The Grid Can't Keep Up: Inside AI's Data Center Energy Crisis

In July 2025, residents near Meta's new data center campus in Temple, Texas, reported that their water taps ran dry. The same week, Meta announced a $27 billion joint venture to build a 5-gigawatt data center complex in Louisiana. That complex would be enough to power roughly 4 million homes. Louisiana handed them a 20-year tax rebate to do it.

These two events capture the tension at the center of AI's infrastructure moment: an unprecedented construction boom, and an equally unprecedented strain on the power and water systems that feed it.

The numbers behind that tension are worth understanding.

A Load the Size of France

Global data centers consumed approximately 415 TWh of electricity in 2024: roughly 1.5% of total global electricity consumption, equivalent to the annual output of about 47 nuclear power plants running continuously. In the United States alone, data centers consumed 176–183 TWh, representing more than 4% of national electricity demand.

AI workloads are the accelerant. A single GPU server commonly draws 3–5 kilowatts, up to 15 times more than a traditional CPU server. Large AI training clusters can exceed 10 kW per server. In a single quarter of early 2024, net additional global power demand from AI-oriented data centers alone added 2 GW. That was up 25% from the prior quarter and more than triple the level a year earlier.

By 2030, global data center electricity demand is expected to roughly double to 900–1,065 TWh. In the U.S., official DOE projections see data centers consuming 325–580 TWh by 2028. That is between 6.7% and 12% of the entire national grid, with AI as the primary driver. Deloitte projects U.S. AI-oriented data center load could reach 123 GW by 2035. In 2024, it was 4 GW.

That is a 30-fold increase in eleven years.

A Grid Built for a Different Era

The U.S. power grid was not designed for this. After years of flat electricity demand, utilities are now revising their load forecasts sharply upward. They are discovering that the infrastructure they deferred building for two decades is suddenly urgently needed.

S&P Global's 451 Research forecast U.S. data center grid load at 61.8 GW in 2025, rising to 75.8 GW in 2026 and roughly tripling by 2030. NERC's 10-year peak demand forecast jumped 24% in a single revision cycle, driven almost entirely by data centers. Data centers are now expected to account for more than half of forecast U.S. load growth over the next five years.

The concentration makes this worse. Northern Virginia, home to the world's densest data center cluster, already has a 4 to 7-year wait for new grid interconnections. Virginia data center grid demand is expected to rise from 9.3 GW in 2024 to 12.1 GW in 2025 alone.

In Texas, ERCOT received so many data center interconnection requests that the grid operator issued a moratorium on new large-load connections in some zones. AEP Ohio reported 13 GW of data center requests in its interconnection queue and introduced new data-center-specific tariffs to prevent stranded infrastructure costs from being passed to residential customers.

The result: 33% of hyperscale operators are now evaluating full on-site generation rather than waiting for the grid to catch up.

The Default Answer Is Still Fossil Fuels

When the grid can't deliver, data centers fall back on what works: diesel and natural gas.

Global diesel backup capacity at data centers grew from roughly 20 GW in 2018 to approximately 55 GW by 2024. Virginia alone permitted 27 GW of diesel generator capacity by end-2025, the equivalent output needed to power 20 million U.S. households. These aren't emergency backup systems that sit idle. During a Virginia heat wave in June 2025, operators ran their diesel fleets for hours to reduce grid load during peak demand-response events.

On-site natural gas is filling the gap where diesel isn't sufficient. Small gas turbines and reciprocating gas engines are being deployed as semi-permanent or primary power sources for AI campuses that can't secure firm grid connections on the timeline they need. The tradeoffs are real: natural gas emits roughly 40% less CO₂ per kWh than diesel, but locking in large gas plants for 20-year lifespans creates a carbon liability that clashes directly with the sustainability targets these same companies publish quarterly.

The Environmental and Energy Study Institute documented this plainly: natural gas accounts for more than 40% of U.S. data center electricity as of 2024, with fossil fuels collectively supplying about 56% of the sector's power. New demand growth is largely being met by gas peaker plants. The AI boom, at the infrastructure level, is currently a fossil fuel boom.

Why Solar and Wind Can't Solve This

Tech companies publish ambitious renewable energy targets. Google claims to match 100% of its electricity consumption with renewable purchases. Microsoft, Amazon, and Meta have made similar commitments. The gap between those claims and reality is instructive.

