1. Introduction

In June 2025, Meta CEO Mark Zuckerberg announced that the company would build a 5-gigawatt data center campus in Richland Parish, Louisiana — a facility he said would "cover a significant part of the footprint of Manhattan." The project, named Hyperion, would be powered by three new combined-cycle natural gas plants built by Entergy, generating 2.26 GW of dedicated electricity. The first 2-GW phase is targeted for completion by 2030. The total investment: at least $10 billion for the data center alone, plus $1.2 billion for a 100-mile, 500-kilovolt transmission line.

Hyperion is extraordinary, but it is not unique. It is one of more than 70 data center projects now scoped at 1 GW or more across the United States. OpenAI's Stargate program, a joint venture with Oracle and SoftBank, has reached approximately 7 GW of planned capacity, with cumulative investment estimates topping $400 billion. Amazon alone has approximately 22 GW of data center capacity planned. The four largest hyperscalers — Amazon, Google, Meta, and Microsoft — spent $413 billion in capital expenditures in 2025 and are projected to spend $600–700 billion in 2026.

This is not a cyclical investment wave. It is a structural, multi-year buildout of digital infrastructure at a scale that has no historical parallel. And it is colliding, with increasing force, against the physical and political limits of the American energy system.

The central problem is straightforward: AI data centers consume electricity at a scale and density that the U.S. grid was never designed to accommodate. A single rack of NVIDIA's latest AI training chips draws 120–140 kilowatts — as much as roughly 100 U.S. homes. A hyperscale AI campus runs tens of thousands of such racks. The Electric Power Research Institute estimates that data centers could consume 9–17% of all U.S. electricity generation by 2030, more than double current levels. In Virginia, the world's largest data center market, data centers already consume about 26% of the state's electricity; that share could reach 41–59% by 2030.

The result is a fundamental renegotiation of the relationship between private technology infrastructure and public electricity systems. Who pays for the transmission lines, substations, and power plants needed to serve these facilities? Should data center loads take priority over residential customers during grid emergencies? Can nuclear power — specifically, small modular reactors that have never been commercially deployed — arrive in time to meet the demand? And what happens to local communities, water tables, and carbon emissions when a single corporate campus consumes as much electricity as a small country?

2. The Scale of the Buildout

2.1 From Megawatts to Gigawatts

The transformation of data center scale is best understood through the numbers. According to the Lawrence Berkeley National Laboratory, the entire U.S. data center industry consumed an average load of roughly 8 GW in 2014. By 2025, U.S. data center grid-power demand had risen to approximately 62 GW. S&P Global forecasts this will climb to 76 GW in 2026 and 134 GW by 2030. The International Energy Agency projects global data center electricity demand will more than double to 945 terawatt-hours by 2030 — roughly equal to Japan's total consumption.

The driver is AI. Training a single large language model can require clusters of 100,000 or more GPUs running continuously for months. Inference — the process of serving AI models to users — is less power-intensive per query but far more widespread, and its aggregate demand is growing faster than training. AI server power density increased 11-fold between 2020 and 2025, according to the IEA, and an individual server rack in an advanced data center could have peak power demand equivalent to 65 households by 2027.

The physical manifestation of this demand is the gigawatt-scale campus. Industrial Info Resources reported at the PowerGen International 2026 conference that more than 70 U.S. projects are now scoped at 1 GW or more of peak demand. Bloom Energy's 2026 Data Center Power Report finds that about one in five data center campuses could exceed 1 GW by 2030, rising to nearly one in three by 2035. U.S. data center IT load could grow from roughly 80 GW in 2025 to about 150 GW by 2028 — effectively doubling within three years.

2.2 Capital Expenditure at Unprecedented Levels

The spending to match is staggering. TrendForce's May 2026 analysis projects that the combined capital expenditure of the world's top nine cloud service providers will reach approximately $830 billion in 2026, with annual growth revised upward from 61% to 79%. Among the four major U.S. hyperscalers:

• Amazon (AWS): Expected to exceed $230 billion in CapEx in 2026, with growth of over 50% driven by AI cloud services.
• Microsoft: Increased its CapEx outlook to $190 billion, implying approximately 130% year-over-year growth.
• Google (Alphabet): Raised its guidance to between $195 billion and $205 billion after Google Cloud revenue surged 82% year-over-year in Q2 2026.
• Meta: Revised its CapEx range upward to $125–145 billion, representing approximately 85% year-over-year growth.

These figures include GPU and custom silicon purchases, but they also include land acquisition, data center construction, mechanical and electrical infrastructure, networking, and — increasingly — direct power infrastructure investment. New data center investment grew from roughly $24 billion in 2015 to $320 billion in 2025 for the four hyperscalers combined. McKinsey has estimated that companies across the compute value chain will need to invest $5.2 trillion into data centers by 2030 to meet worldwide demand for AI alone.

2.3 The Geography of the Buildout

Data center development is expanding far beyond its traditional hub in Northern Virginia. IIR's analysis of announced project value by state reveals a dramatic geographic dispersion:

3. The Energy Grid Bottleneck

3.1 Power as the Defining Constraint

For decades, data center development decisions were driven primarily by real estate considerations: land availability, fiber connectivity, and proximity to users. In the AI era, that hierarchy has been inverted. Power strategy is now the front-end driver of site selection, campus design, and project finance.

The core problem is a timing mismatch. Hyperscale data centers can be sited, built, and commissioned in 18 to 36 months. The transmission infrastructure needed to power them takes 5 to 10 years to plan, permit, and energize. Large power transformers now average 128 weeks of lead time, with generator step-up units at 144 weeks. Demand for gas turbines is so intense that wait times have stretched to seven years, with some data centers turning to refurbished jet engines to obtain the turbines they need. The pace of high-voltage transmission construction in the U.S. has declined from 1,700 miles per year between 2010 and 2014 to just 180 miles over the past two years.

