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

Anjney Midha, CEO of Amp, argues that the critical bottleneck for AI labs is no longer access to capital or compute, but internal culture and mission alignment. He observes that many well-funded labs suffer from a frayed connection between leadership's stated mission and daily operational actions, leading to exodus and stagnation. This is not a resource problem but a leadership failure.

The conversation shifts to infrastructure efficiency, where Midha highlights that most GPU clusters operate well below achievable utilization rates (95%+ node utilization, 60-70% model flops utilization). The root cause is organizational: too many degrees of separation between capital allocators and cluster operators creates compounding waste. His prescription is iterative, common-sense infrastructure bring-up, treating AI scaling as a reason for more discipline, not an excuse for sloppy operations. He warns that short-term, hype-driven compute procurement is creating systemic risk, particularly as communities push back on data center expansion.

Midha positions Amp as an independent system operator for compute, analogous to a power grid, aiming to pool supply and demand across clouds and silicon providers. The core insight is that vertical full-stack integration creates alignment through dictatorship, but horizontal pooling creates alignment through market mechanisms and multi-party coordination. The tradeoff is speed and control versus utilization and flexibility.

Underpinning all of this is a philosophy of "output maxing"—optimizing for outcomes given constraints. This manifests in a critique of the venture and research ecosystem that hoards research, creates adverse selection (only unpromising work gets published), and fails to recognize that top researchers possess the raw leadership capability required for executive roles. The call to action for engineering leaders is to treat culture, alignment, and infrastructure discipline as the real competitive moats, not compute hoarding.