Working Paper 06
Powering the boom: AI datacenter capex, compute efficiency, and the oversupply question
September 2026
Abstract
This paper assembles a 5,138-record dataset on hyperscaler capital expenditure, compute efficiency, prices, utilization proxies, capacity pipelines, and four prior infrastructure cycles. Hyperscaler cash capital expenditure grew 31.6% per year from $23.0 billion in 2015 to $357.5 billion in 2025; per-chip FLOPS/W doubles every 3.80–3.98 years, while frontier training compute doubles every 6.2 months, a log-growth rate 7.8 times as high. With the only experimental demand-elasticity estimate at 0.5, efficiency lowers required capacity; staged 2030 US supply of 87–152 GW compares with bear, base, and bull demand of 56, 96, and 132 GW, while the current market records 1.4% vacancy and 92% preleasing.