Inference AI Infrastructure in the World of Test-Time Compute

Reasoning models shift compute from training to inference, and costs climb as apps make far more calls. YC wants a cheaper, better inference stack.

Until recently, YC notes, most compute spending went into pre-training foundation models. Reasoning models such as DeepSeek R1 and OpenAI's o1 and o3 point to a new scaling trend: far more compute needed at inference time, when apps actually use the models.

As AI apps make 10x or even 100x more API calls to complex reasoning models, YC expects infrastructure costs to become a real problem.

The ask: a rebuilt inference stack, with better inference-layer tooling, cheaper ways to run GPU workloads, and optimizations that let AI apps scale without losing money. YC calls it an unglamorous but critical problem, the kind that often creates big opportunities.

Read the full request on YC →