Anjney Midha's Plan to Radically Lower the Price of Compute

Large AI models are fundamentally constrained by physical inputs — power, GPUs and datacenter infrastructure — which drives firms to repurpose infrastructure, consider extreme mitigations (e.g., off‑Earth data centers) and rationalizes outsized spending to be 'first' at AGI; practical competition often requires strategic compute/capital partnerships (example: a $4B Amazon compute/capital partnership).

Severe underutilization inflates effective GPU costs (roughly $25–$28/hr after wasted capacity versus marketed rates near $2.50), creating a large efficiency opportunity: standardizing heterogeneous compute (AMP's 'Borg for everybody' and grid credits), better operations (raising utilization from low‑60s to high‑90s), and workload packing (filling idle pockets with inference) can materially close that gap.

Orchestration and model‑agnostic layers that prioritize cost, efficiency, security and trust will be rewarded by corporate buyers, even as cloud providers (example: Microsoft) build inference chips to control unit economics for serving; AMP’s public benefit corporate structure also allows it to balance for‑profit margins with mission choices.

Frontier capabilities are converging today (examples cited: '4.8' and 'GPT 5.5'), with distillation and information‑sharing propagating similarity and making durable proprietary dominance harder — yet contrarian bets and small, focused teams can still create frontier labs (guest cited early backing of Anthropic), although compute access remains the gating constraint.