Physical bottlenecks

If a human-level software engineer that could run on an H100 equivalent, at current market rates for software engineers, that H100 should rent for over $250k a year. That's 15x today's spot price.

Supply can only grow ~3x/year while model capability lets the same hardware monetize far more work, so compute prices must rise sharply — a fresh mechanism by which physical inputs bind capability progress.

the four largest AI chip designers collectively consumed around 90% of global CoWoS capacity and HBM supply in 2025, while consuming only 12% of advanced logic die production

Pins down empirically which stage of the chip supply chain binds first: packaging and memory are near-saturated and slow to expand, while logic-die capacity leaves ample headroom.

Bottlenecks across the value chain, however, are reducing the likelihood of more aggressive near-term scenarios, despite booming investment and surging project pipelines.

The IEA's follow-up to its 2025 Energy and AI report: grid-connection delays and equipment bottlenecks are already trimming aggressive data-center scenarios, pushing developers toward onsite gas generation.

it would be technically feasible to distribute training for multi-gigawatt-scale clusters across dozens of sites, thousands of kilometers apart

Single-site power availability need not cap training-run scale: multi-site distributed training is technically feasible with minimal overhead — weakening the strongest version of the electricity-binds-capability argument.