A hyperscale data center campus under construction today represents a capital commitment measured in billions of dollars and a power commitment measured in gigawatts, both locked in for years.

Across the industry, companies are making billion-dollar decisions on power purchase agreements, on-site generation, and interconnection commitments. Most of that is based on one shared forecast of how much data center and power demand will grow in the future.

That leads to risk. Overcommit to power infrastructure and an operator strands capital if AI demand normalizes faster than expected. But underinvest and a competitor locks up the interconnection queue position, the power purchase agreement, or the generation site first.

Thankfully, both outcomes are avoidable with more robust forecasting.

The Consensus Forecast Is a Scenario

What The Numbers Say

One number that anchors current power procurement decisions comes from Goldman Sachs Research, which projects that global data center power demand will grow by roughly 165% by 2030 relative to 2023 levels, with an intermediate step of about 50% growth by 2027 as capacity reaches 84 gigawatts.¹ The International Energy Agency’s own modeling arrives at a similar order of magnitude, projecting global data center electricity consumption will more than double to around 945 terawatt-hours by 2030.² Gartner’s independent forecast tracks closely, projecting worldwide data center electricity demand will double between 2025 and 2030, with AI-optimized computing accounting for the large majority of that incremental growth.³

Who Is Actually Driving Them

The figures above largely assume that hyperscaler capital spending and overall industry spend continue to compound at something close to its current rate. A hyperscaler, in industry usage, is a company that owns and operates cloud computing infrastructure at a scale several orders of magnitude beyond a typical enterprise data center, built to serve millions of customers or, increasingly, to train and run its own AI models. Amazon, Microsoft, Google, and Meta are the reference examples, and their combined capital expenditure plans for 2026 sit somewhere around $725 billion, up roughly 77% from 2025 levels that were themselves records.⁴ That type of spending is the demand signal every power infrastructure decision downstream is built on.

Where The Model Can Break

Treating the expected spending trajectory in the data center sector as a fixed input rather than a variable is where financial models can go wrong. Capital expenditure of this magnitude is discretionary, reversible, and already showing signs of investor scrutiny: for example, Meta’s stock fell more than 9% in a single trading session after it raised 2026 capex guidance.⁵

The risks are also beginning to surface in financing markets. As hyperscalers rely on debt and project-level financing to fund data center construction, lenders are demanding higher yields for some projects, raising the cost of incremental capacity and creating another potential constraint on future investment.

For example, an operator pricing a fifteen-year power purchase agreement or a multi-hundred-million-dollar on-site generation asset against a 165% demand growth projection is implicitly underwriting the assumption that four companies keep spending at a historically unprecedented pace for the full life of that asset. A forecast that only runs this scenario, however well-sourced the number, is a model that has not actually stress-tested anything. It has simply formalized a bet.

Power Procurement Is a Capital Structure Decision

Three Instruments, Three Risk Profiles

The instruments available to operators trying to secure power each carry a different risk and return profile, and each is typically evaluated on its own terms rather than as part of a combined position.

Power Purchase Agreements

Long-term power purchase agreements, or PPAs, are contracts under which an operator agrees to buy electricity from a specific generation project, usually wind or solar, at a fixed or escalating price over a term that has lengthened from roughly ten to twelve years historically to fifteen to twenty years today as demand for capacity has outstripped supply.⁶ Some of these are physical, delivering power directly to the facility; many are virtual PPAs, which are financial contracts for differences that let an operator lock in a price and claim renewable energy certificates without the electricity itself ever reaching the building, since the power is generated and consumed on different parts of the grid.

On-Site Generation

On-site generation, sometimes called behind-the-meter power, involves building dedicated generation, gas turbines, fuel cells, or increasingly small modular reactors, directly at or near the facility, reducing dependence on the surrounding grid at the cost of the capital and permitting burden of owning generation assets outright.

Microgrids

Microgrids and “energy parks” go a step further, combining generation, storage, and load behind a single point of interconnection in a configuration that can, in some cases, operate independently of the wider grid entirely.

Why The Instruments Do Not Hedge Each Other

These are not interchangeable hedges. A virtual PPA leaves an operator fully exposed to local grid reliability and interconnection timing since it provides price protection without a physical power guarantee.

On-site generation solves the physical delivery and timing problem but concentrates technology, fuel, and regulatory risk on the balance sheet, and typically requires years of lead time itself. For example, turbine deliveries for new gas plants now carry lead times measured in years, and interconnection wait times for new generation projects have more than doubled over the past fifteen years to an average of roughly five years before commercial operation.⁷

A microgrid can shorten the timeline to power but introduces financing complexity, since lenders are unaccustomed to underwriting integrated generation-and-load structures at this scale.

