Fireworks AI is betting on specialized inference models over general-purpose large language models, carving out a differentiated AI inference enterprise play.
According to Lin Qiao, the future of A.I. lies in millions of specialized models built on proprietary data. The soaring demand for A.I. has given rise to a new category of digital utility companies that sell compute power, access to models, and developer infrastructure. Among the leaders in this space is Fireworks AI, co-founded by former Meta executive Lin Qiao, who led the creation of PyTorch, a popular open-source machine learning framework, alongside a team of engineers from Meta and Google. This week, the startup closed a $1.5 billion Series D round at a $17.5 billion valuation, with annualized revenue surpassing $1 billion.
Fireworks AI is a platform that enables developers to build products faster and at lower cost than with proprietary models by leveraging open-source alternatives. It provides access to many of the most capable open-source models on the market, including Meta’s Llama series, Mistral, Qwen, and DeepSeek. The platform also allows enterprises to upload their own data to train and fine-tune these models. Its client roster includes Cursor, Harvey, Uber, and Shopify, among others.
Lin describes Fireworks as “a specialized intelligence platform,” as opposed to one focused on general intelligence. She notes that specialized intelligence was what A.I. researchers primarily relied on before general intelligence became viable. “Before generative A.I. was a thing, there was no foundation model holding world knowledge together. GenAI changed that,” Lin explained to Observer. “Now, foundation models learn from the public internet and massive labeled datasets, creating a deeper, more generalized knowledge base that you could use directly as a black-box API.”
Yet Lin believes that amid the abundance of public data and general-purpose models, the most valuable A.I. applications will counterintuitively stem from specialization. “Because foundational models do not have access to the private data locked inside applications and enterprises,” she said. “The majority of data is private, locked inside enterprises as proprietary IP and information that would never get shared outside the company.”
Training and fine-tuning models with that private data creates a sustained demand for Fireworks’ services. “This is a continuous process because applications keep evolving, data distribution changes, and base models keep improving,” Lin said. “We have customers tuning once a week, once a day, or even once every few hours.” She predicted that this tuning workflow would soon be fully automated.
Once a model is finely tuned, Fireworks helps optimize it for inference speed and cost. The company delivers some of the fastest inference—the speed at which an A.I. generates a response—in the industry. For example, Cursor’s code editor uses Fireworks’s speculative decoding to deliver code suggestions up to 13 times faster than traditional setups. Fireworks processes more than 40 trillion tokens in daily inference traffic, excluding training, outpacing both OpenAI and Google’s Gemini according to the latest published data.
The company generates revenue by charging users a flat rate per million tokens. Tokens are the basic unit of data that an A.I. reads, processes, and generates; in English, a token represents roughly four characters, or about three-quarters of a word. “We provide one platform covering the whole end-to-end spectrum of model development, from quality to speed and cost,” Lin said. “The end result is our customer gets better quality, much faster speed, and five to ten times lower cost, allowing them to go to production at a massive scale quickly.”
These days, A.I. executives frequently discuss “moats”—competitive edges that allow companies to maintain their lead. In an era when A.I. coding tools make it easier than ever to turn ideas into applications, traditional product moats are disappearing. “Data is the moat, because it cannot be copied,” Lin declared. “The data collected to understand user intent, user preferences and user engagement—what works well, what doesn’t work well and where you should optimize—is all your proprietary information, and that creates the asymmetry needed to compete. Whoever can turn this data into their proprietary intelligence can build on top of that. And that can compound.”
Fireworks competes with both closed-model providers, such as OpenAI, Anthropic, and Google, and infrastructure platforms like Together AI, Replicate, and AWS Bedrock. Its differentiation lies in focusing on open models while tightly integrating training, fine-tuning, and high-performance inference into a single system.
Beyond the data moat, another argument for open models centers on unit economics. By allowing developers to choose from a wide range of open-weight models, platforms like Fireworks can match each task with the most cost-efficient level of intelligence. This flexibility is increasingly critical as companies look to deploy A.I. at scale, since using a single frontier model for every task quickly becomes prohibitively expensive. “We don’t need to drive a Ferrari to go grocery shopping,” Lin said. “There are so many tasks we solve day-to-day at varying levels of complexity. Some are extremely hard, requiring beyond-human-level intelligence to solve. Others are not that hard. If you use a vendor who can help you automatically select the best model suitable for solving a particular task, you get the quality you need at the lowest cost.”
When Lin founded Fireworks two years ago, the company initially focused on inference, treating it as a “one size fits one” solution. Now, it is doubling down on training as well, driven by the rapid improvement and accelerated release cadence of open models. Open model quality has significantly narrowed the gap with closed models, while release cycles have sped up from monthly to weekly. New models frequently top benchmarks and approach frontier-level performance. “This makes training particularly appealing. With your private data and a little bit of tuning, you can stay on top,” Lin said. Concluding her outlook, she added, “We believe specialized and generalized intelligence will coexist, but the world will not be dominated by a few generalized models. There will be millions of specialized intelligence models—one per use case.”