Fireworks reaches $17.5 billion valuation in open-model AI push

Fireworks has raised $1.5 billion in a new funding round that values the Nvidia-backed AI cloud startup at $17.5 billion, the company said Thursday. For investors following the AI buildout, the deal points to rising demand for cheaper model infrastructure beyond the biggest names in cloud and AI labs.

The San Mateo, California-based company runs open-source and open-weight artificial intelligence models for software developers. In plain English, Fireworks helps companies plug AI models into their own apps without relying only on closed systems from companies such as OpenAI or Anthropic.

Fireworks said it has topped $1 billion in annualized revenue, which means its current revenue pace projected over a full year. CNBC reported that figure is five times higher than last year.

CEO Lin Qiao told CNBC that Fireworks is seeing “super-linear demand” and called the market a “once-in-a-lifetime opportunity.”

Why companies are looking at open models

CNBC reported that finance executives have grown more concerned about the cost of leading AI models and have started asking employees to consider open-source alternatives. Open-source models are AI systems whose code or model details are made available for others to use or adapt, depending on the license.

Fireworks operates in what is known as the inference cloud market. Inference is the step where an AI model responds to a user request after it has already been trained. Fireworks manages the computing infrastructure that lets developers send those requests to models at scale.

Qiao told CNBC that Fireworks’ cost is five to 10 times cheaper than an equivalent-quality closed model. She said the company gives customers a way to use their own data to tune models for specific tasks, describing that as “specialized intelligence” compared with the “generalized intelligence” offered by Anthropic and OpenAI.

The company supports models from Chinese companies including DeepSeek, MiniMax and Z.ai, according to CNBC. It also offers access to open-weight models released by OpenAI last year. Open-weight models make the trained model parameters available, while the underlying training data and full development process may remain closed.

Microsoft partnership expands reach

Fireworks competes with Amazon and Google in model hosting, but it has also started working with large platforms. In March, the company announced a partnership with Microsoft that lets Microsoft customers access models through Fireworks.

Qiao told CNBC the arrangement gives Fireworks a wider distribution channel. Fireworks uses computing power from more than 20 suppliers, including Microsoft, according to CNBC.

The company is also moving beyond inference. CNBC reported that Fireworks has begun offering graphics processing units, or GPUs, for AI model training. GPUs are the chips used for the heavy math behind modern AI systems.

Funding, customers and scale

Atreides Management, Index Ventures and TCV led the new round. Nvidia also joined the financing, along with Evantic and Lightspeed Venture Partners, according to CNBC.

Fireworks was founded in 2022 by Qiao, a former Meta director, and six co-founders. The company has about 200 employees, and Qiao told CNBC she expects that number to reach 600 by the end of 2026.

The startup hired former Salesforce executive George Hu as president in April. CNBC reported that Fireworks plans to use the new capital to build out sales, secure more GPUs and hire more technical staff.

Qiao said Fireworks now processes 40 trillion AI tokens per day. A token is a unit of text used by AI models, equal to about three-quarters of a word, according to CNBC.

Fireworks once got about half of its revenue from AI coding startup Cursor, CNBC reported. Qiao said the company is now more diversified. Other customers include Elastic, GitLab and MongoDB, according to CNBC.

This story draws on original reporting from CNBC.