In ERCOT, the Texas grid operator, large load interconnection requests surged from about 63 GW at the end of 2024 to roughly 226 GW by December 2025 — with approximately three-quarters of these requests from data centers. Many of these are "phantom" requests: speculative or duplicative filings that allow developers to explore multiple site options simultaneously.

3.2 Regional Grid Stress

In PJM, the 2025/2026 Base Residual Auction saw capacity prices surge from $28.92 per MW-day to $269.92 — a more than nine-fold increase. E3's analysis attributes approximately 50% of this increase to load growth (primarily data centers) and 50% to market design changes, power plant retirements, and reduced accreditation of fossil resources. Dominion Energy estimates its requested pipeline totals 70,000 MW, requiring roughly 230 new substations at a cost of $6–12 billion.

Operational stability is under acute pressure. In Fairfax County, Virginia, a cluster of data centers dropped 1.5 GW of load simultaneously during grid testing. In Ashburn, over 3 GW of data center load transferred to backup diesel generators within seconds following a transmission fault — roughly 3% of total PJM demand.

3.3 The "Bring Your Own Power" Model

In response to interconnection delays, hyperscalers are increasingly pursuing behind-the-meter and dedicated generation strategies. Meta's Hyperion campus is funding three new combined-cycle gas plants (2.26 GW) built directly by Entergy, with Meta paying the full annual revenue for 15 years alongside a $1.2 billion, 100-mile 500-kV transmission line. Behind-the-meter co-location addresses timelines but raises deep equity questions about grid backup reliance and resource cannibalization.

4. The Nuclear Pivot

4.1 Hyperscalers Commit to Nuclear at Unprecedented Scale

The most consequential energy development of the AI data center buildout is the hyperscaler pivot to nuclear power. As of mid-2026, the four major U.S. hyperscalers have committed over 9.8 GW of nuclear capacity across 13 announced deals:

• Microsoft & Constellation: 20-year, $16 billion PPA to restart Three Mile Island Unit 1 (835 MW, rebranded Crane Clean Energy Center), targeting first power in H2 2027.
• Amazon & Talen Energy: 17-year, 1.92-GW PPA for Susquehanna nuclear plant in PA — the only active nuclear flow to AI data centers today. Amazon also invested $700M in X-energy for a 12-unit Xe-100 SMR deployment (320–960 MW).
• Meta & Constellation / Vistra: 20-year PPA for 1,121 MW from Clinton Clean Energy Center in IL, plus 2.6 GW across nuclear uprates in OH and PA. Partnerships with TerraPower and Oklo for ~4 GW SMR capacity.
• Google & Kairos Power: World's first corporate multi-SMR agreement for 500 MW of KP-FHR salt-cooled reactors by 2035.

The Axis Intelligence Research Nuclear Readiness Gap Index (NRGI™) stood at 80.4 as of July 2026, indicating that 80.4% of all committed nuclear capacity has yet to reach commercial operation.

4.2 Why Nuclear? The 92% Factor

The fundamental appeal of nuclear is captured in a single number: 92%. That is the capacity factor U.S. nuclear plants averaged in 2024. Wind averaged 34.3%. Solar averaged 23.4%. For an AI training cluster whose operating cost runs primarily on electricity, and whose revenue depends on uninterrupted GPU utilization, firm baseload is paramount.

4.3 The SMR Timeline Problem

Every Western SMR design remains in licensing or demonstration. First-of-a-kind SMR costs remain high ($80–150/MWh), and the IEA projects that natural gas and coal will meet more than 40% of additional data center demand through 2030, leaving a multi-year gap before SMRs deliver at volume.

5. The Political Economy of Cost Allocation

When a 1-GW campus requires billions in transmission upgrades, who pays? In Virginia, State Corporation Commission staff testified that 25 of 59 transmission lines built since 2021 were directly tied to data centers ($2.8B direct, $6.2B total indirect). In response, regulators have launched landmark tariff reforms:

• Virginia GS-5 Tariff: Requires customers ≥25 MW to sign 14-year contracts, pay minimum demand charges of 85% for transmission and 60% for generation regardless of usage, post $1.5M/MW collateral, and pay exit fees.
• Texas SB 6 & Governor Directives: Mandates strict cost-allocation on incremental grid loads, requires load curtailment during ERCOT grid emergencies, and moves to repeal data center sales tax exemptions.
• Georgia & Missouri: 12–15 year contracts, premium pricing, and site-specific upstream cost enforcement. Over 30 new large-load tariffs enacted in 2025–2026 alone.

6. Community Friction and Environmental Impact

Communities that once welcomed data centers for tax revenue are pushing back against noise, water consumption, and residential rate spikes. Loudoun County and Fairfax County in Virginia have enacted strict setback and generator noise rules. Meta's Hyperion site drew intense scrutiny over well-water runoff and traffic.

On the climate front, Nature projects U.S. data center emissions will reach 24–44 million metric tonnes of CO2 annually through 2030. EPRI estimates gas generation build rates must double or quadruple (6.6–13.7 GW/year) to meet load growth. Water consumption for direct-to-chip liquid cooling adds heavy localized stress to arid municipal aquifers.

7. Outlook and Conclusion

The AI data center buildout is entering the execution era. The decisive metric is no longer announced gigawatts, but the capacity that can be permitted, energized, and operated under socially acceptable terms.

Five imperatives will determine whether this expansion succeeds: (1) Aligning cost responsibility strictly with cost causation; (2) Accelerating supply-side permitting for transformers and transmission; (3) Maintaining realistic expectations for SMR commercialization; (4) Proactively addressing community externalities like water and noise; and (5) Expanding granular, utility-level research to protect residential ratepayers.