The growing demand for AI has created a new category of digital services companies that sell computing power, access to models and engineering infrastructure. Among the leaders of this pack is Fireworks AI, co-founded by former Meta CEO Lin Qiao, who led the development of PyTorch, a popular open machine learning framework, and a team of engineers from Meta and Google.
Fireworks AI is a platform for developers to build products faster and at lower cost than proprietary models, using open source models. It has access to many open source models on the market, such as Meta’s Llama series, Mistral, Qwen and DeepSeek. It also allows businesses to upload their own data to train and fine-tune these models. Its clients include Cursor, Harvey, Uber and Shopify, among others.
Lin describes Fireworks as a “special field of intelligence,” contrary to conventional wisdom. Special intelligence is what AI researchers relied heavily on before general intelligence worked. “Before generative AI became a thing, there was no foundational model that held the world’s knowledge together. GenAI changed that,” Lin told the Observer. “Now, the underlying models learn from the public Internet and large labeled data sets, creating a deep, general knowledge base that you can use directly as a black box API.”
But Lin believes that, amid the abundance of public data and general intelligence models, the most important use of AI, on the contrary, will come from experts.
“Because basic models can’t access the private data locked inside applications and companies,” he said. “A lot of information is private, locked inside businesses like proprietary IP and information that can never be shared outside of the company.”
Training and fine-tuning models with that private data creates an ongoing demand for Fireworks services. “This is an ongoing process because applications continue to evolve, data distribution changes, and underlying models continue to evolve,” Lin said. “We have customers who tune in once a week, once a day, or once every few hours.” He predicted that this tuning process would work automatically.
Once the model is fine-tuned, Fireworks helps optimize it for computing speed and cost. The company offers the fastest feedback—the speed at which AI generates feedback—in the industry. For example, the Cursor code editor uses Fireworks’ predictive modeling to deliver code suggestions that are up to 13 times faster than a traditional setup.
Explosives process more than 30 trillion tokens in daily inference traffic (without training), more than OpenAI and Gemini of Google, according to the latest published data.
The company makes money by charging users a low price per million tokens. Tokens are the basic unit of data that AI reads, processes and generates; in English, a token is about four letters, or about three-quarters of a word.
“We offer a single platform that covers the entire spectrum of end-to-end model development, from quality to speed and cost. The end result is that our customer gets better quality, faster speed, and five to ten times lower costs, which allows them to go to large-scale production quickly,” said Lin.
The new moat
These days, AI executives like to talk about the “moat,” or the competitive edge that allows a company to stay ahead of the competition. At a time when it’s easier than ever to turn an idea into an app thanks to AI coding tools, the traditional product funnel is disappearing.
“Data is a resource, because it cannot be copied,” Lin said. “The data collected to understand user intent, user preferences and user engagement — what’s working well, what’s not working well and where to optimize it — is all your proprietary information, and that creates the asymmetry needed to compete. Anyone who can turn this data into their own proprietary intelligence can build on that. And that can add up.”
Fireworks competes with both closed model providers (such as OpenAI, Anthropic and Google) and infrastructure platforms such as Together AI, Replicate and AWS Bedrock. Its difference lies in its focus on open models while tightly integrating training, optimization and high performance guidance into a single system.
“We don’t need a Ferrari when we buy groceries.”
Besides the data moat, another argument for open models is the economic unit. By allowing engineers to choose from a wide range of open-weight models, platforms like Explosives can match each job with the most cost-effective level of intelligence. This flexibility is becoming increasingly important as companies look to deploy AI at scale. Using a single, parameterized model for every job quickly becomes too expensive.
“There’s no need to drive a Ferrari to go grocery shopping,” said Lin. “There are many tasks that we solve every day with varying levels of complexity. Some are very difficult, requiring ingenuity beyond the human level to solve. Others are not so difficult. When you use a vendor that can help you automatically choose the best model to solve a particular task, you get the quality you need at a low cost.”
When Lin founded Fireworks two years ago, the company initially focused on understanding, taking it as “one size fits all.” Now, it is doubling down on training again, driven by the rapid development and release cadence of open models. The quality of open models has significantly reduced the gap with closed models, while release cycles have been accelerated from monthly to weekly. Newer models tend to be higher and closer to borderline performance.
“This makes training especially attractive. With your private data and a little tuning, you can stay on top,” Lin said.
He went on to conclude, “We believe that special and general intelligence will exist, but the world will not be dominated by a few general models. There will be millions of special models of intelligence—one of each.”
