed a $250 million Series C that valued the company at $4 billion. This week, as reported by CNBC, it closed a $1.505 billion Series D at $17.5 billion, with Atreides Management, Index Ventures, and TCV leading the round. Nvidia, Lightspeed, Bessemer, Menlo Ventures, Insight Partners, and Ontario Teachers' Pension Plan also participated. That's a more than fourfold valuation jump in under a year - and it's not being driven by hype. Fireworks has crossed $1 billion in annualized revenue run rate: five times what the company reported at its last raise. The numbers are real.
Lin Qiao built the platform on a straightforward premise: general-purpose frontier models are expensive to run at scale, and most enterprise use cases don't need the full capability of a GPT-4 or a Claude. What they need is a fine-tuned, faster, cheaper version of a capable open-weight model that actually knows their data. Fireworks handles the infrastructure to get there. Notion is a clear example of what that looks like in practice. The company partnered with Fireworks to fine-tune its models and cut AI feature latency from two seconds to 350 milliseconds - a 5.7x improvement that users feel on every keystroke. That kind of gain isn't available from a frontier API call.
Who's actually using it #
The customer list tells you where the demand is coming from. Cursor, Uber, DoorDash, Samsung, Shopify, Perplexity, Vercel - over 10,000 companies in total - are running production workloads on Fireworks. These aren't companies doing AI experiments. They're companies with high-volume, latency-sensitive applications where cost per token actually shows up in the margin structure. At 40 trillion tokens processed daily, Fireworks sits at a scale where even fractional price efficiency compounds into something real.
The broader dynamic here is one the industry has been circling for a while. Frontier model providers - OpenAI and Anthropic foremost among them - have built a business renting access to the most capable models at prices that reflect the cost of training them. That's fine for low-volume tasks: a legal team drafting an occasional memo, or a startup prototyping a feature. It doesn't work when you're Uber routing millions of trips or DoorDash personalising feeds for tens of millions of orders. The maths gets uncomfortable fast. Fireworks is where companies go when they've done it.
What the valuation actually says #
What's notable about the $17.5 billion valuation isn't the size of it. It's the revenue multiple underneath. Sacra puts Fireworks' annualized revenue at more than $1 billion as of this month, which prices the Series D at roughly 17 times forward revenue. That's not a speculative bet on future markets. It's a market willing to pay a growth premium on a company that already has the customers, the volume, and the trajectory. Five times revenue growth year-over-year tends to earn that kind of treatment.
Qiao spent years at Meta leading PyTorch, which became the default framework for training and running neural networks across the research and industry worlds. That background matters more than it might look. Fireworks isn't a reseller of third-party compute or a thin wrapper on Hugging Face - it has built its own inference engine, optimised for throughput and cost at scale, fine-tuning models on customer data to hit latency and accuracy targets that commodity inference can't match. The headcount sits at around 200. Qiao has said she expects that to reach 600 by year's end. That suggests the Series D is going directly into engineering capacity, not marketing.
Frankly, the raise is as much a data point on where AI infrastructure economics are heading as it is on Fireworks specifically. The story of the past two years in enterprise AI has been companies discovering that buying frontier model access is the beginning of the infrastructure journey, not the end. Running models efficiently in production - at the latency and cost their applications actually require - turns out to be a hard, separate problem that frontier labs aren't incentivised to solve for you. Fireworks has built a business on exactly that gap.
There's a lot of gap left to fill. Total funding raised now exceeds $1.8 billion. The company declined to comment on specific deployment timelines beyond the headcount target. Fireworks' next question isn't whether the market is real. It's whether a $17.5 billion infrastructure business can stay ahead of the frontier labs themselves as they push harder into enterprise tooling.
Also read: China signs 29 nations into a rival AI bloc while open-source models do the real work • BloombergNEF just revised its US data center power forecast 83% higher in seven months • Claude Fable 5 helped crack the Jacobian Conjecture after 87 years of failure