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

The AI industry is facing a hidden crisis: massive waste in compute usage, with data centers operating at less than 70% utilization and researchers paying up to $28 per GPU hour despite advertised rates of $2.50. Anjney Midha, founder of AMP PBC, argues that this isn't a hardware bottleneck but a systemic inefficiency rooted in outdated leasing models and fragmented infrastructure. His solution? A software-powered 'grid' that standardizes compute across chips, clouds, and manufacturers, turning idle capacity into a fungible resource. By abstracting hardware complexity and enabling dynamic allocation, AMP aims to slash costs and democratize access—making it possible for any lab to reach the frontier without needing to build its own data center. Midha also challenges the myth of a single 'winner-takes-all' AI model, instead advocating for a 'jagged frontier' where specialized models thrive in distinct domains. The real value, he insists, lies not in raw scale but in verifiable feedback loops, technical literacy, and co-design between models and tools—transforming AI from a speculative race into a predictable, productive utility. The episode reveals that the most expensive AI failures aren't technical—they're economic. Companies like Uber are blowing through token budgets because of poor model routing and lack of internal expertise. Midha’s vision is not just to lower prices, but to rewire how we think about AI infrastructure: as a public utility, not a private fortress.