Skip to content
Open Doors Partners

The New Constraint: Why AI Infrastructure Is Now an Energy Problem

The capital required to build the AI era has arrived. The question now being answered unevenly, and with consequence is whether the power exists to run it.

A perspective from Open Doors Partners

_____________________________________________________________________________________________________________________________________

Amazon’s projected capital expenditure for 2026 exceeds the entire annual investment budget of the US energy sector. One company, in a single year, is committing more to infrastructure than all of American energy exploration, refining, and delivery combined. That comparison is not offered as a statement about Amazon. It is offered as a statement about the scale of what is being built, and the particular constraint that scale has now encountered.

The five largest US cloud and AI infrastructure providers have committed collectively to between $600 and $630 billion in capital expenditure this year. Morgan Stanley estimates roughly $2.9 trillion in global data centre construction will flow through the global economy by 2028, with more than 80 percent of that spending still ahead. The Stargate project — a joint undertaking between OpenAI, SoftBank, and Oracle  has announced commitments toward a 10-gigawatt US buildout, with nearly $400 billion in capital pledged within the first three years.

These figures are consequential. But they are not the most consequential thing happening in this market.

From a Capital Constraint to a Power Constraint

Throughout 2024 and into 2025, the primary challenge facing AI infrastructure was financing. Matching the capital intensity of frontier AI workloads with instruments suited to long-duration, asset-heavy deployment required new thinking. The market answered with institutional depth. Industry-wide debt issuance reached $182 billion in 2025, nearly doubling the prior year. GPU-collateralised financing emerged as a new instrument class: Nscale secured $1.4 billion in delayed-draw facilities backed by compute hardware. Google issued roughly $29 billion in debt for data centre expansion, including a century bond. The full spectrum of institutional capital, from secured and unsecured to structured and securitised, moved into AI infrastructure financing. Morgan Stanley has identified a $1.5 trillion financing gap for the global buildout; sovereign and private capital are both moving to close it.

Capital found its form. What remains constrained is not money. It is electrons.

Microsoft disclosed an $80 billion backlog of Azure orders it cannot fulfil due to power constraints. Demand exists. Capital exists. Infrastructure cannot be deployed because the grid cannot absorb it. Meta has secured agreements for more than 6 gigawatts of nuclear-linked generation, sufficient to power five million homes, specifically to obtain firm, dispatchable power for AI workloads. Three of the largest hyperscalers have committed to tripling global nuclear production by 2050. The capital intensity of these businesses, Oracle at 57 percent of revenue and Microsoft at 45 percent, is at levels that have no historical precedent in technology. These are not technology companies that need electricity. They are, structurally, energy companies that produce compute.

The Scale of the Power Problem

The IEA’s projections make the constraint legible. Data centres consumed 415 terawatt-hours globally in 2024, roughly 1.5 percent of world electricity. By 2030, that figure is projected to reach 945 terawatt-hours, equivalent to Japan’s current total consumption, reached in six years. In the United States specifically, AI data centre power demand is projected to reach 123 gigawatts by 2035, a 30-fold increase from current levels. BloombergNEF revised its 2035 global data centre power demand forecast upward by 36 percent in the span of seven months, to 106 gigawatts. The revision is itself a signal: the constraint is not static. It is accelerating.

The geography of AI infrastructure is reorganising around this reality. Northern Virginia, the traditional centre of US data centre concentration, is approaching saturation. Development is shifting to central and southern Virginia, Georgia, and Texas, where available grid capacity offers operational room. Internationally, Southeast Asia and the Middle East are gaining share not on the basis of demand proximity or labour cost, but on the basis of what the constrained US grid cannot reliably offer: power at scale. The selection criteria for where AI infrastructure is built have changed. Power availability now ranks above connectivity, real estate cost, and in some cases regulatory environment.

