The growing demand for AI has given rise to a new class of digital utility companies that sell computing power, access to models and infrastructure to developers. Leading the way is Fireworks AI, co-founded by former Meta CEO Lin Qiao, who led the creation of PyTorch, a popular open source 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 a lower cost than proprietary models, using open source models. It has access to many of the most capable open source models on the market, such as Meta’s Llama, Mistral, Qwen, and DeepSeek series. It also allows companies to upload their own data to train and fine-tune these models. Its clients include Cursor, Harvey, Uber, Shopify and others.
Lin describes Fireworks as a “specialized intelligence platform,” rather than general intelligence. Specialized intelligence was what AI researchers relied on primarily before general intelligence became applicable. “Before the advent of generative AI, there was no underlying model that brought the world’s knowledge together. But GenAI changed that,” Lin explained to Observer. “Now, it learns the underlying models from the public Internet and large datasets, creating a deeper, more general knowledge base that you can use directly as a black box API.”
But Lin believes that, given the abundance of public data and general intelligence models, the most valuable uses of AI, counterintuitively, will come from specialization.
“Because foundational models cannot access private data locked within applications and organizations,” she said. “The majority of data is private, locked within organizations as proprietary intellectual property and information that can never be shared outside the company.”
Training and tuning models using that proprietary data creates a constant need for Fireworks services. “This is an ongoing process because applications continue to evolve, data distribution changes, and underlying models continue to improve,” Lin said. “We have customers who tune in once a week, or once a day, or even once every few hours.” I expect this tuning process to be fully automated soon.
Once the model is fine-tuned, Fireworks helps optimize it in terms of inference speed and cost. The company offers some of the fastest inferences – the speed at which AI generates a response – in the industry. For example, Cursor’s code editor uses Fireworks’ speculative decoding to deliver code suggestions up to 13 times faster than traditional settings.
Fireworks processes more than 30 trillion tokens in daily inference traffic (excluding training), more than OpenAI and Google’s Gemini, according to the latest published data.
The company makes money by charging users a fixed price per million tokens. Codes are the basic unit of data that AI reads, processes and generates; In English, a token is approximately four letters long, or about three-quarters of a word.
“We provide a single platform that covers the entire spectrum of model development from start to finish, from quality to speed and cost,” Lin said. “The end result is that our customers get better quality, much faster speed and five to ten times lower cost, allowing them to go to full-scale production quickly.”
New trench
These days, AI executives like to talk about the “moat,” or the competitive advantage that allows a company to stay ahead of the competition. While turning an idea into an app is easier than ever thanks to AI coding tools, the traditional product moat is disappearing.
“Data is the trench, because it cannot be copied,” Lin declared. “The data collected to understand user intent, preferences, and interaction — what’s working well, what’s not working well and where you need to improve — is all your proprietary information, and that creates the asymmetry needed to compete. Anyone who can turn that data into their own intelligence can build on top of that. And that can multiply.”
Fireworks competes with closed model providers (such as OpenAI, Anthropic, and Google) and infrastructure platforms such as Together AI, Replicate, and AWS Bedrock. Its distinction lies in its focus on open models while integrating training, fine-tuning, and high-performance inference into a single system.
“We don’t need a Ferrari to buy groceries.”
Besides the data moat, another argument for open models is unit economics. By allowing developers to choose from a wide range of open-weight models, platforms like Fireworks can match each task to the most cost-effective level of intelligence. This flexibility is increasingly important as companies look to deploy AI on a large scale. Using a single parametric model for each task quickly becomes expensive.
“We don’t need to drive a Ferrari to go shopping,” Lin said. “There are a lot of tasks we solve every day with varying levels of complexity. Some are very difficult, requiring superhuman intelligence to solve. Others are not that difficult. If you use a vendor that can help you automatically choose the best-suited model to solve a given task, you will get the quality you need at the lowest cost.”
When Lin founded Fireworks two years ago, the company initially focused on heuristics, treating it as “one size fits all.” Now it’s doubling down on training as well, driven by rapid improvement and a cadence of open model releases. The quality of open models has significantly narrowed the gap with closed models, while release cycles have accelerated from monthly to weekly. New models often outperform benchmarks and approach borderline performance.
“This makes training particularly attractive,” Lin said. “With your own data and a little tuning, you can stay on top.”
She continued in conclusion: “We believe that specialized and generalized intelligence will coexist, but the world will not be dominated by a few generalized models. There will be millions of specialized models of intelligence – one for every use case.”
