Discovered Materials is hunting cooler chips with AI—because the heat wall is now as real as the compute wall.

A startup is using AI to explore the space of possible semiconductor and cooling materials. At 1,000W per chip, thermal management is becoming the limiting factor for datacenter density.

In plain terms: Discovered Materials applies AI to find new semiconductor and thermal materials for chips. The business case: as AI training chips push past 700-1,000W per unit, heat is now as big a constraint as compute density.

The problem they're solving

Modern AI accelerators (H100, B200-class) generate heat at a scale that limits how densely you can pack them in a rack. Liquid cooling helps but doesn't solve the fundamental problem—the materials themselves. Better thermal conductors, phase-change materials, and novel semiconductor substrates could let chips run hotter without throttling, or cooler at the same performance envelope.

Discovered Materials runs an AI-in-the-loop search: generate candidate material compositions, predict properties computationally, filter by manufacturability constraints, iterate. TechCrunch describes it as "AI whack-a-mole"—the search space is vast and the useful candidates are sparse.

Under the hood

The AI-for-materials-discovery space includes DeepMind's GNoME (2.2M crystal structures), Microsoft's MatterGen, and Meta's Open Catalyst Project. Discovered Materials focuses specifically on thermal applications for semiconductors—a narrower, commercially closer target than the broader crystal structure discovery work.

So what

The compute scaling problem is increasingly a physics problem. Materials AI is a credible long-horizon lever—but the gap between "discovered in simulation" and "manufacturable at yield" is wide. This is a 5-year infrastructure bet, not a next-quarter fix. Worth watching if you're tracking the hardware layer of the AI stack.