Solar and wind generate power roughly 25–40% of the hours in a year, depending on location. AI training clusters and inference workloads run continuously, at high power density, 24 hours a day. Matching intermittent renewable generation to that load profile via renewable energy certificates and power purchase agreements means the company is buying credits for power that was generated somewhere else, at a different time, and often delivered to a different grid node entirely. Critics describe this accurately as "additionality theater."

The physics of storage compounds the problem. A 1 GW data center would need approximately 20 GWh of battery storage to bridge a single cloudy, windless day. That is an installation costing $1–2 billion, using technology that degrades after 4–10 hours of discharge. Utility-scale batteries work for peak shaving. They don't work for baseload AI infrastructure.

Actual U.S. data center grid power as of 2024: 56% fossil fuels, 21% nuclear, 22% renewables. Renewables are growing via PPAs, but interconnection queues exceeding 2,000 GW globally mean solar and wind projects are waiting years to connect. AI campuses are coming online now.

The Nuclear Bet

The limitations of renewables for baseload power explain why every major hyperscaler made a nuclear deal in 2024 or 2025.

Microsoft signed a 20-year, $16 billion power purchase agreement with Constellation Energy to restart Three Mile Island Unit 1. The 837 MW Pennsylvania reactor was shut down in 2019 and will be brought back exclusively to power its AI data centers. Target: online by 2028.

Google partnered with Kairos Power for up to 500 MW of small modular reactors, starting with a 50 MW Hermes 2 reactor targeted for 2030, feeding the Tennessee Valley Authority grid to supply data centers in Tennessee and Alabama.

Amazon invested over $20 billion to build an AI data center campus directly adjacent to the Susquehanna nuclear site in Pennsylvania for direct carbon-free power access, while backing X-energy's SMR program for up to 5 GW across multiple future projects, including a 320–960 MW facility in Washington state.

Meta signed a 20-year deal for 1.1 GW from Illinois' Clinton nuclear plant and issued RFPs for an additional 1–4 GW of new nuclear capacity by the early 2030s. Oracle is planning a 1 GW data center powered by three SMRs.

Collectively, Big Tech signed contracts for more than 10 GW of U.S. nuclear capacity over roughly 18 months. That's not a hedge. It's a structural shift in how the sector thinks about power.

Small modular reactors are the longer-term bet. SMRs produce 50–300 MW per module and are factory-built for lower upfront cost ($6,000–$12,000 per kW versus $8,000–$15,000 for conventional large reactors). They can also be sited closer to load centers than traditional plants. NuScale's 462 MW design received U.S. Nuclear Regulatory Commission design approval in May 2025. Analysts estimate SMRs could meet 10% of data center power demand by 2035 if the current project pipeline materializes.

The Harvard Business Review noted the timing tension: SMR deployments target 2030–2035, but AI data center demand is accelerating now. Nuclear is the right long-term answer running on the wrong timeline for the near-term problem.

The Smarter Alternative: Using the Heat You Already Have

Every data center is a large heat source. Servers convert virtually all their electrical input into heat. A 1 GW facility generates roughly 1 GW of thermal waste that is typically blown into the atmosphere or evaporated through cooling towers. That waste heat represents an enormous amount of energy that is simply discarded.

A growing number of facilities, particularly in Northern Europe, are capturing it instead.

Denmark's atNorth data center feeds waste heat directly into the Vestforbrænding district heating network, warming more than 8,000 homes annually and replacing oil and gas boilers as part of Denmark's national fossil fuel phase-out. Finland's Fortum uses heat pumps to extract 2,200 GWh annually from a data center for 7,000 residential customers, while returning chilled water to the facility for cooling. That closed loop produces net benefit for both systems. In January 2026, CNBC documented an AWS campus in Ireland heating an entire university campus year-round through the same principle.

The enabling technology is the heat pump: a device that lifts the low-grade waste heat from servers (typically 30–40°C) to the 60–80°C required for district heating networks and hot water systems. The European Union now mandates heat recovery for data centers larger than 1 MW. The U.S. has no equivalent requirement.

The National Renewable Energy Laboratory's ESIF facility in Colorado achieves a Power Usage Effectiveness of 1.04, meaning it uses almost no energy beyond what the servers themselves consume. It does this by routing waste heat from a 10 MW data center to heat adjacent offices and laboratories. At scale, researchers at Rice University are piloting organic Rankine cycle systems that use rooftop solar collectors combined with server waste heat to generate electricity equivalent to about 5% of site consumption.

