The Numbers First
Don’t bury the projection. NC State University’s FREEDM Systems Center, citing International Energy Agency data, projects that global AI data center energy consumption will grow from approximately 485 terawatt-hours in 2024 to roughly 945 terawatt-hours by 2030. That’s a 95% increase over six years, nearly doubling, and it represents approximately 3% of projected global energy demand by the end of the decade.
Three percent sounds modest. It isn’t. Total global electricity consumption in 2024 was roughly 29,000 terawatt-hours. By 2030, IEA projects growth to approximately 31,000 to 33,000 TWh depending on electrification pace. Three percent of that is a dedicated industrial load the size of several large European countries’ total annual electricity consumption, committed entirely to AI infrastructure. No single industrial sector has added this much demand to global grids this fast in recent memory.
These are projections, and projections carry uncertainty. IEA has revised its data center energy estimates upward multiple times over the past five years as AI deployment accelerated beyond earlier forecasts. The 945 TWh figure could be conservative if training runs scale faster than expected, or optimistic if hardware efficiency gains outpace deployment growth. Both scenarios have real precedents. Treat it as a directional signal with meaningful uncertainty bounds, not a precise forecast.
Where the Power Has to Come From
The energy source question is the constraint beneath the constraint. Building more data centers requires power. Powering data centers cleanly requires renewable generation at scale. Building renewable generation at scale requires permitting, construction, and grid interconnection, each of which has its own timeline, and none of which is fast.
The current U.S. grid interconnection queue exceeds 2,000 gigawatts of proposed generation capacity, the majority of it solar and wind. Average interconnection timelines have extended to five or more years in many regions. That timeline mismatch, data centers that can be built in two to three years waiting for power that takes five or more years to connect, is already showing up as a binding constraint in hyperscaler siting decisions. Northern Virginia, which hosts the world’s largest data center concentration, has seen Dominion Energy signal capacity limits to new customers. That’s not a future risk. It’s a current operational condition.
AI companies are responding with two strategies. Some are contracting directly with nuclear operators, Microsoft’s deal with Constellation Energy to restart Three Mile Island’s Unit 1 is the most visible example. Others are co-locating with natural gas generation, which offers reliability but undercuts clean energy commitments. Neither path scales to 945 TWh at current deployment velocity. The grid infrastructure problem is structural, and it runs ahead of any individual company’s solution.
Who Bears the Cost: A Stakeholder Map
Infrastructure investors face the most direct financial exposure. Data center assets underwritten with optimistic power purchase agreement assumptions, say, flat or declining power costs through 2030, are carrying unpriced risk. As grid access tightens and power costs rise, operating cost projections built on those assumptions will compress margins. The deals most exposed are those in high-demand markets with constrained grid capacity: Northern Virginia, Phoenix, Chicago, and the Texas ERCOT zone.
AI hyperscalers and cloud providers have enough scale and procurement sophistication to manage near-term exposure through long-term power contracts and direct generation investments. The risk for them is less about input costs and more about permitting timelines and community opposition. A data center that can’t break ground because local utility infrastructure isn’t ready isn’t a financial loss, it’s an opportunity cost that compounds as AI capacity demand grows faster than supply.
Enterprise AI buyers sit one level removed but aren’t insulated. If hyperscalers face rising energy costs, those costs migrate into compute pricing. The widely-cited declines in inference cost-per-token over 2024 and 2025 were driven partly by hardware efficiency and partly by favorable energy pricing. A tighter energy market slows that cost curve. For enterprises building AI-dependent products around current cost assumptions, that matters.
Grid operators and utilities face the opposite problem: a demand surge they didn’t plan for, arriving faster than their capital programs can respond. Investor-owned utilities in data center corridors are filing rate cases and infrastructure investment plans to address this, but regulatory approval timelines are long. The gap between when the demand arrives and when the infrastructure investment is approved creates reliability risk – which is precisely the community-level concern driving the legislative proposals now emerging.
Communities hosting data centers bear costs that rarely show up in investment analyses: grid stress that affects all local ratepayers, water use for cooling systems, and land use displacement. These are the constituencies that proposed moratorium legislation is designed to protect. Whether any specific bill passes is secondary. The political dynamics it reflects, local backlash against large-scale AI infrastructure, are real and will shape the regulatory environment regardless of federal legislative outcomes.
The Policy Environment Taking Shape
According to reporting from Cornell University’s news service, which could not be independently verified in this cycle, lawmakers have reportedly introduced a bill proposing a moratorium on new AI data center construction. The details of that bill, including its sponsors, scope, and legislative posture, are unconfirmed pending source resolution. The hub will update this coverage when the Cornell reporting can be confirmed.
What is verifiable: the legal and policy sector is actively engaged. Foley and Lardner published a summary of key questions shaping the AI energy and data center market in March 2026, reflecting significant client interest in the regulatory landscape. Major law firms don’t publish client briefings on speculative topics. Their engagement here is a confirmation signal that energy and permitting constraints are active commercial risk factors for their data center and AI infrastructure clients.
State-level action is also accelerating. Several states with significant data center concentration, Texas, Virginia, and Georgia among them, have active proceedings at their public utility commissions related to large load interconnection requests. The outcomes of those proceedings will shape where new capacity can realistically site before federal legislation reaches a floor vote.
The Investment Implication
For investors tracking AI infrastructure as an asset class, the energy constraint introduces a differentiation signal that hasn’t been fully priced. Not all data center markets face the same exposure. Markets with ample renewable generation, utility capacity headroom, and favorable interconnection timelines, parts of the Southeast, the Pacific Northwest, and certain Midwest markets, are structurally better positioned than constrained high-demand corridors. Site selection and power contract terms are becoming alpha generators in a way they weren’t three years ago.
The broader capital implication: AI infrastructure investment thesis documents written in 2022 and 2023 that don’t address energy cost risk as a primary variable are underpriced for the current environment. That’s not a crisis, it’s a repricing event that will take several years to fully work through valuations. The companies and funds that address it explicitly now will have better decision frameworks than those that treat it as a background assumption.
TJS Synthesis
The 945 TWh projection is a directional signal, not a balance sheet item. But the directional signal is clear enough that treating AI energy demand as a future problem is no longer defensible. The constraint is present, compounding, and already affecting siting decisions, power procurement strategies, and regulatory engagement at every level of the data center economy.
The stakeholders best positioned through 2030 are those who priced the energy question early: investors who built power cost escalation into underwriting models, operators who secured long-term power purchase agreements before grid tightness became visible in spot markets, and AI companies that invested in efficiency at the hardware and software layer before margin compression forced the issue. For everyone else, the repricing is coming. The IEA data says so. The utility commission filings say so. And if the Cornell reporting confirms a federal moratorium proposal, the legislative calendar will say so too.