Why Power Is the New Data Center Bottleneck: Grid Queues, Transformers and Behind-the-Meter

The data center power grid bottleneck is the gap between how fast AI campuses can be built and how fast utilities can connect, generate and deliver electricity to them. This article explains where the queues are longest, why transformers and turbines are scarce, how behind-the-meter generation became mainstream, and what regulators are doing about it. It is part of Kurums’ Data Center Economy hub; for the downstream effect on household bills, read our companion piece on whether data centers raise electricity prices, and for the demand numbers behind the queues see our data center electricity consumption statistics.

Why is power now the main bottleneck for data centers?

Power is the bottleneck because demand requests have grown far faster than grids can add generation, wires and transformers. A single AI training campus can need as much electricity as a mid-sized city, and utilities typically need four to six years to deliver that capacity.

The International Energy Agency estimates that data centers consumed about 415 TWh of electricity in 2024, roughly 1.5% of global use, and projects that figure could more than double to around 945 TWh by 2030. In its 2026 follow-up, Key Questions on Energy and AI, the IEA said data center electricity demand rose 17% in 2025, well ahead of the roughly 3% growth in total global electricity demand, and it explicitly named slow grid connections and tight gas turbine and transformer supply chains as the main brakes on growth.

Money is not the constraint: the IEA notes that capex by five large technology companies exceeded $400 billion in 2025 and is set to rise a further 75% in 2026 (see the AI capex supercycle). Chips can be bought; a 500 kV transmission line or a turbine factory slot cannot be bought quickly.

How big are the interconnection queues in PJM and ERCOT?

The queues are enormous and largely speculative. ERCOT tracked about 438 GW of large-load requests in May 2026, against an all-time peak demand of roughly 85.5 GW, while PJM expects large-load additions of 39 GW by 2030 in its 2026 forecast.

ERCOT: a queue five times the size of the grid

Texas has become the clearest example of demand outrunning process. According to Latitude Media, ERCOT’s large-load queue stood at about 63 GW at the end of 2024 and reached about 226 GW by November 2025, with roughly 77% of it coming from large data centers seeking connection by 2030. By late May 2026, ERCOT vice president Jeff Billo told the board the total had climbed to about 438 GW, and that more than 300 GW of those requests had no studies submitted at all. Billo acknowledged ERCOT simply did not have much information on many of them.

ERCOT’s April 2026 preliminary forecast, reported by the Texas Tribune, showed peak load of about 367,790 MW by 2032, against a record of 85,508 MW in August 2023. That sum of often-duplicated requests will not materialize, but it forced change: ERCOT moved to batch studies in July 2026, starting with a “Batch Zero” of projects that already have financing and land.

PJM: the capacity market sounds the alarm

PJM, which serves more than 67 million people across 13 states and Washington, D.C., expresses the bottleneck through prices rather than queue size. Its 2028/29 Base Residual Auction, announced in July 2026, cleared at the $325/MW-day price cap in every zone, procured about 138 GW and still fell roughly 6.8 GW short of the reliability requirement. PJM said a roughly 2 GW rise in forecast demand, largely from data center development, shaped the result.

PJM’s 2026 forecast includes large-load additions of 39 GW by 2030 and 87 GW by 2040, per an LBNL briefing. In May 2026 Ohio testimony, PJM projected summer peak rising from about 156,400 MW in 2026 to roughly 224,000 MW by 2041, alongside about 40 GW of 2022–2030 retirements, mostly coal.

Source: ERCOT; PJM; Lawrence Berkeley National Laboratory (Queued Up 2026 edition and Speed to Power briefing); Texas Tribune; Utility Dive.

Why does it take five years to connect new generation?

It takes about five years because every new plant must pass sequential grid studies, secure land and permits, and wait for transmission upgrades. LBNL measured a 61-month median from request to operation for projects completed in 2025, up from 22 months in 2008.

