Hyperscaler capex has entered a $500B+ annual regime with no deceleration signal yet visible in disclosed backlogs -- power, not chips, is the binding constraint on AI deployment timelines through 2028

Microsoft ($190B), Meta ($125-145B), Google ($180-190B guidance for 2027), and Amazon (scaling aggressively with $225B in Trainium commitments alone) are collectively deploying over $500B annually in AI infrastructure capex. This is not a one-year spike -- Dell's $51.3B AI backlog has a pipeline that is 'multiples of backlog, growing sequentially,' and CoreWeave's $100B remaining performance obligation extends visibility through 2028. The capital intensity is so extreme that MSFT attributed $25B+ of its capex to higher component costs alone. The critical finding from both methods is that power availability, not semiconductor supply, has become the primary bottleneck. Bloom Energy's CEO stated that 'time to power has gone from a procurement consideration to an existential necessity.' BIP CEO confirmed 'effectively no data center inventory remaining for 2026, and even 2027 is quite scarce.' SMCI cited 'customer site readiness delay' (power and networking not yet equipped) as its primary revenue headwind. Grid interconnection queues of 3-5 years are forcing hyperscalers into behind-the-meter generation, with Oracle Jupiter (2.45 GW, 100% fuel cell) and Microsoft-CVX West Texas defining a new no-grid paradigm. The physical infrastructure implications are staggering: each GW of data center capacity requires approximately $5-8B in IT equipment, meaning the 100+ GW utility pipeline alone implies $500B-800B in semiconductor and server demand. MLM reports data center aggregate volumes up 62%, PLD sees DC suppliers taking 10% of new logistics leasing (up from 5%), and CBRE says 'we can't hire enough people.' This is no longer a technology cycle -- it is an industrial mega-cycle comparable to postwar electrification.