Artificial intelligence is colliding with electrical infrastructure that takes years to permit and build. The celebrity in this story is the chip. The throttle is steel, copper, cooling water, and a queue at the utility.
You can buy compute on a product cycle. You cannot buy a high-voltage transformer, a substation, an interconnection study, or a firm megawatt on a product cycle. Countries that can deliver electrons plus permits will capture training and inference capacity even if they do not win the model-lab race.
Software ships on a sprint. A large power transformer ships on a factory calendar. A transmission line ships on a political calendar. Anyone who treats those as the same problem will misread the AI race.
The GPU is famous for a reason. It is scarce, expensive, and photogenic. It is also the part of the stack that still moves closest to industrial time. Allocated buyers can still get racks in months. The rest of the campus is not on that clock. High-voltage equipment, interconnection studies, substations, cooling, and firm generation are. The IEA has already said the quiet part: unless grid bottlenecks are addressed, a meaningful share of planned data-center projects is at risk of delay. Treat the ranges below as ranges. A single “average wait” is usually marketing.
| Asset | Verified time range | What the clock actually measures | How solid is the number? |
|---|---|---|---|
| GPU cluster (allocated buyer) | Months, not years—if you have allocation | Factory slot plus memory. The IEA flags high-bandwidth memory tightness through at least end-2027. | Qualitative. Vendor “weeks” quotes vary by customer; not a national statistic. |
| Large power transformer | GAO 2023: roughly 12–18 months moving toward 18–36 months. DOE-funded 2026 lab analysis: up to four years; extra-high-voltage up to about 60 months | Windings, grain-oriented steel, bushings, tap changers, test bays. Bushings quoted at up to 130 weeks. | Solid as a range. Do not flatten it into one magic number. |
| Transmission (advanced economies) | IEA: typically four to eight years | Siting, right-of-way, substations. NERC counted about 41,000 miles of >100 kV lines in the 10-year stack, with at least 390 of nearly 900 projects delayed. | Solid as a typical range. Some projects never finish. |
| Large gas turbine | Five-plus years for large frames (EPRI via Reuters); Wood Mackenzie has described six-year lead times | A manufacturing slot, not a commissioned plant. The IEA reported a roughly 70% surge in gas-turbine orders in 2025. | Solid directionally. Dollar-per-kW forecasts are still forecasts. |
| Large nuclear (recent U.S. example) | Vogtle Units 3 and 4: construction start 2013, commercial operation 2023 and 2024 | Licensing, first-of-a-kind construction, labor, and a decade of schedule reality | Public record. It is one plant, not a production line. |
| Small modular reactor | None commercial in the U.S. as of August 2026. IEA put first SMRs around 2030. | Design certification, construction, fuel, and a supply chain that is not yet a fleet | Qualitative near-term. A 2026–2028 SMR campus is a press release. |
Sources for this table are listed at the end of the report. Queue megawatts are requests, not built load. Transformer months are order-to-receipt, not “idea to energized campus.”
Training made the headlines. Inference is the factory. A factory that never powers down still has to be plugged in.
Lawrence Berkeley National Laboratory’s 2024 report—the study DOE asked for—put U.S. data-center electricity at about 176 terawatt-hours in 2023, or roughly 4.4% of U.S. electricity, up from about 58 TWh in 2014. DOE relayed the same study’s 2028 range: about 325 to 580 TWh, or roughly 6.7% to 12% of U.S. electricity, depending on shipments, utilization, and cooling. That band should stay a band. Anyone quoting a single 2030 gigawatt figure as a meter reading is selling certainty the labs do not have.
The IEA’s April 2025 Energy and AI report put global data-center electricity at about 415 TWh in 2024—around 1.5% of world electricity—with the United States at 45%, China at 25%, and Europe at 15%. The 2026 update kept the shape: about 485 TWh in 2025 toward about 950 TWh in 2030. Data-center electricity rose about 17% in 2025; AI-focused sites surged about 50%. Capital expenditure at five technology companies exceeded $400 billion in 2025. That is industrial plant, not a software budget.
National percentages are a lullaby. The EIA documented Virginia’s commercial electricity sales jumping by nearly 30 million megawatt-hours between 2019 and 2025, largely from data centers. PJM’s Dominion zone—the densest cluster on earth—saw summer peak 23% higher in 2025 than in 2019, and winter peak 45% higher than 2019–20. Ireland already showed the small-grid version: data centers took more than 21% of metered electricity in 2023, and Dublin-area connections were frozen into the late 2020s. The bill, the peak, and the backlash land on a map.
Photonics can cut the energy cost of moving bits inside the building. It does not invent megawatts. See the earlier report on photonics in AI data centers and on rare-earth and critical-mineral bottlenecks. This page is the missing middle: the transformer on the pad, the study in the queue, and the plant that has to run at 2 a.m.
A large power transformer is not a commodity SKU. It is a custom machine that can take years to wind, test, and haul, and the United States does not make most of the ones it installs.
More than 90% of the power consumed in the United States passes through a large power transformer. In 2019, the Commerce Department figure cited by DOE was blunt: about 82% of large power transformers consumed here were imported. There is one domestic producer of grain-oriented electrical steel in the official record: Cleveland-Cliffs. DOE has said for years that mill cannot cover domestic demand at the quality and price of imports. Bushings and on-load tap changers are their own chokepoints. The 2026 DOE-funded gap analysis put bushing lead times at up to 130 weeks. You can finish a tank and still wait on a part that bolts onto the top of it.
Capacity is being added. Not on a GPU calendar. Hitachi Energy announced a $457 million large-transformer plant in South Boston, Virginia, with operations expected in the second half of 2028 and a ramp after that. Siemens Energy has been expanding in North Carolina. Those are real factories, and they are late relative to the load already in the queue. Winding crews, test bays, and electrical steel do not appear because a hyperscaler signed a letter of intent.
