Most of the world's useful data lives inside individual companies, not on the open internet. A model fine-tuned on that private data will beat a general purpose model at the specific job a company actually needs done, and Fireworks has built its entire platform around making that possible at scale.
Today, the company is announcing a $1.5 billion Series D at a $17.5 billion valuation, led by Atreides Management, Index Ventures, and TCV, alongside existing investors like Lightspeed and Nvidia. I'm delighted to share that Operator Collective has joined this round, having been first introduced to CEO and Co-Founder Lin Qiao via founding OpCo LP and Board of Advisors member Erica Ruliffson Schultz at one of our AI Luminaries dinners last year.
The Company
Fireworks trains and runs the models behind production AI applications. Instead of stitching together separate tools for training a model, serving it, and finding compute to run it on, the company put all three on one platform: a training system, an inference platform, and a virtual GPU cloud spanning 20 clouds and more than 30 regions.
Fireworks was founded in October 2022 by seven engineers who led PyTorch at Meta and AI serving at Google Cloud, including CEO Lin Qiao and CTO Dmytro Dzhulgakov. Cursor, Cognition, GitHub, Meta, Genspark, and Doximity all run production workloads on the platform today.
Why You Should Pay Attention
For the first few years of the AI boom, most of the attention went to training bigger models. Today, the bigger cost center for most companies is inference, the actual running of a model once it's built. As more companies build AI agents that work all day instead of answering one question at a time, that running cost keeps climbing.
At the same time, open models (ones anyone can download and fine tune, rather than a closed model you rent by the token) have gotten good enough to compete with, and sometimes beat, the big proprietary options on the tasks that matter to a specific business. A few years ago, betting a product on an open model felt risky. Today, it's a serious option, and Fireworks has built the infrastructure that makes it practical at scale. Today, 65% of tokens served through Fireworks are from models that have been specialized.
And, Fireworks went from roughly 15 trillion tokens served per day at its last funding round to more than 43 trillion today, and its annualized revenue run rate crossed $1 billion, up from about $280 million at that previous round.
The Details
Running a model well is more complicated than it sounds. There are thousands of ways to configure how a model runs, different chips, different levels of compression, different setups for speed versus cost, and the best combination shifts every week as new hardware and models come out. Rather than build one serving system and hope it works everywhere, Fireworks benchmarks its workloads across leading open-weight model-serving frameworks, picks whichever performs best for that specific job and hardware, and then improves it further with its own code. A few examples:
- Speculative decoding: a small, fast model drafts an answer and a bigger model checks it, which speeds things up without hurting quality.
- KV cache reuse: the system saves work it's already done so similar questions don't start from zero.
- Quantization: running the model in a lighter numerical format to save memory and increase speed.
None of these are exclusive to Fireworks. But tuning all of them together, reliably, at scale, is the hard part, and that's exactly what Fireworks gets right.
On the training side, Fireworks offers everything from fully managed fine-tuning to full control over custom training loops, and customers like Cursor and Cognition use it for both training their models and serving them in production. Cursor's Composer 2 model, trained on Fireworks, scored higher on a widely used coding benchmark than Anthropic's Opus 4.6. Once a company trains its model on Fireworks and serves it there too, switching providers is a quick decision.
Why We're Invested
Lin Qiao ran engineering for the PyTorch and Caffe2 frameworks at Meta, tools that a huge share of the AI industry is now built on. Dmytro Dzhulgakov was a core maintainer of PyTorch itself, and co-founder Chenyu Zhao led Google's Vertex AI serving infrastructure.
The company also recently brought on George Hu as President, who previously helped take Salesforce's revenue from zero to $5 billion and served as COO at Twilio through its highest growth years. Bringing on someone with that exact track record shows Fireworks is serious about moving beyond developer signups into large enterprise deals. A new partnership that embeds Fireworks directly within Microsoft's Azure AI Foundry gives it a distribution channel most infrastructure startups never have access to.
This is also the kind of company our Collective Venture Model™ is built for. Plenty of our 250+ Operator LPs have built or bought infrastructure like this themselves, and when they look at Fireworks' numbers, they recognize a problem they've lived through.
What's Next
Fireworks is putting this new funding into growing its engineering team, expanding global compute capacity, and going deeper with partners like Microsoft and Nvidia. As open models keep improving and more companies decide they'd rather own their AI than rent it, Fireworks is positioned to be the infrastructure that decision runs through.
If your company is weighing whether to keep renting intelligence or start building a version that's actually yours, this is a good time to take a look. Learn more at fireworks.ai.
