Investor Anjney Midha on AI compute waste, AMP’s 1.2GW grid plan, and frontier systems efficiency

Anjney Midha, CEO of AMP, argues that the AI industry’s focus on acquiring more GPUs obscures a deeper problem: underutilization of existing compute. He points to reports that a frontier lab like xAI operated at sub-10% Model FLOPs Utilization (MFU), while older training runs achieved far higher rates—GPT-3 at ~21%, Gopher at ~32%, Megatron-Turing NLG at ~30%, and PaLM at ~46%. Best-in-class MFU today is closer to 60–70%, he says.

Midha attributes low utilization to misaligned incentives across the AI supply chain. He describes a "radian metaphor": small misalignments in goals between capital allocators, cluster operators, and end users compound into large inefficiencies at scale. At Google, where his co-founder Seb built the Borg/PBorg-GQM scheduler, node utilization targets of 95% were considered standard, with 96% treated as an outage threshold.

AMP’s proposed solution is a compute grid designed to "make FLOPs flow like megawatts." The company plans a 1.2GW base-load capacity with an additional 6GW of spike capacity. To achieve this, AMP advocates for open protocols, compute markets, and the role of an independent system operator—an entity akin to a grid operator—to coordinate supply, demand, and reliability across heterogeneous hardware.

Midha also critiques data center backlash and power grid constraints as external pressures shaping AI scaling. He argues that frontier AI is increasingly a systems problem: scheduling, networking, kernels, frameworks, data pipelines, parallelism, and cluster reliability determine whether theoretical FLOPs translate into real training progress. "Move fast and break things" approaches, he says, are incompatible with AI data centers, where reliability and community buy-in are critical.

The interview situates AMP’s vision within a broader critique of frontier lab dynamics. Midha highlights research hoarding—citing unpublished DeepMind work—as a market failure that reduces collective efficiency. He also emphasizes the fragility of culture in AI labs and the importance of "output maxing" as a discipline for frontier systems, where every FLOP must be aligned with measurable outcomes.