Fireworks hits $17.5 billion valuation and $1B in annualized revenue

The cost of the latest AI models is creating increasing concern among finance executives, who are beginning to direct employees to consider open source alternatives.

This speeds up cloud startup fireworkcompeting with Amazon And Google to host models that developers can incorporate into applications. NvidiaThe -backed company said Thursday that its annual revenue surpassed $1 billion, five times last year, and is now up $1.5 billion at a $17.5 billion valuation.

“We’re seeing super linear demand,” Fireworks co-founder and CEO Lin Qiao told CNBC in an interview at the company’s headquarters in San Mateo, California. “This is a once-in-a-lifetime opportunity to have a market like this.”

Fireworks are much smaller anthropic And OpenAIThis year, investors are not close to the top names in the tech world, each worth over $800 billion and whose market capitalization is in the trillions. But the startup’s revenue milestone shows that companies are not entirely satisfied with the models coming out of the best labs.

This success also allows Amazon to Microsoft and Google hasn’t completely dominated cloud computing.

Shares of the easy-to-use cloud infrastructure provider DigitalOcean Growth is up 149% so far this year accelerated. CoreWeaveThe company, which rents out Nvidia graphics processing units, or GPUs, raised $1.5 billion in its IPO last year and is now valued at $42 billion.

By managing the compute infrastructure for models, Fireworks has business in the inference cloud market alongside startups like Baseten and Together AI. Additionally, neoclouds CoreWeave has also begun providing GPUs for training Lambda and AI models. nebius.

Instead of going it alone, Fireworks began forming alliances.

He made a statement in March partnership With Microsoft offering its own Foundry service for running open models. The arrangement allows Windows and Office company customers to leverage models through Fireworks, which rely on computing power from more than 20 vendors, including Microsoft.

“Through Microsoft, we can achieve a much broader reach,” Qiao said.

Fireworks offers developers an easy way to adopt models from Chinese companies like DeepSeek. MiniMax And Zai. OpenAI, which was released last year, also has open-heavy models. The goal is for customers to bring in their own data that is not available in frontier labs and refine the models until they deliver state-of-the-art performance for specific tasks, Qiao said.

While Anthropic and OpenAI offer “generalized intelligence,” Fireworks can unlock “specific intelligence,” he said.

This claim may sound familiar to those who follow the rhetoric of technology leaders.

Microsoft CEO Satya Nadella wrote on Sunday: blog post “A company should be able to use a model without giving up the knowledge that makes it unique.”

What Nadella is talking about palantir CEO Alex Karp’s remarks on CNBC earlier this month.

Karp said technical customers “want to know they have the means to produce.” “It is not transferred to anyone else.”

Dollars and cents are also a factor. Cryptocurrency exchange operator coinbase CEO Brian Armstrong wrote in a June article that he was embracing cheaper models where it made sense. x post.

“Compared with the closed model of equivalent quality, our cost is five to 10 times cheaper,” Qiao said.

an old one Meta director Qiao and his six co-founders founded Fireworks in 2022. Approximately 200 people work in the company. Qiao expects the number of people to reach 600 by the end of 2026.

“This is the year we really hit the gas,” Qiao said.

Fireworks ex-hired sales force Manager George Hu was appointed chairman in April. The startup plans to build a formidable sales team after years of customers signing them up. The new money will also help Fireworks buy more GPUs and hire more tech experts.

Developers are increasingly relying on Fireworks to process requests.

Qiao said Fireworks now processes 40 trillion AI tokens per day. Google In May, it announced that artificial intelligence models processed approximately 19 billion coins For developers, this works out to more than 27 trillion per minute per day. OpenAI announced In March, we stated that developer tools were working on 15 billion tokens per minute, which means approximately 22 trillion tokens per day. Each symbol is approximately three-quarters of a single word.

As of last year, about half of Fireworks’ revenue came from Cursor, an AI coding startup that has become less reliant on OpenAI and Anthropic and created a proprietary model called . Composer.

“We are much more diverse now,” Qiao said. Elon Musk in June SpaceX It agreed to acquire Cursor in a $60 billion all-stock deal, and the transaction will close this quarter. Other Fireworks clients include: Elastic, GitLab And MongoDB.

Atreides Management, Index Ventures and TCV led Fireworks’ new round. In addition to Nvidia, Evantic and Lightspeed Venture Partners also participated.

WRISTWATCH: OpenAI CEO: China’s open source models are getting better