How Sovereign Capital Is Responding

Sovereign capital has recognised this dynamic before most institutional frameworks have caught up to it. Saudi Arabia has doubled its US investment target to approximately $1 trillion, with AI infrastructure as a primary focus. Sovereign wealth funds, including GIC, Temasek, and MGX, are embedded as co-investors in the largest AI model development rounds. Kuwait Investment Authority is named in Brookfield’s $100 billion AI infrastructure programme. The pattern is consistent: sovereigns are moving from enabling infrastructure investment to directing it, acquiring ownership positions in the physical systems on which the next decade of economic activity will run.

This is the second-order implication that private markets have been slower to price than credit markets have been. The asset underneath the AI infrastructure trade is not primarily the GPU cluster. It is the land-plus-power configuration that determines whether that cluster can be built and operated at scale. The valuation framework that applies to companies occupying structural positions in this configuration is not the framework venture has historically applied. These businesses carry capital intensity ratios that belong to infrastructure and industrial underwriting, the logic of utilities, transmission networks, and energy systems, not to growth-multiple technology investment. The firms that recognise this distinction early are underwriting a different asset than the firms that have not yet made the shift.

What This Means for Private Market Positioning

The constraint has moved. Capital arrived, found its instruments, and is now deployed at a scale the prior decade of technology investment would not have predicted. What is being underwritten now, at the institutional level, at the sovereign level, at the level of firms building durable positions in this category, is control over the physical systems that make the AI era operable.

The buildout is real. The question being answered, imperfectly and with consequence, is who controls the electrons.

_____________________________________________________________________________________________________________________________________

Frequently Asked Questions

Why has the primary constraint on AI infrastructure shifted from capital to power?
 Financing found its instruments — debt issuance, GPU-collateralized loans, and structured capital closed much of the funding gap through 2025. What remains scarce is grid capacity: demand for power now regularly exceeds what can be delivered, independent of how much capital is available to deploy.

Why are hyperscalers increasingly described as energy companies rather than technology companies?
 Their capital intensity ratios — Oracle and Microsoft each committing a large share of revenue to infrastructure — now resemble utility and industrial underwriting more than growth-technology investment. Securing firm, dispatchable power, including nuclear-linked generation, has become as central to their operating model as compute itself.

How is the geography of AI infrastructure development changing as a result of power constraints?
 Development is shifting away from traditionally saturated hubs like Northern Virginia toward regions and countries with available grid capacity, including central and southern US states and parts of Southeast Asia and the Middle East. Power availability now outranks connectivity and real estate cost as a site-selection criterion.

What does the power constraint mean for how private investors should underwrite AI infrastructure positions?
 The underlying asset is increasingly the land-and-power configuration that determines whether a compute cluster can be built and operated at scale, not the GPU cluster alone. That reframes the appropriate valuation logic toward infrastructure and energy-systems underwriting rather than venture-style growth multiples.

Sources

  • Futurum Research, AI Capex 2026: The $690B Infrastructure Sprint, Feb 2026
  • Global Data Center Hub, Q1 2026: The Quarter AI Infrastructure Became Energy-Constrained, Apr 2026
  • Morgan Stanley Research, AI Is Now a Macro Variable, Mar 2026
  • IEA, Global Energy and AI Report, 2025 (415 TWh / 945 TWh projections)
  • BloombergNEF, Data Centre Power Demand Forecast, 2025–2026
  • Deloitte, US AI Power Demand 2035 (123 GW / 30x figure)
  • Nscale press release, $1.4B delayed-draw term loan, 2026
  • CNBC / Bloomberg, Meta nuclear power agreements (6+ GW)
  • Investing.com, Microsoft Azure backlog ($80B, power-constrained)
  • CreditSights / Data Center Richness, hyperscaler capex verification

Open Doors Partners LLC | Registered Investment Adviser | This post is for informational purposes only and is not an offer or solicitation. Past performance is not indicative of future results. All investments involve risk, including loss of principal. Read full disclosures here.