The demand side is only half of the problem. Data centers need new generation to serve them, and that generation sits in its own queue. LBNL’s Queued Up 2026 edition counted about 2,061 GW of active generation and storage capacity seeking interconnection at the end of 2025, roughly 1,312 GW of generation and 749 GW of storage. That total actually fell about 10% during 2025, as more than 750 GW withdrew while over 600 GW of new requests arrived.

Gas capacity in the queues jumped about 86% to roughly 253 GW, while solar, wind and storage each declined. And few queued projects get built: of requests filed from 2000 to 2020, only about 13% of capacity had reached operation by the end of 2025.

Even cleared projects stall. PJM told Ohio legislators in May 2026 that delays among approved-but-unbuilt projects stemmed from permitting (29%), supply chains (23%), transmission owners (11%) and testing (9%), among other causes. Its new Expedited Interconnection Track is capped at 10 large projects a year.

How severe is the transformer and switchgear shortage?

It is severe. PwC analysts cited by pv magazine and Reuters Events say lead times for high-capacity power transformers have stretched to as long as four years, and industry experts estimate transformer prices have risen roughly 80% over five years.

Transformers sit at every step between power plant and server rack. Wood Mackenzie data show demand for generator step-up transformers rose 274% between 2019 and 2025, and demand for substation transformers rose 116%. Domestic manufacturing capacity, and the supply of grain-oriented electrical steel that transformer cores require, has not kept pace.

Manufacturers are expanding: Hitachi Energy has committed more than $1 billion to US capacity, including a South Boston, Virginia plant due in 2028, and Siemens $421 million to a Charlotte, North Carolina factory. Until they ramp, Reuters reported in July 2026, utilities are reserving equipment years earlier than they used to.

Switchgear, cable and UPS systems face similar pressure, which is why electrical equipment makers are big winners of the build-out (see the data center supply chain).

Why are gas turbines sold out, and what does that mean for data centers?

Gas turbines are sold out because utilities, power producers and data center developers all want firm capacity at once. GE Vernova reported 122 GW of gas turbine backlog and slot reservations at the end of Q3 2026 and expects at least 130 GW by year-end.

GE Vernova’s numbers are a useful barometer. Its gas turbine backlog plus reservations rose from about 83 GW at the start of 2026 to 100 GW after Q1, 116 GW after Q2 and 122 GW after Q3. Power Engineering reported that in Q1 roughly 20% of contracted gigawatts were tied directly to data center load, with the rest held by utilities, power producers and industrials, many building for that same demand.

Delivery slots tell the same story. Heavy-duty gas turbines carry lead times of about three years, and only around 10 GW of GE Vernova capacity remained across 2029 and 2030 combined early in 2026. GE Vernova is lifting annualized output to about 20 GW in the second half of 2026 and targets 24 GW by 2028. Prices have followed demand: Wood Mackenzie projects gas turbine prices could reach about $600/kW by the end of 2027, nearly triple 2019 levels, and GE Vernova said new orders in the first half of 2026 were priced well above late-2025 orders.

What is behind-the-meter power and why are AI campuses using it?

Behind-the-meter power is generation built on or next to a data center that serves the load directly rather than through the utility grid. AI developers use it to energize campuses years before the grid can, accepting higher costs and permitting fights in exchange for speed.

Cleanview, an energy data firm, tracks 59 behind-the-meter data center projects with about 90 GW of planned self-generation capacity. The IEA’s 2026 report identified a pipeline of about 47 GW of on-site gas power projects for US data centers, although it cautioned, based on satellite tracking, that many are still at early stages.

xAI’s Colossus: the template and the controversy

At xAI’s first Colossus site in Memphis, April 2025 aerial imagery showed 35 temporary gas turbines totalling about 422 MW. For Colossus 2, Mississippi’s environmental permit board on 10 March 2026 approved 41 permanent gas turbines at a power site in Southaven, just across the state line, forming a roughly 1.2 GW plant. After an April 2026 lawsuit by the NAACP and the Southern Environmental Law Center, the company agreed to remove 69 mobile turbines by July 2027.

