The race to build advanced AI models has historically focused on acquiring more computing power, but emerging evidence suggests the true constraint may lie in how efficiently frontier labs utilize the GPUs they already own. According to Anjney Midha, who has worked with leading AI companies including Anthropic and Mistral, xAI's recent training runs achieved sub-10% Model FLOPs Utilization (MFU)—a stark decline from previous generations where GPT-3 reached 21%, Gopher hit 32%, and PaLM achieved 46%, with best-in-class systems now approaching 60-70% efficiency. The inefficiency reflects a broader systemic challenge: frontier AI has become fundamentally a systems engineering problem rather than a capital allocation problem. Success depends on optimizing scheduling, networking, data pipelines, parallelism, cluster reliability, and countless technical decisions that determine whether theoretical computing capacity translates into actual training progress. Midha, who is now building AMP's independent compute grid, argues that simply increasing capital expenditure on GPU purchases won't automatically yield better models without addressing the underlying operational infrastructure—a problem Google historically considered so serious that 95% utilization would have triggered an outage alert. The implications extend beyond individual laboratories. As AI infrastructure demands intensify, the industry faces a market failure in how compute resources are allocated and utilized. Midha's work suggests the next generation of AI infrastructure must prioritize alignment, efficiency, and responsibility, potentially evolving compute markets toward independent system operator models where FLOPs flow like regulated megawatts rather than through fragmented, inefficient purchasing patterns.