Who Pays

Key Takeaways

• $1.1 trillion in announced investment represents the largest private infrastructure buildout in US history
• Private equity dominance replaced REITs as primary capital source
• Debt levels create systemic risk: $600B+ in loans backed by AI demand assumptions
• Hyperscalers (Microsoft, Google, Amazon) increasingly invest directly

The Trillion-Dollar Buildout

In January 2025, President Trump announced Stargate—a $500 billion joint venture between OpenAI, Oracle, SoftBank, and NVIDIA to build AI infrastructure. The same week, Microsoft confirmed accelerated data center spending. Google, Amazon, and Meta each disclosed tens of billions in planned 2025 capex.

Add it all up—the mega-projects announced at White House photo ops, the regional expansions in Virginia and Texas and Ohio, the colocation operators adding capacity in Phoenix and Atlanta, the behind-the-meter installations and grid-scale battery facilities—and the total exceeds $1.1 trillion in announced investment across 604 documented projects.

To put that in perspective:

• The entire Interstate Highway System cost $500 billion in 2025 dollars (built over 35 years)
• The 2009 stimulus package totaled $787 billion
• The Apollo program cost $257 billion in 2025 dollars
• The Manhattan Project cost $30 billion in 2025 dollars

This is the largest private infrastructure buildout in American history. And it's happening in five years.

But who is actually paying for it? Where does $1.1 trillion come from? And what happens if the AI demand assumptions underpinning these investments don't materialize?

The Four Eras of Data Center Finance

Era 1: REITs and Public Markets (2010-2020)

For most of the 2010s, data center development was dominated by publicly-traded Real Estate Investment Trusts. Equinix, Digital Realty, CyrusOne, and others built facilities and leased capacity to enterprise customers and hyperscalers. The REIT structure offered tax advantages (no corporate income tax if 90% of income is distributed as dividends) and access to public debt and equity markets.

The model worked well for the "cloud era" demand profile: steady growth, 5-10 year lease commitments, relatively predictable capacity needs. Investment during this period totaled about $48.6 billion—substantial, but nothing compared to what came next.

Era 2: Private Equity Boom (2021-2022)

In 2021, private equity discovered data centers. The sector offered everything PE firms love: essential infrastructure, long-term contracts, predictable cash flows, and fragmented ownership ripe for consolidation.

Blackstone acquired QTS Realty Trust for $10 billion in 2021. KKR and Global Infrastructure Partners bought CyrusOne for $15 billion. DigitalBridge, Brookfield, and others launched multi- billion dollar funds specifically targeting data center assets.

Why the sudden interest? Valuation arbitrage. Public REITs traded at 15-18x EBITDA. Private equity was willing to pay 22-25x, betting they could optimize operations, consolidate platforms, and either sell to another PE firm at a higher multiple or recapitalize through refinancing.

During 2021-2022, private equity deployed $57.8 billion into data center acquisitions and developments. The model was aggressive: 60-70% debt financing, ambitious growth targets, and expectations of continued enterprise demand growth.

Then interest rates rose. The debt became more expensive to service. And AI changed everything.

Era 3: The AI Transition (2023)

CoreWeave, a cryptocurrency mining operation, pivoted to offering GPU cloud infrastructure in 2020. By 2023, they were signing billion-dollar contracts with Microsoft and raising capital at valuations exceeding established operators. Why? They had GPUs—and GPUs were suddenly the most valuable commodity in technology.

NVIDIA's data center revenue grew from $10.3 billion in fiscal 2022 to $47.5 billion in fiscal 2024 to a projected $120+ billion in fiscal 2025. The GPU shortage became so acute that startups raised funding based primarily on securing H100 allocations.

This validated two critical assumptions for investors:

1. AI inference was becoming a massive, sustained workload (not a hype cycle)
2. Traditional data center capacity was inadequate for AI power densities

Capital availability exploded. Projects that would have struggled to raise $500 million in 2020 were suddenly getting $3 billion commitments. The constraint wasn't capital—it was power, equipment availability, and construction capacity.

Era 4: Mega-Deals and Direct Investment (2024-2025)

By 2024, the scale of individual projects had grown so large that traditional financing structures couldn't accommodate them. A $10 billion data center campus requires capital beyond what most single operators or PE firms can deploy.

This created the consortium model: combinations of hyperscalers, sovereign wealth funds, private equity, and strategic investors pooling capital for mega-projects. Stargate exemplifies this— SoftBank, OpenAI, Oracle, and NVIDIA combining resources with Abu Dhabi's MGX sovereign fund.

But the most significant shift is hyperscaler direct investment. Microsoft, Google, Amazon, and Meta collectively plan to spend over $200 billion on capex in 2025, much of it for owned-and- operated data centers rather than leased capacity.

