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

What It Covers

Anjney Midha, founder of AMP PBC and early Anthropic backer, explains how software-based compute orchestration can reduce effective GPU costs from $25–28 per hour to the marketed rate of $2.50, by standardizing fragmented chip infrastructure into a unified grid modeled on electricity distribution.

Key Questions Answered

• •Compute Utilization Gap: Most independent data centers run below 70% node utilization, and model flop utilization (actual chip usage during workloads) can fall below 11%. Elon Musk's Colossus 2 cluster in Memphis ran at under 60% node utilization. Researchers should measure output efficiency, not chip headcount, when evaluating AI infrastructure investments.
• •True Cost of Leased Compute: Long-term GPU leases appear priced at $2.50–3.00 per hour, but because research demand is spiky and teams over-provision for peak loads, the effective cost balloons to $25–28 per hour. AMP's grid reallocates idle capacity to other users, returning the actual price paid closer to the marketed rate.
• •Verifiable Feedback Drives Model Progress: AI models improve fastest where task outcomes can be objectively verified — software passing unit tests and pull request reviews, or materials science predictions confirmed by X-ray diffraction. Subjective feedback like "that answer was wrong" produces minimal improvement; structured verification loops are what separate fast-progressing domains from stagnant ones.
• •Multiple Frontiers, Not One Winner: The AI landscape contains at least 17 distinct frontiers — software engineering, consumer chat, video generation, scientific discovery — each with different leaders. Anthropic leads coding with under 5,000 employees while Google's 60,000-person team remains close but behind. Corporate AI buyers will increasingly route queries to whichever model is cheapest for a given task, abstracting away brand entirely.
• •Model-Harness Co-Design: Breakthroughs like Claude Code result from simultaneous development of model capabilities and the surrounding tooling harness, not harness innovation alone. Teams build the harness to anticipate specific model improvements three months out, then remove third-party tool dependencies once the model internalizes those capabilities — collapsing task completion time by one to two minutes per operation.