Fireworks AI hits $17.5 billion valuation as companies hunt for cheaper alternatives to OpenAI and Anthropic

Key points

• Fireworks AI, a cloud startup that hosts and runs AI models for developers, surpassed $1 billion in annualised revenue in 2025, five times its revenue from the year before.
• The company raised $1.5 billion in new funding in mid-2025 at a valuation of $17.5 billion.
• Fireworks processes 40 trillion AI tokens, the tiny chunks of text that AI models read and write, every day, outpacing Google's disclosed developer traffic of roughly 27 trillion per day.
• The company's CEO says its models cost five to ten times less than comparable closed models from the major AI labs.
• Fireworks currently employs around 200 people and expects to reach 600 staff by the end of 2026.

Company finance chiefs are getting nervous about AI bills. That anxiety is proving to be very good news for a startup most people have never heard of.

Fireworks AI, headquartered in San Mateo, California, hosts AI models and runs them on behalf of developers building apps and products. Rather than build their own models from scratch, developers pay Fireworks to handle the heavy computing work. This week, first reported by CNBC Tech, the company announced it has crossed $1 billion in annualised revenue and closed a $1.5 billion funding round that values it at $17.5 billion. One year ago, revenue was a fifth of where it sits today.

Why would a company choose Fireworks over OpenAI or Google?

Cost is the short answer. CEO Lin Qiao told CNBC that running a model through Fireworks costs five to ten times less than running an equivalent closed model from one of the major labs. Closed models are AI systems where the internal workings are kept private by the company that built them, such as OpenAI's GPT-4o.

Fireworks specialises in hosting open-weight models, AI systems where the underlying design is published publicly for anyone to inspect and adapt. Names like DeepSeek and Meta's own releases fall into this category. Businesses can bring their own private data and fine-tune, meaning carefully adjust, these models for specific tasks, without handing that data to a big lab. That's the protection enterprises are now paying for.

"A company should be able to use a model without giving up the knowledge that makes it unique," Microsoft CEO Satya Nadella wrote recently, a sentiment that captures the exact anxiety Fireworks is profiting from. Microsoft is now a Fireworks partner, letting its own customers tap Fireworks models through a shared arrangement.

Fireworks now handles 40 trillion tokens per day. One token equals roughly three-quarters of a word. Google disclosed in May 2025 that its developer AI tools handled about 27 trillion tokens per day; OpenAI's developer tools processed roughly 22 trillion per day as of March 2025. As our 15 July story on the AI cost shift found, companies are quietly moving away from the most powerful models toward cheaper, targeted ones. Fireworks is where a lot of that traffic is landing.

Until recently, a single customer, the AI coding tool Cursor, accounted for about half of Fireworks' revenue. Cursor has since been acquired by SpaceX in a $60 billion deal. Qiao says the business is now far more spread out, with clients including Elastic, GitLab, MongoDB and database infrastructure companies beyond that group.

Fireworks was founded in 2022 by Qiao, a former Meta director, alongside six co-founders. The fresh funding will go toward buying more GPUs, the specialised chips that power AI model inference, and building out a proper sales team for the first time.

Watch the customer concentration number. Cursor was half the business not long ago. If one or two of the new names follow the same path, that diversification story gets complicated fast.