Keeping an AI model fast enough to use is turning out to be the expensive part of building one.
Fireworks AI just raised $1.5 billion at a $17.5 billion valuation. The company doesn't build AI models. It takes other companies' models and makes them run faster and cheaper. That round is a clear signal about where the money in AI is actually moving right now: not just toward smarter models, but toward making the models that already exist usable at scale.
Most people size up an AI company by how smart its model seems. That's the wrong first question.
Fireworks is now doing more than $1 billion a year in revenue, up five times from last year. It moves more than 40 trillion tokens (roughly, pieces of text) through its systems every day, nearly triple what it was moving a year earlier. Its investors, including Index Ventures, TCV, and Nvidia, didn't back the company because the underlying models got smarter. They backed it because someone had to solve the unglamorous problem of making AI usable at scale, once it already works.
In my experience, that's the question most people skip. Everyone asks whether a model is good. Almost nobody asks, up front, whether it can be served fast enough and cheap enough to keep a paying user around.
