A Demand Shock With No Modern Precedent

US data centres consumed approximately 183–192 TWh of electricity in 2024 — roughly 4–5% of total US electricity supply — according to estimates from the IEA and EPRI's Powering Intelligence 2026 report. That figure was already notable. What forecasters now project for the rest of the decade is in a different category entirely.

EPRI's 2026 report — a 60% upward revision from its own 2024 baseline — projects US data centre electricity consumption of 380–790 TWh by 2030, equal to 9–17% of the entire US electricity supply. The IEA's base case, cited by Pew Research Center, puts US consumption at 426 TWh by 2030, a 133% increase from 2024. Goldman Sachs Commodities Research projects active data centre power demand rising from 31 GW in 2025 to 66 GW by end-2027, with data centres' share of US summer peak demand rising from 4.1% to 8.5% over that period.

The primary engine is artificial intelligence. AI-dedicated accelerated servers — the GPU and custom-silicon clusters underpinning large language models and inference workloads — grow at approximately 30% per year in the IEA's base case, roughly double the rate of overall data centre load growth. A typical AI-focused hyperscaler consumes electricity equivalent to 100,000 households annually; the largest facilities under construction today are expected to use roughly 20 times that amount.

The Supply Side Cannot Keep Up

The scale of the demand signal would be less consequential if the electricity supply system could respond at equivalent pace. It cannot, for structural reasons that are unlikely to resolve quickly.

Interconnection queues. Lawrence Berkeley National Laboratory's Queued Up 2025 report counted over 2,060 GW of generation and storage capacity actively seeking US grid connection at end-2024 — roughly double the entire installed US generation fleet. By early 2026, industry tallies placed the figure at approximately 2,600 GW. Only 13% of projects submitting applications between 2000 and 2019 had reached commercial operation by end-2024; 77% had been withdrawn. Average queue wait times have more than doubled since 2008, reaching approximately 5 years for the 2022 project cohort nationally, and over 8 years in PJM — the largest US grid operator.

Transformer bottlenecks. Standard power transformer lead times reached 128–144 weeks (2.5–2.8 years) as of Q2 2025, per Wood Mackenzie, against a pre-pandemic norm of 7–14 months. Some specialised orders now extend to four years. Demand for generator step-up units has risen 274% since 2019; substation power transformer demand is up 116%. Wood Mackenzie estimated a roughly 30% supply shortfall for power transformers in 2025, with shortages expected to persist through 2030. The US domestic industry covers only about 20% of national demand.

Transmission siting. High-voltage transmission projects routinely require 5–10 years from planning through permitting to energisation — while a modern data centre can be constructed in 18–24 months. This timing mismatch is not a temporary friction; it reflects decades of underinvestment in grid infrastructure that cannot be resolved through policy alone. FERC's Order 1920 (May 2024) mandates long-term regional transmission planning reform, but implementation is staggered and contested, with first planning cycles commencing by June 2026.

NERC's 2025 Long-Term Reliability Assessment, released in January 2026, projected North American summer peak demand growth of 224 GW over the next ten years — a 69% upward revision from the prior year — and found that 13 of 23 North American assessment areas now face elevated or high resource adequacy risks over the next five years.

Electricity Markets Are Already Repricing

The stress is not theoretical. PJM — which serves 65 million people across 13 eastern states — saw its 2025 capacity auction clear at $333.44/MW-day, up from $28.92/MW-day the year before, with total supplier costs rising from $2.2 billion to $14.7 billion in a single year. The auction also recorded the first capacity shortfall in PJM's history, falling 6,625 MW short of reliability requirements for the 2027/28 delivery year. Data centres account for 94% of PJM's projected 32 GW peak load growth from 2024 to 2030.

In Texas, ERCOT received 226 GW of large-load interconnection requests by November 2025 — up from 63 GW just 12 months prior — against a system peak of approximately 85.9 GW and total installed capacity of roughly 103 GW. As of Q1 2026, 198 GW of new large-load applications were filed in that single quarter alone.

