Why AI Labs With Unlimited GPUs Still Fail — Anjney Midha, AMP

TL;DR

• 1

  95% node utilization should be table stakes: Midha says single-tenant GPU clusters should run above 95% node utilization and best-in-class MFU should be around 60 to 70%, otherwise the issue is usually leadership and alignment, not hardware scarcity.

• 2

  Amp wants to be the PJM of compute, not another neo-cloud: He describes Amp as an independent system operator for AI, pooling about 1.3 gigawatts of trusted supply and demand across clouds so labs get guaranteed base load plus flexible spike capacity.

• 3

  Anthropic's edge was preparation, not luck: Midha says the company spent four years building a culture around efficiency, safety, and a P0 on coding, which is why its October 2024 era breakthrough looked sudden from the outside but was earned internally.

• 4

  Culture is fragile, not a moat you can buy: His blunt diagnosis for cash-rich labs that still cannot ship state-of-the-art systems is that they skipped hardship, never defined the real P0, and let culture fray because mission-aligned actions stopped matching rhetoric.

• 5

  Community backlash is becoming a real compute bottleneck: Citing a Stanford talk from General Matter founder Scott Nolan, he says data centers should consider charging something like $4.50 instead of $4 per GPU-hour and returning the extra $0.50 directly to local communities.

• 6

  His most personal AI bet is end-of-life prediction: Drawing on work with Stanford Med professor Nigam Shah and a 12 million-patient dataset, Midha argues even relatively simple models can improve palliative-care decisions, but regulation still blocks shifting liability from doctors to AI systems.