Jensen Huang Called One AI Company the ‘TSMC of AI Factories.’ Its CTO Sees What Comes Next

Months after Nvidia (NASDAQ:NVDA) CEO Jensen Huang dubbed Fireworks AI the “TSMC of AI factories,” the startup’s co-founder and CTO Dmytro Dzhulgakov foresees a market shift away from single frontier models toward millions of specialized, open-weight systems.

Fireworks AI builds the infrastructure and operating layer to help companies deploy, customize, and run AI models at scale efficiently.

From One Model to Millions

Huang’s TSMC comparison highlights Fireworks’ core mission: operating and optimizing AI models for speed and cost efficiency rather than building giant proprietary ones.

But Dmytro foresees a larger shift as open-weight models rapidly catch up to frontier labs, which face a growing hurdle: lacking proprietary customer data.

According to Dmytro, AI-native startups like Cursor, Cognition, Perplexity, Genspark, and Harvey build products around general-purpose models, then use internal data to fine-tune specialized open-weight models.

That shift can elevate target performance from 80–90% to 95–99% while cutting costs tenfold—making user-generated data far more valuable than the underlying code.

Fireworks Sees An Opportunity

That fragmentation does not make AI infrastructure simpler. It makes it more complicated.

Dmytro described inference as a stack spanning GPU kernels, hardware orchestration, model sharding, traffic routing and load balancing. Even the same model may require radically different configurations depending on whether the customer prioritizes speed or cost.

Fireworks operates across multiple cloud providers, effectively creating a "virtual cloud" that abstracts heterogeneous compute from customers. Dmytro said Fireworks has deliberately stayed out of hardware and data-center management, focusing instead on compute orchestration and the software platform.

Why it Matters

Huang’s TSMC comparison may ultimately prove useful for a reason beyond inference.

If AI moves from a handful of giant models toward millions of specialized systems, the winners may include the companies quietly operating, optimizing and connecting those models underneath the applications investors actually see.

Fireworks’ bet is that AI factories will need their own operating layer — just as chip factories do.

Image via Shutterstock