The key points are as follows. Fireworks is a startup that provides cloud deployment services for open-source large language models to software developers. It has just completed a $1.5 billion funding round, achieving a post-money valuation of $175 billion.
The company, which has received investment from NVIDIA, has seen its annualized revenue surpass $1 billion, marking a fourfold increase year-over-year.
The company's CEO, Qiao Lin, revealed that the code development tool startup Cursor previously contributed over half of its revenue. Now, as more enterprises pivot towards open-source models, the company's client base has become significantly more diversified.
Currently, the high costs associated with mainstream closed-source large language models are putting pressure on corporate financial officers, who are increasingly instructing their technical teams to switch to open-source alternatives.
The demand-driven market tailwinds are propelling the growth trajectory of cloud service provider Fireworks. The company competes with Amazon and Google, primarily offering hosting services for large models to developers, enabling them to integrate these models into their own applications.
This NVIDIA-backed startup announced on Thursday that its annualized revenue has exceeded $1 billion, which is five times the figure from the same period last year. The company has just concluded a $1.5 billion funding round, valuing it at $175 billion post-money.
"The market demand is growing at a super-linear rate," said Qiao Lin, co-founder and CEO of Fireworks, in an interview at the company's headquarters in San Mateo, California. "Catching such an industry wave is truly a once-in-a-lifetime opportunity."
While Anthropic and OpenAI have valuations exceeding $800 billion this year, making Fireworks significantly smaller by comparison, and dwarfed by tech giants with trillion-dollar market caps, this startup's revenue milestone strongly indicates that enterprise clients are not entirely satisfied with the closed-source models from leading AI labs.
This achievement also demonstrates that the cloud computing market is not entirely dominated by Amazon, Microsoft, and Google.
DigtalOcean, which focuses on lightweight cloud infrastructure, has seen its stock price surge 149% year-to-date, with its business growth accelerating. CoreWeave, a company specializing in renting out NVIDIA Graphics Processing Units (GPUs), completed a $1.5 billion IPO last year and is currently valued at $42 billion.
Fireworks' main business is inference cloud services, providing compute infrastructure for hosting large models, competing in this space with startups like Baseten and Together AI. It has also expanded into GPU rental services, similar to CoreWeave, Lambda, and Nebius, offering compute power for large model training to its clients.
Fireworks is no longer operating in isolation and has begun forming industry alliances.
In March of this year, the company announced a partnership with Microsoft. Microsoft's own Model Foundry platform also offers open-source model hosting services. Through this partnership, enterprise customers within the Microsoft Windows and Office ecosystems can directly access various models hosted by Fireworks. With over 20 compute power suppliers, Microsoft is a key partner for Fireworks.
"Leveraging Microsoft's channels allows us to significantly broaden our customer reach," Qiao Lin stated.
Fireworks provides developers with access to open-source models from several Chinese companies, including DeepSeek, MiniMax, and Z.ai, while also hosting the open-weight models released by OpenAI last year. Qiao Lin explained that enterprise clients can import their proprietary industry data to fine-tune and optimize these open-source models, ultimately producing custom models that are highly tailored to specific vertical scenarios and deliver top-tier performance.
While Anthropic and OpenAI focus on general-purpose large language models, Fireworks specializes in intelligent models for specific vertical domains, creating a differentiated market positioning.
This perspective aligns with recent commentary from Silicon Valley tech leaders.
Microsoft CEO Satya Nadella stated in a blog post on Sunday, "Enterprises should have the right to use AI models while retaining their unique core business data and intellectual property."
Nadella's remarks echoed the views expressed by Palantir CEO Alex Karp earlier this month.
Karp noted that enterprise technology teams "want to maintain control over their data assets and are reluctant to entrust core assets to third-party platforms for hosting."
The cost differential is also a crucial factor driving companies towards open-source solutions. Coinbase CEO Brian Armstrong posted on social platform X in June, stating that the company has fully switched to low-cost open-source models for suitable business scenarios.
"For equivalent performance, the inference cost of our open-source models is 5 to 10 times lower than that of closed-source commercial models," Qiao Lin explained.
A former Meta executive, Qiao Lin co-founded Fireworks with six other founders in 2022. The company currently has around 200 employees, and she expects the team to grow to 600 people by the end of 2026.
"2026 will be the year we go all out, with our business reaching full-scale momentum," Qiao Lin said.
In April of this year, Fireworks hired former Salesforce executive George Hu as its President. For years, the company's customer acquisition relied primarily on self-service sign-ups. It now plans to build a professional sales team. Proceeds from this funding round will be used to expand GPU compute capacity and hire more technical research and development personnel.
An increasing number of developers are choosing Fireworks to handle their AI inference needs.
Qiao Lin disclosed that Fireworks processes 40 trillion tokens per day. In May, Google disclosed that its large models for developers process approximately 19 billion tokens per minute, equating to over 27 trillion tokens per day. In March, OpenAI announced that its developer tools process 15 billion tokens per minute, or about 22 trillion tokens daily. The industry-standard conversion is roughly 1 token equals 0.75 English words.
Last year, the code AI startup Cursor contributed nearly half of Fireworks' revenue. Cursor has now significantly reduced its reliance on OpenAI and Anthropic by developing its own proprietary code model, Composer.
"Our client structure is now highly diversified," Qiao Lin stated. In June, Elon Musk's space exploration company SpaceX announced a $60 billion all-stock acquisition of Cursor, with the transaction expected to close this quarter. Other key clients of Fireworks include Elastic, GitLab, and MongoDB.
This funding round was led by Atreides Management, Index Ventures, and TCV, with participation from NVIDIA, Evantic, and Lightspeed Venture Partners.
Comments