AI Coding Agents for Hardware-Optimized Code

Nvidia's lead rests largely on hand-optimized CUDA code. YC wants AI that writes kernels and other low-level code so alternative chips can compete.

YC argues that AI hardware is still held back by software. In its view, Nvidia dominates largely because AI models use CUDA's hand-optimized code. Competing hardware, such as AMD or custom silicon, often underperforms not only because of the chips but because system-level code like kernels and drivers is very hard to write and too few engineers work on it.

Why now: YC thinks reasoning models such as DeepSeek R1 and OpenAI's o1 and o3 could generate hardware-optimized code that matches or beats human-written CUDA.

The ask: AI-generated kernels that make more hardware alternatives work for AI. YC frames this as breaking dependencies, not just improving performance.

Read the full request on YC →