Robotics investor's bet: 2027 is an infra year; latency is the biggest gap for LLMs on real robots

paigeinsf · x · 2026-11-20

An investor whose team runs weekly evals on frontier models with real robots offers a bet on the next 1-3 years of robotics, noting there is no consensus and most takes rest on demos and sim scores.

Core thesis: 2027 will be an infrastructure year, driven by LLMs overhauling the model landscape, with open weights catching up after. General LLMs are already good at memory and long-horizon tasks; latency is the biggest remaining gap plus system 0 optimization. Once solved, data, evals, and post-training become where the value lands almost overnight.

The quoted addendum cites Mercor ($1M → $50M run-rate in 2024, $500M by late 2025 as labs bought expert data for RL) and Applied Compute (founded May 2025 after DeepSeek R1, betting on GPUs + RL post-training on enterprise data) to argue timing relative to model progress matters most for data/infra companies.

Original post →

More from AGI Musings

AGI Musings channel →