Fireworks raises $1.5B as specialised AI infrastructure reaches scale
Fireworks AI has raised a $1.505B Series D at a $17.5B valuation as enterprises spend more on the infrastructure needed to customise and run AI models in production.
The AI infrastructure race is moving beyond training the largest models. Fireworks AI has raised a $1.505B Series D at a $17.5B valuation, reflecting investor demand for the systems that make models fast, affordable and reliable enough to run inside real products.
What happened
Fireworks was founded in 2022 by engineers with backgrounds at Meta’s PyTorch organisation and Google’s Vertex AI team. Chief executive Lin Qiao previously led PyTorch at Meta. The company provides infrastructure for deploying, fine-tuning, evaluating and serving AI models, with a particular emphasis on high-speed inference for open and specialised models.
The round was led by Atreides Management, Index Ventures and TCV, with participation from existing and new investors including Nvidia, Lightspeed Venture Partners, Bessemer Venture Partners, Insight Partners, Menlo Ventures and Ontario Teachers’ Pension Plan. It follows a $250M Series C at a $4B valuation in October 2025.
Fireworks says it has passed a $1B annualised revenue run rate and now processes more than 40T tokens per day. Disclosed customers include Uber, Shopify and Doximity. These operating figures are company-reported rather than independently audited.
Why it matters
The repeated cost of generating AI outputs can eventually exceed the cost of training the original model. Companies therefore need infrastructure that can route workloads, reduce latency, control spending and choose the most suitable model for each task. Fireworks is betting that enterprises will use a mix of frontier, open and internally customised models rather than depend on one closed provider.
Its acquisition of compute-orchestration startup Hathora also gives Fireworks more control over how workloads are distributed across global GPU capacity.
The bigger picture
AI infrastructure is becoming a strategic layer between model developers and application companies. Fireworks could benefit even if leadership at the model layer keeps changing, because customers still need deployment, optimisation and monitoring.
The risk is that this remains a capital-intensive business exposed to GPU availability, declining inference prices and competition from Amazon, Microsoft, Google and other specialist providers. Fireworks must prove that its software and performance advantages are strong enough to protect margins as compute becomes cheaper and more widely available.