Why build instead of lease? Three reasons:

• Availability - No one else can deliver the scale and timeline needed
• Customization - AI workloads require specialized power, cooling, and networking
• Economics - At $10B+ per campus, ownership beats leasing on net present value

The Blackstone Thesis

When Blackstone acquired QTS in 2021, they published an investment thesis that has shaped the entire sector: data centers are "digital infrastructure" comparable to cell towers, fiber networks, and energy transmission—essential, long-lived assets that will grow regardless of economic cycles.

The thesis had several components:

1. Demand Durability - Digital transformation is structural, not cyclical
2. High Barriers to Entry - Grid constraints and capital requirements limit competition
3. Inflation Protection - Lease contracts often include power pass-throughs and escalators
4. Consolidation Opportunity - Fragmented market with room for platform building

Most importantly, Blackstone projected the total addressable market would reach $1 trillion by 2030. In 2021, this seemed aggressive. By 2025, it looks conservative—we're already at $1.1 trillion in announced projects.

Other institutional investors followed Blackstone's lead. Pension funds, endowments, insurance companies, and sovereign wealth funds piled into the sector, viewing data centers as infrastructure with equity-like returns.

Investment Growth Curve

The acceleration of capital deployment is remarkable:

• 2010-2020: $48.6 billion total investment (~$4.4B/year)
• 2021-2022: $57.8 billion total investment (~$28.9B/year)
• 2023: ~$85 billion (estimated)
• 2024: ~$140 billion (estimated)
• 2025: $200+ billion (projected based on announced commitments)

This represents a 50x increase in annual deployment from 2015 to 2025. NVIDIA's revenue growth provides a parallel indicator: as GPU sales explode, the infrastructure to house those GPUs must expand proportionally.

But capital deployment is concentrated in time and geography. Northern Virginia, Phoenix, Dallas, Atlanta, and Columbus are receiving tens of billions each. Some states have zero announced projects over $100 million. This creates regional boom dynamics—construction labor shortages, electrical equipment lead times, and grid constraints—that further concentrate investment where infrastructure already exists.

Debt and Systemic Risk

Here's the uncomfortable reality: if $1.1 trillion in projects are being built, and typical financing uses 60-70% debt, that implies $600-770 billion in loans backed by the assumption that AI demand will continue growing.

Who holds this debt?

• Commercial banks (construction loans, term loans)
• Pension funds (through private debt funds)
• Insurance companies (infrastructure debt portfolios)
• Asset managers (CLOs, structured products)

This starts to look familiar. In 2006-2007, financial institutions held hundreds of billions in loans backed by the assumption that housing prices would continue rising. When that assumption failed, the system seized up.

Data center debt is different in important ways—these are productive assets generating revenue, not speculative residential construction. The underlying demand (computation) is real. But the scale and growth rate of that demand is uncertain.

What happens if:

• AI capabilities plateau, reducing inference demand growth?
• Algorithmic efficiency improves, doing more with less infrastructure?
• A recession reduces enterprise IT spending?
• Electricity costs rise faster than revenue, squeezing margins?

A facility financed at 65% debt with lease commitments covering 80% of capacity has limited margin for error. If utilization drops to 60%, or if lease rates decline 20%, the debt service coverage ratio falls below 1.0. Defaults cascade. Lenders tighten. New construction stops.

We don't know if this will happen. But the debt exposure is real, and it's held by systemically important institutions. That creates correlation risk: if AI demand disappoints, it affects $600B+ in loans simultaneously.

Hyperscaler Direct Investment

The hyperscalers' shift to direct ownership is both a validation of the sector and a competitive threat to independent operators.

When Microsoft announces $80 billion in capex for fiscal 2025, mostly for data centers, that's $80 billion in demand that won't go to Equinix, Digital Realty, or CoreWeave. The hyperscalers are vertically integrating, bringing infrastructure in-house rather than leasing.

This has several implications:

1. Market bifurcation - The largest customers are becoming competitors, leaving independent operators to serve mid-market and specialized workloads
2. Capital competition - Hyperscalers can finance at lower cost (corporate bonds at 4-5%) than independents (project finance at 7-9%)
3. Grid priority - Utilities may prioritize hyperscaler projects given scale and credit quality

But hyperscaler ownership isn't necessarily more efficient. Operating data centers is a different competency than developing cloud services. Google, Amazon, and Microsoft are all learning that real estate development, utility negotiations, and facilities management require expertise they don't naturally have.

This creates opportunity for hybrid models: hyperscalers providing capital and long-term commitments, operators providing development and management expertise. The consortium structures emerging in mega-projects reflect this division of labor.