Why American Data Centres Are Unplugging From the Grid
Data Infrastructure / Infrastructure Management | 6 min read
The AI infrastructure boom has exposed a hard constraint that no amount of capital expenditure can quickly solve: America's power grid cannot keep pace with the demand for data centre electricity. As of early 2026, US data centres consume approximately 176 TWh of electricity annually — 4.4% of the nation's total power — with the trajectory steeply upward. The result is a growing movement among hyperscalers to abandon grid connectivity entirely and generate their own power, reshaping the economics of AI infrastructure, the architecture of the US electricity system, and the relationship between Big Tech and utility regulators.
The Scale of the Power Problem
Data centre power requests now exceed 200 GW of new grid capacity — yet interconnection queues can delay connections by years. AI-optimised server racks now reach between 50kW and 100kW per rack, compared to the 5–10kW of traditional infrastructure, fundamentally redefining power delivery and cooling requirements. With 550 planned data centre projects totalling 125 GW of capacity in the global pipeline, and analysts warning of a 49 GW US generation shortfall by 2028, power availability has replaced chips as the primary scaling constraint for AI infrastructure. The combined capital expenditure from just five Big Tech companies in 2025–2026 exceeded US$320 billion — the technology sector is now outspending the entire US utility industry on energy-adjacent infrastructure by a factor of two.
The grid strain is already reaching households. Retail electricity prices have risen 42% since 2019, outpacing the 29% rise in the Consumer Price Index over the same period. Capacity market prices in PJM Interconnection — the grid operator serving 65 million people — have spiked nearly tenfold. Goldman Sachs projects that data centre power consumption will boost core inflation by 0.1% in both 2026 and 2027.
Off-Grid as a Strategic Response
Faced with congested interconnection queues and conditional or curtailable grid access, data centre operators are increasingly opting for "island-mode" or fully off-grid systems — generating and managing their own power independently. In the near term, natural gas is playing a central role: operators require dispatchable power to meet the "five nines" reliability standard (99.999% uptime), which intermittent renewables cannot currently deliver at scale without extensive storage. A 1GW gas plant may occupy around 40 acres, whereas equivalent renewable capacity requires far more land — often distant from the fibre networks data centres depend upon for connectivity.
"We're seeing this increased demand for it because the grid-tied interconnect queue is congested and this alternative may provide a way to bypass it. We're going to see this demand for off-grid projects because the juice is now worth the squeeze."
— Maura Yates, CEO, Mothership Energy
"What we've seen over the last couple of years is that movement from chips being the primary scaling constraint to power generation being the primary scaling constraint."
— Peter Perri III, Managing Partner, Jupiter Island Capital
Grid Operators Push Back: Stability vs. Access
NERC has identified sudden data centre disconnections as one of its most important emerging grid reliability risks — large facilities can disconnect simultaneously during grid disturbances, threatening system balance. ERCOT has warned that losing roughly 2,600 megawatts of demand at once could destabilise the Texas grid. PJM Interconnection is proposing that data centres either bring their own generation or accept being disconnected during demand spikes — a condition hyperscalers are actively resisting, citing reliability requirements and air-quality restrictions on diesel backup generators. The Southwest Power Pool is developing a "conditional" connection model — offering priority timing in exchange for curtailment rights during periods of grid stress.
Regulatory Frameworks Evolving at Speed
Policy responses are emerging across multiple levels. In Texas, Senate Bill 6 introduces stricter requirements for large energy users above 75MW, including mandatory on-site generation disclosure and grid support participation during emergencies. At the federal level, proposed legislation including the Decentralized Access to Technology Alternatives Act of 2026 aims to exempt fully off-grid systems from certain regulatory oversight, provided they remain electrically separate from the main grid. Virginia, Georgia, Indiana, and Washington have all enacted or proposed legislation requiring data centre operators to fund infrastructure improvements proportional to their electricity consumption. The Department of Energy is accelerating permitting for grid expansion, and FERC is evaluating reforms to interconnection queue processes that currently delay new generation by years.
Key Takeaways
- • US data centres consume ~176 TWh annually (4.4% of national power), with requests exceeding 200 GW — far beyond what current grid queues can accommodate. Analysts warn of a 49 GW US generation shortfall by 2028.
- • AI-optimised racks now operate at 50–100kW per rack vs. 5–10kW for traditional infrastructure — power has overtaken chips as the primary AI scaling constraint.
- • Hyperscalers are pursuing off-grid ("island-mode") natural gas power to bypass congested interconnection queues and guarantee 99.999% uptime for AI workloads.
- • Grid operators (NERC, ERCOT, PJM, Southwest Power Pool) are pushing data centres toward self-generation or curtailable "conditional" grid access — generating pushback from hyperscalers citing reliability and air-quality concerns.
- • Policy responses include Texas Senate Bill 6, the proposed federal Decentralized Access to Technology Alternatives Act of 2026, and state-level legislation in Virginia, Georgia, Indiana, and Washington requiring data centre operators to fund proportional grid infrastructure improvements.