GPU underutilization is the real compute killer, not GPU scarcity. Bryan Bischof from Theory Ventures points out that training trillion-parameter models creates massive orchestration nightmares—keeping thousands of GPUs saturated with work is harder than just buying more hardware. The bottleneck isn't silicon availability, it's infrastructure engineering: managing distributed workloads, minimizing idle time during gradient synchronization, and handling stragglers in multi-node clusters. When you're running 10k+ GPUs for months, even 10% idle time burns millions in wasted FLOPS. The real challenge is building systems that keep every tensor core fed with data while coordinating across fault-prone network fabrics.