The AI Infrastructure Buildout: A Comprehensive Framework for the Datacenter and Power Cycle
Note: Just a preview of a few reports coming, and we know these have been long and in-depth but we felt it important to build a baseline of data and modeling to work off going forward. Upcoming in the next few weeks.
Two growth thesis reports one on Amazon AWS, and one one Google GCP.
Weighing in on the death of SaaS narrative
Expert interview with a Sr. Fellow from Rambus on Memory trends
We are in the early stages of one of the largest infrastructure buildouts in modern history. The construction of AI-scale computing infrastructure is reshaping the datacenter, power generation, and labor markets simultaneously and there is no sign demand is slowing. This is reflected in signed leases, committed capital expenditure, and operational facilities that are only now reaching their initial capacity targets.
What makes this cycle structurally different from prior infrastructure buildouts is the natural hedge that supply constraints provide against overbuild risk. In a typical demand cycle, the risk of excess capacity increases as more projects break ground. In this cycle, every layer of constraint — from power availability to labor to equipment lead times — functions as a physical check on the pace of new supply. The result is an unusually durable pricing environment for operators who can deliver capacity.
Couple those dynamics with the reality that all the main hyperscalers and neoclouds are specifically building new shells for AI compute, rather than ripping and replacing existing infrastructure, and the greenfield demand for power for AI data centers is incredibly strong.
The Demand Signal
The demand case begins with this framing: AI model performance continues to improve with scale, and scale requires compute, and compute requires power. Industry estimates suggest that datacenter capacity required to support GPU deployments alone could reach approximately 12 GW in 2026, roughly double the 6 GW required in 2025, with incremental needs approaching 20 GW by 2027.
Hyperscaler capital expenditure is expected to reach approximately >$700 billion in 2026, up from roughly $437 billion in 2025, with many Wall Street estimates now pointing toward >$850 billion in 2027. We decompose the full $716 billion across compute hardware, physical infrastructure, networking, and power systems — revealing a structural shift from the historical 50/50 equipment-to-facility split to roughly 66/34 in favor of silicon, with significant implications for where value accrues across the infrastructure stack. Over the last decade, prior-year datacenter demand has explained virtually all of the variance in current-year hyperscaler capex. The demand already absorbed in 2025 implies continued capex acceleration through at least 2027 (conservatively), with Google, Microsoft, and Meta each individually expected to spend as much or more than the entire hyperscaler cohort spent collectively in 2023.
A critical detail that is often overlooked: the major GW-scale AI facilities (xAI Colossus 2, Oracle-OpenAI Stargate, Amazon Rainier, Meta Prometheus, Microsoft Fairwater) are only now reaching their initial operational capacity targets. The models released to date have been largely trained on Hopper-generation systems. Blackwell-trained models are forthcoming, and Rubin GPU shipments beginning in the second half of 2026 will push rack-level power requirements toward 1 MW and beyond. We are still in the early phase of a multi-year capacity absorption cycle.
The Constraint That Matters Most
We detail three compounding supply constraints in the full report, but as we have come to fully appreciate over the last year, the headline is power. Grid interconnection queues have extended to six or more years. The total interconnection queue stands at approximately 2.3 TW roughly double the entire installed generation capacity of the United States. Total generation capacity additions reached only 80% of industry estimates in 2025, with shortfalls concentrated in exactly the categories that matter most for datacenter baseload power.
The timing mismatch we identify between renewable and natural gas capacity pipelines creates a potential power supply gap in the 2027–2028 timeframe, precisely when the largest AI facilities are scaling toward full operational capacity. This gap, and how it resolves, has direct implications for datacenter operator positioning and power generation investment.
The Investment Implication
The supply-demand imbalance is flowing directly into datacenter economics. Primary metro vacancy rates sit below 2% at all-time lows. The pricing environment has shifted structurally from historical volume discounts for large tenants to a market where large power blocks command premium pricing precisely because they are scarce. For the largest operators, lease expirations through 2029 carry average rental rates 15–25% below current market, creating a multi-year embedded earnings growth tailwind independent of new capacity additions.
The demand and pricing dynamics are compelling parts of the story. But a growing share of datacenter capacity is being absorbed by a class of tenant — the neoclouds — whose financial profiles bear little resemblance to the investment-grade hyperscalers that anchor the sector's underwriting assumptions. In the full report, we build a tenant credit framework covering duration mismatches, implied cash runways, circular financing dynamics, and recovery economics that institutional investors need to underwrite the REIT, noting not all neoclouds are REITs but the key here is positioning correctly.
We believe this setup rewards those who understand where the bottlenecks are tightest and which operators are best positioned to deliver against them.
What Subscribers Get in the Full Report
Capex waterfall decomposition — the ~$716 billion disaggregated across compute hardware (~$430B), physical infrastructure (~$214B), and networking (~$72B) with per-MW unit economics, the historical shift from 50/50 to 66/34 equipment-to-facility ratio, facility cost breakdown by subsystem (cooling, electrical, backup power), and the liquid cooling and 800V DC architecture cost premiums reshaping build economics
Complete demand architecture — customer concentration analysis across the six hyperscalers driving 80%+ of U.S. leasing demand
The capacity pipeline in detail — 78 GW of projected U.S. IT load additions through 2029, geographic shift analysis showing the tenfold expansion into tertiary markets, and the primary market pricing trajectory from 2015 through 2025
The Bitcoin miner pivot — 15 GW of contracted power pivoting to AI/HPC hosting, with deal economics ($118–188/kW/month, 12–19% development yields, 75–97% NOI margins) and specific transaction analysis across Hut 8, Applied Digital, and Core Scientific
Power generation deep dive — technology-by-technology pipeline analysis covering solar, wind, natural gas, and battery storage with execution data, delay rates, and the critical renewables-to-gas handoff in 2028
Labor constraint framework — quantified analysis of the ~500,000-worker deficit facing the U.S. power industry, the electrician workforce bottleneck, and the coming 800V DC power architecture transition that compounds the skilled labor shortage
Datacenter pricing model — three-tier pricing breakdown (hyperscale, wholesale, retail) with capex per MW, development yields, and the mark-to-market opportunity embedded in expiring leases through 2029
Tenant credit framework — implied cash runway analysis across CoreWeave, Applied Digital, Crusoe, Lambda, Nebius, and xAI with duration mismatch quantification, circular financing risk assessment, the Nvidia reflexivity problem, and recovery economics showing why the asset holds value even when the tenant doesn't
REIT positioning analysis — distinct risk-return profiles across the hyperscale leasing play, the retail/interconnection franchise, and the high-beta development-stage exposure, with valuation context and catalyst timelines
Full risk catalogue — scaling law discontinuity, efficiency-driven demand destruction, execution gaps, financing and credit risk (including CoreWeave dynamics), customer concentration, regulatory headwinds, and trade policy impacts