Table of Contents
Grid delays, on-site generation and denser racks are reshaping AI infrastructure
In sum – what to know:
- Grid queues – Interconnection studies stretch for years, with projects completed in 2025 waiting a median of more than five years.
- On-site power – Developers are building their own gas generation, though only around 2GW of the 90GW announced was operating by mid-2026.
- Denser racks – Rack density is rising fast, and air cooling stops being practical around 40kW, pushing operators toward liquid cooling.
The AI buildout has hit a power wall. After two years in which access to accelerators determined who could train the biggest models, electricity has become the harder, more immediate physical delivery problem. Developers now talking about “speed to power” instead of “speed to compute” and Nvidia describing grid interconnection as a primary bottleneck for AI scaling. Gartner predicted in 2024 that 40% of existing AI data centers would be power-constrained by 2027, while Morgan Stanley estimates a US shortfall of 32–33GW through 2028, or about 34% of the 97GW of incremental power needed, even after counting generation built on site.
The limit could move again, with SemiAnalysis arguing that memory and logic chips could become the bottleneck in 2027–28 if buildings and power arrive faster than semiconductor supply. That’s a useful warning against treating today’s acute shortage as permanent, though it doesn’t diminish the immediate electrical constraint.
Power has become political too, with Gallup finding in March that 71% of Americans opposed a local data center, versus 53% for a nuclear plant, and voters favoring limits on data center development in 40 of 41 local ballot measures on November 3. The Senate was scheduled to vote at 5:30 p.m. ET Monday on a procedural cloture vote to advance the BAAJA permitting bill, not final passage. The bill would require new data centers of 20MW or more to pay the full incremental cost of the infrastructure built to serve them, but the outcome wasn’t known at the time of writing.
Demand is outrunning the grid
Interconnection, the technical and regulatory process of studying and approving a grid connection, is where much of that enormous demand gets stuck — often for years. Lawrence Berkeley National Laboratory counts roughly 8,200 projects seeking to connect 1,312GW of US generation, while projects completed in 2025 waited a median of more than five years. At PJM, which serves much of the Mid-Atlantic and Midwest, RMI finds the average timeline grew from under two years in 2008 to over eight in 2025. In June, the Federal Energy Regulatory Commission ordered all six grid operators under its jurisdiction to justify their current large-load tariffs or propose changes.
Texas shows both the extraordinary scale and the considerable uncertainty — ERCOT, the state’s main grid operator, has large-load requests totaling about 474GW, more than five times Texas’s record peak, with roughly 90% tied to data centers. Yet press coverage of a Wood Mackenzie analysis puts only 23% of disclosed requests at high confidence, making the headline figure a less reliable indicator of eventual operational demand. PJM’s capacity prices, which pay generators to be available in future years, rose from $28.92 per megawatt-day for 2024/25 to $325 for 2028/29. The market monitor’s $29.4B data-center attribution is a counterfactual model rather than a meter reading. Virginia’s Rider T-1 itemizes transmission costs partly driven by data centers on household bills, and studies range from a 2.1% residential price increase (MIT) to no significant effect (an industry-aligned think tank). China has a relative advantage in state-directed siting, though the Oxford Institute for Energy Studies finds its electricity prices comparable to US prices.
Developers are building their own power
Developers are responding with behind-the-meter generation, meaning power produced on site instead of drawn entirely from the grid, and Cleanview tracks about 90GW announced across 59 data centers, though only around 2GW was operating by mid-2026. Texas leads this uneven buildout, five states hold 83% of capacity, and Cleanview reports that OpenAI and Anthropic have signed leases for more than 10GW.
The equipment shows the commercial urgency, with developers using truck-mounted gas generators, large piston engines and turbines derived from jets. A new US gas turbine takes about six years to deliver, GE Vernova has sold slots into 2032, and used units cost roughly $1.5–1.7 million per megawatt, versus $2–2.5 million new. Anthropic’s $3.4B deal for 1.7GW of Caterpillar generators, which are reciprocating units rather than turbines, bypassed that queue — by taking over equipment ordered years earlier. Cleanview says speed now matters more than generation efficiency.
On-site power doesn’t remove the need for permits, and New Mexico twice denied a gas pipeline for Oracle’s Project Jupiter, while Oracle sent a force majeure notice to Blue Owl Capital, which owns the project developer, to delay payments if the campus misses its 2028 online date. Texas’s September 21 permit halt also covers on-site generation. Nuclear deals cover only part of future demand, while Bloom Energy’s claim of 9% lower five-year total ownership costs for a 1GW fuel-cell site, against an AC baseline it defines, comes from a vendor selling on-site power.
Inside the rack, power and cooling are changing
Inside the buildings, the hardware is changing just as quickly — Nvidia’s GB200 rack draws about 120kW, GB300 around 135–140kW, Vera Rubin 180–230kW (analyst estimates), and Kyber roughly 600kW. Blackwell raised GPU power 75% over Hopper but rack density 3.4 times, while AFCOM says average rack density rose 69% in a year, from 16kW to 27kW.
Air cooling stops being practical around 40kW, and GB200 and GB300 racks ship liquid-only. Direct liquid cooling, which circulates coolant through plates attached to hot components, reached 22% of operators in Uptime Institute’s 2024 survey and 36% in AFCOM’s 2026 survey, which uses a different measure. Nvidia specifies 800VDC, or 800-volt direct current, for Rubin Ultra and Kyber, slated for 2027 (SemiAnalysis reported in July that Kyber slipped to 2028, which Nvidia disputes), and says the same wire gauge carries 157% more power than at 415 volts AC — a substantial electrical difference at these densities.
There’s a catch — generator step-up transformers now have lead times beyond 160 weeks, while volatile AI training loads can swing from 30% to 100% in milliseconds. On July 22, about 3,800MW of load transferred to backup power after a 230 kV transmission line fault in Northern Virginia. The Congressional Research Service says data centers don’t appear to have caused blackouts to date, but electrical reliability concerns are real. Roughly 40% of US data centers sit in the most water-stressed areas, according to Business Insider. It’s still unresolved whether grids and water supplies can keep pace with increasingly dense racks.
