Inference Chips for Agent Workflows
YC partner asks founders to take on NVIDIA, Cerebras, Groq, Tenstorrent, and Etched on a $500K seed check. Chip industry responds with receipts.
“Inference Chips for Agent Workflows. Most AI chips are designed for "prompt in, response out." Agents don't work that way. They loop, branch, and call tools. We want to fund founders building inference silicon purpose-built for agentic workflows.”View on X ↗
Why this is wrong
- 01. Chip economics aren't seed-stage.
A competitive inference chip is a $200M–$500M raise problem, not a $500K one. Etched raised $120M for a single transformer ASIC. Cerebras, Groq, and Tenstorrent all sit in the multi-hundred-million tier before first silicon. YC writes $500K and a few partner office hours. Telling founders to commit five years to silicon under that capital structure is structurally incoherent.
- 02. "Prompt in, response out" is a strawman.
Modern inference systems already handle KV-cache reuse, branching, speculative decode, paged attention, and tool-call orchestration across chips. The bottleneck isn't a fixed prompt-response shape baked into silicon. It's KV-memory bandwidth, interconnect topology, and scheduler design — system-level problems, not chip-level ones. Chip founder @PMV_InferX put it bluntly: "inference is not a CHIP problem. It's a system problem."
- 03. YC has no partners with chip credentials.
No VLSI, no fab, no EDA, no analog/mixed-signal, no packaging. The author of this RFS bullet, Diana Hu, also authored "Supply Chain 2.0 for Semiconductors" the same day — meaning the same partner is asking founders to take on TSMC AND NVIDIA simultaneously, while YC has no demonstrated expertise in either. As @gauravisnotme put it: "strongly recommend against taking advice from people who know nothing about it."
- 04. Same RFS announcement covered "Software for Agents" and "AI OS for Companies."
Both are within YC's actual expertise envelope (consumer/SMB SaaS distribution, founder-led GTM). Asking the same batch to also produce inference silicon is a tell that the chip bullet was a marketing-driven thesis, not a returns-driven one.
The reception
“If you are really doing a chip startup, I would strongly recommend against taking advice from people who know nothing about it. Especially if you're trying to compete in a market where the competition has billions of dollars and decades of experience.”
“Just FYI, inference is not a CHIP problem. It's a system problem.”
“Here's $500k, go take on NVIDIA bud. But seriously, imagine trying to raise money + optimize custom hardware for a fad.”
“This was my first impression (basically she described CPU). And i'm doing a chip startup.”
“Competitive inference chip is also $200m to $500m fund raising problem. Better to solve this at system level.”
“too based, yc's been on a nosedive lately”