Stargate, Hyperion and the hybrid model

OpenAI’s first Stargate campus in Abilene, Texas, is grid-connected, but builder Crusoe acquired 29 GE Vernova LM2500XPRESS aeroderivative turbines of about 34 MW each, close to 1 GW if all are installed, for grid stability and backup in place of diesel generators. Meta’s Hyperion campus in Louisiana took a utility-led route: Louisiana regulators in August 2025 approved Entergy’s plan to build three new combined-cycle gas plants largely to serve it.

What are large-load tariffs and how do they change the economics?

Large-load tariffs are special utility rate classes for very big customers such as data centers. They typically require minimum demand charges, long contract terms, collateral and upfront payment for grid upgrades, so that other ratepayers are not left with stranded costs if a project is cancelled.

The LBNL Speed to Power briefing catalogued how fast these tariffs are spreading. Dominion Energy Virginia’s GS-5 class applies to large loads above 25 MW with high load factors; Florida Power & Light uses a 50 MW threshold; We Energies in Wisconsin uses 150 MW; Arizona Public Service starts at 5 MW; and El Paso Electric targets data centers above 50 MW. Size thresholds range from none to 500 MW, while contract durations are converging on 10 to 15 years. Ohio regulators approved a data-center-specific tariff for AEP Ohio in 2025 that obliges new data center customers to pay for a minimum share of their contracted demand whether or not they use it.

These rules change the economics in two ways. First, they make speculative or duplicate requests expensive, which should eventually shrink headline queue figures to something closer to reality. Second, they shift grid-upgrade risk onto the data center owner, raising upfront cost but improving certainty. Texas went further with Senate Bill 6 in 2025, directing the Public Utility Commission to set interconnection standards, disclosure rules for duplicate requests, demand-reduction requirements during grid emergencies and new transmission cost allocation, with five rulemakings scheduled through 2026. The fairness question of who pays for new wires is covered in depth in our article on electricity prices and data centers.

What are FERC and grid operators doing to speed up connections?

Federal regulators are forcing every major US grid operator to write clear large-load interconnection rules. On 18 June 2026, FERC issued six show-cause orders requiring PJM, MISO, SPP, CAISO, NYISO and ISO-NE to justify or reform their tariffs on study processes, cost allocation, co-location and flexible load.

The process began in October 2025, when the Department of Energy used a rarely invoked authority to direct FERC to consider a rulemaking on large-load interconnection (Docket RM26-4). Rather than one national rule, FERC chose region-specific proceedings covering study processes, protection against cost shifting, co-location and behind-the-meter generation, new services for flexible loads, and studies of generation built to serve nearby loads. Grid operators had 60 days to respond.

Earlier, in December 2025, FERC ordered PJM to create three new transmission services for co-located generation and load, after years of dispute that included its November 2024 rejection of an expanded Talen Energy and Amazon co-location deal at the Susquehanna nuclear plant. Flexible loads that can curtail during a few dozen peak hours a year are increasingly offered faster, non-firm connections.

Which regions have power available, and where is the build-out heading?

Power availability is pushing data centers away from saturated hubs toward places with spare transmission, cheap gas or stranded generation, such as West Texas, the US Midwest and Southeast, the Nordics and the Gulf states, while traditional hubs impose waits measured in years.

Northern Virginia faces multi-year waits for large new connections, and in Europe Dublin has effectively paused new large data center grid connections, with Frankfurt and Amsterdam also constrained (see our map of Europe data center investment). By contrast, Gulf states are marketing abundant, low-cost power as a strategic advantage, as covered in our article on the Gulf data center race.

Proximity still matters: the IEA found 19 of 21 large US AI data centers under construction sit within 65 km of a city of 100,000-plus people. For how Texas, Ohio and other states compete, see our coverage of US data center investment.

Growth will now hinge on four physical factors: reformed queues, transformer and turbine factory ramp-ups, on-time nuclear restarts (see nuclear power for data centers) and wider use of flexible-load agreements.

Frequently Asked Questions