The practical limits are geographic proximity (heat piping is economical within roughly 10 km) and climate (Scandinavian winters make waste heat more useful than Arizona summers). But as European regulations tighten and U.S. scrutiny of water and power use grows, waste heat recovery is moving from optional sustainability initiative to operational cost advantage.

The Incentive Gold Rush

Meanwhile, states and cities are competing aggressively for data center investment with packages that have reached extraordinary scale.

At least 41 U.S. states now offer some form of tax exemption, abatement, or subsidy for data center construction and operation. The federal government added 100% bonus depreciation on qualified equipment through 2029 via the 2025 budget legislation. The cumulative value of these incentives runs to billions of dollars annually.

Texas handed out more than $1 billion in data center subsidies in 2025, with sales tax exemptions available for facilities investing $200 million or more. Virginia provided $732 million in data center tax abatements in 2024. Northern Virginia already hosts roughly 70% of U.S. data center capacity. Illinois made $370 million available for facilities meeting a $250 million investment threshold.

Louisiana passed HB 827 in 2024, offering a 20-year (extendable) sales and use tax rebate for facilities investing $200 million and creating 50 jobs. Indiana provides a 25 to 50-year sales and use tax exemption covering power infrastructure, equipment, and construction for facilities investing $10 million or more.

The rationale is economic development: data center construction creates thousands of peak construction jobs, and operational facilities generate significant ongoing tax revenue and local spending. Studies suggest roughly 9 out of 10 data centers would not locate in a given jurisdiction without incentives. That makes the competition self-reinforcing.

The Other Side of the Ledger

The incentive packages don't account for the full cost.

U.S. data centers consumed approximately 17 billion gallons of water in 2023, a figure projected to reach 68 billion gallons by 2028: a 300% increase. A single large AI facility can draw 500,000 to several million gallons per day, equivalent to thousands of households. In Texas, data centers are projected to consume 25 billion gallons in 2025 (0.4% of the state's total water supply), rising to 2.7% by 2030 as construction accelerates. The Texas Tribune reported that state water planners were not adequately accounting for this demand in their long-range projections.

The New York Times documented in July 2025 what happens when data center water demand meets residential supply in a constrained area: wells ran dry near Meta's Temple, Texas campus. Residents in Mesa, Arizona reported similar conditions near Google facilities. The Clean Wisconsin analysis found that a single Vantage 3.5 GW data center campus could indirectly demand 54 million gallons per day through power plant cooling alone. That is water consumption comparable to 970,000 residents, and it doesn't appear in the facility's direct reporting.

On electricity, the costs flow through utility rate structures. Grid upgrades required to serve large data center loads are recovered through rate increases distributed across all ratepayers. Dominion Energy customers in Virginia saw rate increases of 10–20% partly attributable to data center infrastructure investment. Residents are effectively subsidizing grid expansion for commercial facilities that negotiated significant tax exemptions on the back end.

In response, states including California, Minnesota, and eight others passed 2025 legislation requiring water reporting and supply impact assessments before data center approvals. Oregon suburbs proposed construction moratoriums. Developers are now facing lawsuits over indirect water use that never appeared in their disclosed consumption figures.

What Is Actually Being Built

Despite the constraints, the construction pipeline is enormous.

As of Q4 2025, the global data center construction pipeline stands at $2.3 trillion: 32% in active execution, 68% in planning. North America accounts for $1.29 trillion of that total. The top five hyperscalers (Microsoft, Google, Meta, Amazon, and Oracle) collectively plan $710 billion in capital expenditure in 2026 alone. North America has 35 GW currently under construction, with 92% of that capacity already pre-leased.

The standout projects define the scale of what "hyperscale" means in 2026:

• Meta's Hyperion campus in Louisiana: 2 GW initially, targeting 5 GW, funded through a $27 billion joint venture with Blue Owl Capital. On-site power generation is included because waiting for the grid is not an option at that scale.
• Oracle and OpenAI's Stargate campus in Abilene, Texas: 1.2 GW targeting $100 billion in total investment, housing more than 450,000 NVIDIA GPUs. First buildings were operational in 2025.
• Microsoft's Wisconsin expansion: $13 billion across 15 new facilities at and around the former Foxconn site in Mount Pleasant.
• Vantage's Shackelford County campus in Texas: 1.4 GW at $25 billion, projecting 5,000 jobs.

