AI's $7 Trillion Build-Out: Data Centers, Power Bottlenecks, and a Fragile Capital Loop

AI's $7T Build-Out: What Real Estate and Construction Need to Price In Now

AI is no longer a tech story-it's a capex story. The infrastructure bill for data centers, power, and grid upgrades is heading toward $7 trillion this decade. McKinsey estimates $5.2T for AI-specific infrastructure and another $1.5T for traditional IT by 2030.

The catch: revenue hasn't caught up. Generative AI is tracking $30-$55B in 2025 revenue and only $85B by 2029 per S&P Global-far behind the pace of spend. That gap is the risk you must underwrite.

Demand Narrative vs. Cash Reality

AI adoption is real across consumer and enterprise. Compute intensity is rising as models push into reasoning, multimodal, and agentic workloads. But the business model is still forming.

The dominant view: AI will be huge, but the timing is uncertain. That uncertainty is colliding with a fixed-cost base that grows every quarter-power, chips, sites, grid, and people.

Data Centers at the Tip of the Spear

Generative AI data centers are being greenlit on forward assumptions about pricing and adoption, often before grid access is firm. JPMorgan's base case: 122 GW of capacity additions between 2026-2030-a step-change from current run-rates.

Execution is constrained by power systems, qualified labor, permitting, and capital markets. Without risk-aligned ROI (and realistic delivery timelines), growth breaks.

Where the Risk Is Accumulating

• Monetization risk: Revenues may lag capex. If cash generation disappoints, expect pressure on hyperscaler cash flows, data-center valuations, and borrowing costs-especially inside highly leveraged off-balance-sheet structures. A Nasdaq survey shows ~71% of corporate AI initiatives are still early, with fuzzy ROI.
• Circular capital risk: Money is looping among hyperscalers, AI labs, developers, and chipmakers. Examples: cloud credits as "investment," chipmaker equity funding labs that then buy GPUs and cloud from the same firms, and SPVs that finance GPUs and lease them back to AI labs. It accelerates builds in good times and amplifies downside if any node pulls back.
• Funding and balance-sheet risk: Free cash flow can't carry this alone. More capex is moving into capital markets, SPVs, and third-party equity-raising sensitivity to rates, spreads, ratings, and equity valuations.
• Concentration risk in RPOs: Hyperscalers carry an estimated ~$1.5T in remaining performance obligations (RPOs) tied to a handful of AI labs. Add Google Cloud's $240B backlog, largely AI-led, and the exposure deepens. Counterparty risk now sits inside what many view as "high-quality" forward revenue.
• Execution bottlenecks: Power, grid interconnections, specialized trades, long-lead equipment, and permits can push schedules even when demand is contractually committed.
• Single-point-of-failure risk: If a major AI lab stumbles on funding, the ripple effect touches revenues, site builds, SPV debt service, and rental streams across markets. The system is tightly coupled.

How Concentrated Are the Backlogs?

Selected estimates across hyperscaler RPOs and backlogs (AI-heavy and circular in places):

• Microsoft: $625B RPO; OpenAI ~45% (~$281.3B). ~25% due within 12 months; ~2.5-year weighted duration.
• Oracle: $523B RPO; ~$300B tied to OpenAI, with material exposure to Meta, NVIDIA, and xAI. ~25% due within 12 months.
• Google Cloud: $155B RPO. ~20% due within 12 months; 9 of 10 top AI labs reportedly build on GCP.
• AWS: $160-$180B RPO. Anthropic is both a customer and partially owned; ~25% due within 12 months.
• Total: ~$1.5T in AI-linked RPOs concentrated in a small group of labs.

Case Study: Why a Single Lab Matters

OpenAI functions like an anchor tenant for the ecosystem. It is lining up more than $1T in spend over the next decade and is reportedly targeting up to $100B in the first half of 2026 ahead of a potential filing later in the year. Revenue ambition: $100B by 2028. Projected cash burn to 2029: $140B.

Shifts-real or perceived-in commitment from key partners can jolt public markets, widen spreads, and slow ground-up projects. Confidence is a construction input here.

System Constraints You Must Price

• Power and grid: Continuous, highly reliable power at scale is non-negotiable. Oxford Economics expects U.S. data-center demand to outpace supply additions through 2030, with bottlenecks moving from generation to grid integration and transmission.
• Renewables and storage: Fast to deploy but intermittent; storage becomes integral. Contracts increasingly look utility-like-steady cash flows but high barriers and long paybacks.
• Site obsolescence: Older facilities face expensive upgrades; some risk becoming stranded assets as compute densities climb.
• Policy and geopolitics: Export controls, industrial policy, energy regulation, and national security add friction. Geography matters as hyperscalers shift from saturated hubs (for example, Northern Virginia) to secondary and remote metros with uneven costs and timelines.

What This Means for Developers, Operators, and Capital Allocators

• Power-first underwriting: Don't sign until interconnection glidepaths, redundancy, and curtailment risk are mapped. Treat megawatts as your true critical path.
• Stage-gate capex: Tie spend to hard milestones-permits, transformers on order, switchgear delivery, capacity reservations, and contracted offtake with pricing floors.
• Tenant concentration limits: Cap exposure to any single lab or model family. Require security packages that persist through model pivots or funding shocks.
• Scrutinize RPO quality: Look through headline backlog to counterparty health, circularity, and near-term maturity schedules.
• SPV leverage stress tests: Model rate shocks, GPU resale values, and rental holidays. Add cash sweeps and maintenance covenants that bite early.
• Delivery buffers: Add float for transformers, fiber, and grid upgrades. Lock labor through multi-year frameworks; invest in training for high-voltage and liquid cooling.
• Spec vs. build-to-suit: In tight grids, prioritize BTS with minimum-take clauses, escalators indexed to power costs, and step-up pricing tied to density.
• Retrofit economics: For legacy sites, price liquid cooling, higher rack densities, and switchgear upgrades against greenfield alternatives. Don't strand cash in half-measures.
• PPAs and onsite generation: Explore hybrid stacks-renewables, storage, and firming-where policy and interconnection timelines justify it.
• Geographic barbell: Blend proven metros (with known constraints) and secondary markets with better interconnection math but higher execution risk.

2026 Watchlist for Real Assets

• Fundraising and liquidity events for major AI labs; any delay or down-round will echo into capex plans and credit spreads.
• Hyperscaler capex guidance vs. realized AI revenue run-rates.
• Interconnection queues, transformer lead times, and regional transmission approvals.
• GPU pricing and supply cadence; shifts in vendor financing and SPV terms.
• Colocation pricing floors, density premiums, and escalation mechanics.
• Secondary-market valuations for stabilized DCs and power-adjacent assets.
• Policy signals on data sovereignty, export controls, and energy-market design.

The Bottom Line

The AI build-out is massive, real, and increasingly sensitive to execution, capital costs, and timing. As spend scales, so do systemic linkages-and the chance that a single wobble transmits across leases, RPOs, SPVs, and valuations.

Play for durable cash flows, not headlines. Price counterparty risk. Secure power early. Build buffers into schedules and covenants. That's how you stay solvent through the cycle.

Need to upskill your team on AI basics to evaluate vendors, contracts, and ROI? Explore job-specific programs at Complete AI Training.