Fireworks AI raises $1.5B to build the specialized AI inference and training platform for enterprise

Fireworks AI raises $1.5B to build specialized inference and training platform

Fireworks AI has raised $1.5 billion to expand its AI inference and training platform, which CEO and co-founder Lin Qiao describes as a "specialized intelligence platform" designed to let enterprises own and operationalize their proprietary data rather than rely on generic API wrappers.

The core product is a co-designed training and inference stack. Qiao argues the platform can deliver five to ten times lower operating costs compared to standard approaches, which matters because cost, not capability, is increasingly what determines whether an AI-native business survives at scale.

“We raised 1,500,000,000. We focus on building specialized intelligence platform — we want to make sure every single company has a tool to protect their alpha and to turn their alpha into their own intelligence. In AI time, once you have product market fit, you're likely to scale into bankruptcy. We want to give our customer the best tool to build a specialized intelligence, to have full control of their own intelligence, to stand on top of and have full control of the cost for them to scale in the long run.”

Scaling into bankruptcy

Qiao's sharpest argument is that the SaaS-era growth playbook breaks down for AI companies. In SaaS, product-market fit is the green light to scale as fast as possible. In AI, she argues, scaling before solving the cost structure can push companies toward insolvency even when the product is working. She says this isn't just a startup problem — large public companies with winning AI products are getting stuck because deploying those features to their full user base would be prohibitively expensive.

The customer base

Named customers include Cursor and Harvey, both training customized models on Fireworks' platform. Qiao says the customer mix extends across coding, legal, finance, recruiting, marketing, and customer support — all verticals where companies have proprietary data that, in her framing, should be powering their own models rather than enriching a generic foundation model provider.

Vertical integration and the stack

Qiao declines to signal that Fireworks will build down into custom silicon or power infrastructure. The philosophy is to partner with specialists at each layer of the stack and focus narrowly on training and inference. Qiao frames the platform's edge as switching from "token maximization" to "value maximization" — using the most economical approach to solve a specific task rather than maximizing compute throughput.

Capital deployment

The round is being used primarily to expand capacity across compute, research, and commercial headcount. Qiao is explicit that demand is growing super-linearly and the bottleneck is supply, both human and infrastructure.