Texas is on track to overtake Virginia as the largest U.S. data center market by 2030, with a projected 142% increase in market share. "Frontier" markets like Texas, Ohio, and Georgia now account for 64% of new U.S. capacity additions.

The reality check: Sightline Climate's February 2026 analysis found that of the 777 data center projects totaling 190 GW announced since 2024, only 5 GW of the 16 GW slated for 2026 delivery is actually under construction. The remaining 11 GW has been announced without progress despite claimed timelines of 12–18 months. Grid interconnection delays and permitting backlogs are the primary constraints, compounded by transformer shortages. Analysts estimate that only half of announced projects may ultimately materialize on their stated schedules.

Announced Is Not the Same as Built

The gap between announcement and operation is wider than the press releases suggest. A typical hyperscale data center follows a sequence: 6–12 months of site prep and permitting, 12–18 months of construction (compressed to 9–12 months with modular prefabrication), then 3–6 months of commissioning and grid synchronization. Under ideal conditions, a project announced in early 2026 comes online in late 2027 or 2028. Under current conditions, ideal is rare.

Power is the primary bottleneck, responsible for roughly 50% of pipeline delays. Virginia's grid interconnection queue runs 4 to 7 years. Texas processes connections in 12–18 months, which explains why Texas and Louisiana are leading 2026 construction starts. The 33% of hyperscalers planning on-site gas turbines or nuclear aren't doing it for sustainability optics. They're doing it to bypass a queue they can't afford to wait in.

The second bottleneck is hardware supply chain. Transformer lead times now run 2 to 4 years, up from roughly 6 months pre-pandemic. Tariffs introduced in 2025 raised construction material costs 10–20%. Generator and switchgear orders require 12–18 months of lead time. These are the long-pole items that anchor every schedule, and no amount of modular prefabrication shortens them.

Labor is the third constraint. The U.S. construction industry faces an estimated 499,000-worker shortage in skilled trades, with data center work particularly affected by the specialized electrical and controls expertise required. Factory-built modular designs help by moving labor off-site, but they don't eliminate the installation bottleneck.

The Regulatory Counterpressure

As the construction pipeline expanded in 2025, state legislatures began pushing back. By early 2026, at least ten states had proposed or enacted moratoriums on new data center approvals, including New York, Michigan, Virginia, Ohio, Louisiana, Georgia, Vermont, and Maryland. Moratorium lengths ranged from six months to three years.

Frederick County, Maryland imposed a six-month pause. Baltimore County extended its review process through 2027. Georgia's Senate proposed a one-year moratorium effective July 2026. In Michigan, a township enacted a moratorium directly linked to a $1 billion Meta-associated project, and the project was subsequently canceled after sustained public opposition over water and grid strain.

Good Jobs First documented the spread: moratorium bills are now a standard legislative response in any state receiving large data center proposals. The pattern is consistent. Incentives attract announcements, announcements attract opposition, and opposition produces regulation.

Axios reported in February 2026 that approximately 26% of data center projects that began in 2025 slipped their timelines, with regulatory delays cited alongside power and labor as primary causes. WilmerHale's February 2026 analysis documented the patchwork that results. Maryland and Virginia now require water and energy impact studies. Ohio, Texas, and Georgia require grid impact consultations before permitting. No federal framework exists to standardize any of it.

The developers who fared best are the ones who engaged municipalities and utility companies before filing permits, not after receiving approvals. Early community engagement and transparent disclosure of water and power demands don't eliminate opposition, but they reduce the organized kind. The projects that moved fastest in 2025–2026 followed that pattern. The ones stalled in moratoriums largely didn't.

What Comes Next

The energy picture for AI data centers in 2026 is this: demand is growing at 22% per year and the grid cannot expand fast enough to meet it. Fossil fuels are filling the gap near term. Nuclear restarts and SMRs are being contracted for the medium term. Waste heat recovery is emerging as the underappreciated option for facilities willing to invest in the integration work.

The incentive race will continue because the economic math works for jurisdictions with available land and power capacity. But the backlash from water-stressed communities and ratepayers bearing grid upgrade costs will produce regulatory friction that slows the permitting pipeline. That friction is already visible in the gap between announced and actually-under-construction capacity.

The companies that will build the most reliably are the ones that treat power and water as primary design constraints from the site selection stage, not problems to solve after the lease is signed. That means co-locating with existing power infrastructure and engaging with utility and community stakeholders before breaking ground.

The grid can't keep up. The companies that understand that early will build faster. The ones that don't will spend years in interconnection queues.