Anjney Midha, founder of AMP, discusses the real bottleneck in AI scaling: not GPU acquisition but GPU utilization. Frontier labs often run at sub-10% MFU while best-in-class is 60-70%. AMP is building a compute grid modeled after electric grid independent system operators (like PJM Interconnect), pooling supply and demand across clouds and silicon to make FLOPs flow like megawatts. The conversation covers data center community backlash, research hoarding at DeepMind, AMP's 1.2GW base-load ambition with 6GW spike capacity target, the concept of 'output maxing' as a discipline, non-NVIDIA chip standardization via NVIDIA reference architecture, why researchers make great CEOs, how Anthropic cracked coding, and Anjney's long-standing interest in AI-powered end-of-life prediction in healthcare.
