A startup Nvidia's Jensen Huang calls "the TSMC of AI factories" says the next wave of AI money will go to the companies running and tuning models, not the ones building the biggest ones — good news for infrastructure players, awkward for the top labs.

• Fireworks AI runs the plumbing that lets companies deploy, customize and serve AI models cheaply and fast, rather than building its own giant model.
• Its CEO, Dmytro Dzhulgakov, expects the market to move from a handful of frontier models to millions of specialized open-weight ones.
• The big labs face a growing problem: they don't own their customers' private data, which is what makes a model actually good at a specific job.
• AI-native firms like Cursor, Perplexity and Harvey already fine-tune open models on their own data, pushing accuracy from 80-90% to near-perfect while cutting costs tenfold.
• Fireworks deliberately stays out of chips and data centers, stitching together several cloud providers into one virtual layer instead.

Outlook: If AI really splinters into millions of tailored models, the quiet orchestration layer — an operating system for AI factories — becomes the place where value piles up.