Summary
Latent Space sits down with Anjney Midha, CEO and founder of Amp, at Periodic Labs to discuss what he calls the biggest bottleneck in the AI industry: infrastructure waste. Midha — whose co-founder Seb built the BorgX/BorgGQM scheduler at Google — argues that most AI clusters are dramatically underutilized, with best-in-class MFU (model FLOP utilization) sitting at 60–70% and node allocation targets needing to reach 95%+ (a figure Google treats as an outage threshold). He frames this not as a technical problem but a culture and alignment problem: too many degrees of separation between the capital funding a cluster and the engineers measuring its output.
The conversation develops a broader philosophy Midha calls “output maxing” or frontier systems engineering — the idea that the AI industry has a license to say “this time is different” on capabilities, but not on infrastructure discipline. He draws on lessons from the semiconductor and DSA industries to argue for iterative cluster bring-ups over big-bang deployments, citing Anthropic’s transformer architecture bet as an example of productive standardization over premature diversification.
Midha also discusses Amp’s public benefit corporation structure and two personal missions: AI-assisted end-of-life clinical decision-making to reduce Medicare waste, and net-positive data centers as a prerequisite for the compute scale needed to make those models possible. The episode is particularly valuable for engineers and infrastructure leaders thinking about GPU cluster economics and organizational alignment.
📺 Source: Latent Space · Published June 18, 2026
🏷️ Format: Interview
