Inference Chips for Agent Workflows

Agents loop through tool calls, branches and long contexts, and GPUs sit 30–40% utilized on that work. YC wants silicon and compilers designed for the agent loop.

Most AI chips assume "prompt in, response out", but agents loop: calling tools, branching, backtracking and holding context across many steps. YC says GPUs reach only 30–40% of peak utilization on this bursty work, which keeps switching between model calls limited by memory, tool calls limited by I/O, and orchestration limited by the CPU. No one is building hardware around the agent loop: quick switching between models, built-in speculative decoding, and memory for KV caches that last a whole run.

The insight: as with Groq, the compiler matters as much as the chip. Founders who understand both chip architecture and how agents execute are well placed.

Read the full request on YC →