The Covenant Analogy

Procurement, in other words, is a capital structure question. A lender evaluating covenant exposure on a leveraged loan does not look at each debt instrument in isolation; it looks at the combined maturity schedule, the combined exposure to a single counterparty or commodity, and the combined downside if conditions deteriorate.

Power portfolios deserve the same treatment. An operator holding a stack of fifteen-year physical PPAs, a virtual PPA book concentrated in one regional grid, and a partially built on-site generation asset has a specific, quantifiable combined exposure to regional power price moves, interconnection delay, and counterparty credit risk. Few operator models currently produce that combined view. Most produce a project-by-project return calculation and stop there.

Four Scenarios, Not One

A financial model built to withstand scrutiny from a sophisticated capital partner needs to run, at minimum, four distinct scenarios rather than a single consensus case with sensitivity toggles bolted on.

Growth and Normalization

The first is sustained AI demand growth, roughly the Goldman Sachs and IEA base cases described above, in which hyperscaler capital spending continues compounding and utilization rates stay high.

The second is demand normalization, in which AI infrastructure spending decelerates faster than current guidance suggests, whether from efficiency gains in model training and inference, a slowdown in enterprise AI adoption, or a repeat of the market reaction that followed the emergence of more efficient training approaches like DeepSeek’s in early 2025, which briefly raised broad questions about the return on current AI capital spending.⁸

Regulatory and Grid Disruption

The third scenario is regulatory and grid disruption, distinct from demand risk because it can strand an asset even if demand materializes exactly as forecast. Interconnection queues nationally hold nearly 2,300 gigawatts of generation and storage capacity awaiting connection, more than the country’s entire installed power capacity, and wait times in constrained regions can stretch to roughly seven years.⁹ The Federal Energy Regulatory Commission’s 2023 reform of the interconnection process, along with state-level responses like Texas’s Senate Bill 6, which now requires large loads exceeding 75 megawatts to prove site control and post transmission cost commitments before entering the queue, are actively reshaping the economics and timing of new capacity in ways a static model will miss entirely.¹⁰

The interconnection saga in PJM illustrates the potential consequences particularly clearly. Rapid data center load growth is colliding with transmission constraints and a slow pipeline for new generation, contributing to sharply higher capacity-market prices across the region. Those increases have intensified concerns that the cost of serving large new loads could ultimately be borne in part by residential and commercial ratepayers, adding political and regulatory risk to projects even where underlying data center demand remains strong.

Energy Price Volatility

The fourth scenario is energy price volatility, which operates independently of both demand and regulatory risk. Even with demand tracking forecasts and interconnection proceeding on schedule, spot power prices in constrained regional markets can move sharply based on weather, fuel prices, and the pace at which new generation actually comes online relative to new load. A portfolio heavily weighted toward variable-price grid exposure, or toward virtual PPAs whose value depends on a spread between contract and market price, carries meaningfully different downside in a high-volatility environment than one anchored in physical, fixed-price delivery.

The Decision That Has To Get Made Anyway

The capital in this sector is already being committed at a rate approaching three-quarters of a trillion dollars a year among the largest hyperscalers alone, and the operators who secure power positions now, whether through PPAs, on-site generation, or interconnection queue placement, will hold a structural advantage over those who wait.

Operators that build four-scenario models and treat their power procurement instruments as a single portfolio with combined exposure rather than a set of independent bets, and that can produce a credible downside case on demand, put themselves in a materially different position than competitors making the same billion-dollar bet off a single growth curve.

1. Goldman Sachs Research, “AI to Drive 165% Increase in Data Center Power Demand by 2030.”
2. International Energy Agency, “Energy Demand from AI.”
3. Gartner, “Gartner Says Electricity Demand for Data Centers to Grow 16% in 2025 and Double by 2030.”
4. Tom’s Hardware, “Google, Microsoft, Meta, and Amazon capex spending to hit $725 billion in 2026.”
5. Yahoo Finance, “Hyperscalers Hit $700 Billion in 2026 AI Spending Plans.”
6. Data Center Dynamics, “Do PPAs have a future in the data center sector?”
7. Hanwha Data Centers, “Data Center Grid Limitations: The Power Bottleneck.”
8. Goldman Sachs Research / American Public Power Association, “AI to Drive 165% Increase in Data Center Power Demand by 2030.”
9. Hanwha Data Centers, “Data Center Grid Limitations: The Power Bottleneck.”
10. arXiv, “Electricity Demand and Grid Impacts of AI Data Centers: Challenges and Prospects.”