Hyperscalers Are Becoming Energy Companies

The four largest hyperscalers (Amazon, Google, Microsoft, Meta) guided to approximately $725 billion in combined 2026 capital expenditure — up roughly 77% from ~$410 billion in 2025. Goldman Sachs projects $5.3 trillion in combined capex from these four firms through 2030. Confronted with grid queues stretching 5–8 years, they have responded with three parallel strategies:

Nuclear PPAs. As of May 2026, all four hyperscalers had signed nuclear power deals, with 13 announced projects committing over 9.8 GW of nuclear capacity, per the SMR Intel tracker. These include Microsoft's 20-year PPA for the full ~835 MW output of the restarted Three Mile Island Unit 1 (accelerated to H2 2027 following a FERC transmission waiver in June 2026); Meta's commitment to up to 6.6 GW spanning Vistra, Oklo, TerraPower, and Constellation's Clinton plant; Amazon's 1.92 GW expansion with Talen Energy at Susquehanna plus a $700 million investment in X-energy SMRs; and Google's 500 MW Kairos Power deal — the first US corporate SMR fleet transaction — alongside 1,800 MW with Elementl Power.

Behind-the-meter gas generation. Microsoft signed a deal with Chevron for approximately 2.67 GW of dedicated natural gas generation for a West Texas data centre, bypassing grid queues entirely. Meta is building an Ohio campus with on-site gas generation from Williams. Oracle's Stargate project uses modular gas turbines during construction. Roughly one-third of new US data centre projects are exploring private or on-site power solutions.

Renewable PPAs and grid co-investment. The four hyperscalers collectively accounted for 49% of all global clean power purchase agreement volumes in 2025, per BloombergNEF. Google's acquisition of Intersect Power brought in-house renewables development capacity; the company also signed a 1 GW solar PPA with TotalEnergies for Texas, and an agreement with Xcel Energy for 1,900 MW of clean energy in Minnesota. In March 2026, hyperscalers signed a White House pledge to co-fund grid upgrades.

Policy Is Reshaping the Economics

The policy landscape has shifted materially since 2024. The One Big Beautiful Bill Act (OBBBA), enacted in 2025, imposed an accelerated deadline for wind and solar projects to qualify for technology-neutral clean electricity credits: construction must begin before July 5, 2026, or the facility must produce electricity before January 1, 2028. Other energy technologies retain access to credits through 2033–2035. Separately, the OBBBA eliminated residential clean energy credits at end-2025.

On the regulatory side, EPA proposed repeal of all greenhouse gas emissions standards for power plants under Clean Air Act Section 111, sending a final repeal rule to OMB on May 14, 2026. More foundationally, EPA finalised the rescission of the 2009 Endangerment Finding — the legal basis for all federal GHG regulation — on February 12, 2026. Legal analysts note this faces near-certain litigation risk.

Federal permitting reform legislation remained stalled as of mid-2026, though several states — including Oregon, Washington, Massachusetts, and Illinois — enacted or opened new streamlined siting processes in 2026.

What to Watch

The most informative near-term data points are: the PJM 2026 capacity auction result (late 2026), which will confirm or extend the 2025 repricing; Wood Mackenzie's next transformer lead-time survey; NERC's 2026 Long-Term Reliability Assessment (January 2027); and the construction progress at Crane Clean Energy Center (Three Mile Island Unit 1), whose H2 2027 first-power target is the most visible test of the nuclear restart thesis.

The central tension in this story is a straightforward timing mismatch: AI infrastructure can be deployed in 12–24 months; the grid takes 5–10 years to expand. How that gap is bridged — through nuclear PPAs, behind-the-meter gas, demand flexibility, or rationed access — will determine the energy mix and cost structure of the AI era.

This post is produced by 1Oak Research for general educational and informational purposes only. It does not constitute investment advice, a solicitation, or a recommendation to buy, sell, or hold any security or asset. All investments carry risk, including the possible loss of principal. Past performance is not indicative of future results. Readers should conduct their own due diligence and consult a qualified financial adviser before making any investment decision. 1Oak Research is not a licensed or regulated financial adviser.