Fireworks raises $1.5 billion at $17.5 billion valuation
The AI infrastructure company says annualized revenue has passed $1 billion and will use the Series D to add compute capacity and engineers.
By Colin Brandt · Enterprise Reporter
· 3 min read
Fireworks AI has raised $1.5 billion in a Series D financing that values the AI infrastructure startup at $17.5 billion. Atreides Management, Index Ventures and TCV led the round, with Nvidia and more than a half-dozen other investors also participating, according to the company.
The round puts another large private-market price tag on AI infrastructure, where investors are backing companies that can package GPU access, model tuning and inference into developer products. Fireworks said its annualized revenue recently moved above $1 billion, though it did not disclose profitability, headcount or the full list of participating investors.
Fireworks sells a cloud platform for developers building with open-source AI models. Its customers include Samsung Electronics, GitLab and other large technology companies, according to Fireworks. The company says its platform now handles more than 40 trillion tokens a day for users.
What Fireworks sells
The company’s core product gives developers access to managed graphics processing unit clusters under usage-based pricing. Customers use the platform to fine-tune open-source models, then run those models in production through Fireworks’ inference services.
Fireworks also offers an AI agent meant to automate parts of the training process. Developers can specify the task they want a model to perform and provide a training dataset, after which the agent selects technical settings for the training run, according to the company.
Fireworks says the agent searches for hyperparameter combinations that improve model output quality. The company also says the system can supplement a customer’s training data with DPO files, which contain examples or instructions for how a model should respond to prompts.
For larger training jobs, Fireworks supports four parallelization methods, each tuned for a different model type, according to the company. Customers can run those methods side by side or use a single approach.
Inference is part of the pitch
After a model is customized, customers can host it through one of two Fireworks inference products. One is a serverless service, which removes the need for customers to configure the underlying infrastructure. The other, called Deployments, gives users dedicated GPU clusters.
Fireworks says Deployments delivers better performance than its serverless service and provides more controls. Those include autoscaling settings, which determine how capacity is added or removed as demand changes, and quantization, a compression technique used to reduce infrastructure requirements for models.
“Every company holds knowledge no one else has: its data, its workflows, its customers, its definition of quality,” co-founder and Chief Executive Officer Lin Qiao said in the company’s announcement. “Fireworks turns that knowledge into specialized intelligence they own and can keep improving.”
The company said it will use the new capital to expand its infrastructure and hire more engineers. It did not give a timeline for the buildout or specify how much of the funding will go toward compute capacity versus hiring.
This story draws on original reporting from SiliconANGLE.