There is not one queue. There is a generator queue, a load queue, and a political queue. They are easy to mash into a single scare number. That is how you get 410 gigawatts of Texas “demand” that is mostly paper.
LBNL’s Queued Up 2026 counted about 2,061 GW of generation and storage still seeking interconnection at the end of 2025, down from a peak near 2,600 GW at the end of 2023. Median time from request to commercial operation for projects completed in 2025: 61 months. Request to interconnection agreement: 45 months. Of requests filed from 2000 to 2020, only about 13% of capacity was online by the end of 2025. Most of the queue is an option, not a plant.
Data centers do not wait in the same line as a solar farm. Large-load interconnection is still often ad hoc. ERCOT told the Texas legislature in April 2026 it was tracking about 410 GW of large-load requests as of March 26, about 87% data centers. That is requests. NERC’s 2025 Long-Term Reliability Assessment, using mid-2025 forecasts, put ERCOT large-load additions at 45 GW by 2030, of which 23 GW were data centers. The gap between those figures is the honesty problem in this debate.
FERC Order 2023 tried to clear speculative generation with cluster studies and first-ready rules. Load is now the congestion. NERC finds 13 of 23 assessment areas facing resource-adequacy challenges over ten years, with summer peak demand forecast to rise more than 224 GW—69% above the prior year’s projection—driven mostly by data centers. Texas wrote a different option: Senate Bill 6 gave ERCOT tools to curtail new large loads in emergencies. That is a kill switch, not a solved market.
NERC itself warns that some large-load projects slow or fail to materialize in the near term even as later-year requests keep rising. Treat interconnection gigawatts as a pipeline of ambition. Count signed, studied, and financed load as the real queue.
Electricity gets the national hearings. Water gets the county meeting. Evaporative cooling is cheap in dry air and politically expensive in a drought. Air cooling and closed loops use less water and dump more heat into the local air. There is no free rejection of heat. Someone’s river, aquifer, or night-time temperature pays.
Virginia spent its 2026 session turning that fight into statute. The budget directs DEQ to designate Cooling Water Scarcity Areas and, before 2032, requires data centers there—and new facilities in the Eastern Virginia Groundwater Management Area—to certify air cooling, closed-loop, recycled, or similarly efficient systems. Arizona’s attorney general, on August 31, 2026, called for a statewide pause on new data-center approvals after a federal cut to the state’s Colorado River allocation. Pinal County has already killed rezoning requests. Dublin showed the electron version: EirGrid would not take new Dublin-area applications into the late 2020s.
This is why a Gulf desert with seawater cooling and a state-owned utility is not a curiosity. It is a permitting regime. A U.S. county that can veto a cooling tower can veto a training cluster. The model does not care. The site does.
One camp wants firm carbon-free power that does not exist yet at fleet scale. The other camp wants turbines that exist, cost more every quarter, and still take years to deliver. Both are arguing past the interconnection study.
Small modular reactors are a real technology class. They are not a 2026 supply of megawatts. As of August 2026 there is no commercial SMR operating in the United States. NuScale’s Utah municipal project collapsed in 2023 when costs would not hold. TerraPower’s Natrium plant in Kemmerer, Wyoming, received an NRC construction permit in March 2026, with completion targeted for 2030—a first-of-a-kind machine, not a catalog item. The IEA put first SMRs around 2030. A letter of intent with a nuclear developer is a hedge. It is not a substation.
Gas is the bridge people actually order. It is no longer fast. EPRI has described large-frame turbine delivery at five-plus years. Wood Mackenzie has described six-year lead times. The IEA counted a roughly 70% surge in gas-turbine orders in 2025 and warned that onsite gas is not automatically faster at scale: you overbuild generation 30% to 70% to cover variable AI load, and you still need fuel, air permits, and often a grid hook for backup. Behind-the-meter plants are spreading because the queue is slow, not because they are elegant.
The fight is political because the alternatives land on different voters. A peaker in a suburb is noise and methane. A reactor is a twenty-year argument. A transmission line is a property-rights argument. Doing none of them is how you export the inference load.
The United States still hosts the largest data-center footprint and the frontier labs. That is not the same as hosting the next wave of inference. NERC’s map is already flashing: 13 of 23 assessment areas face resource-adequacy challenges this decade, and summer peak demand is projected to rise more than 224 GW in ten years. Transmission miles in the planning stack are up; miles under construction are not up by the same amount. A country that lectures about AI leadership while counties pause campuses over water and transformers sit on a multi-year wait is inventorying speeches.
China does not need to win every model release. It needs to keep connecting load. In the IEA’s 2024 split, China already used about a quarter of global data-center electricity. IEA Electricity 2026 expects China to add about 2,600 TWh of electricity demand over five years—on the order of today’s European Union—while the United States adds more than 420 TWh, with data centers about half of that U.S. growth. Generation build in China is a state project. Permitting is not a decade-long hobby. If Washington slows sites and Beijing does not, inference follows the busbar.
The Gulf is making the same bet with fewer pretenses. Microsoft and G42 announced a 200-megawatt UAE expansion through Khazna, expected to start coming online before the end of 2026. TAQA and EWEC closed financing on a 1-gigawatt Al Dhafra open-cycle gas plant to serve data centers, alongside a giga-scale solar-plus-storage block. Developers have described Stargate UAE as sitting inside a 5-gigawatt UAE-U.S. AI campus in Abu Dhabi. Read those as announcements, not energized racks. Then notice the pattern: the state procures the plant, the campus, and the permit as one package.
Argentina is running a cousin of that strategy with law instead of oil: sell speed and legal stability to builders tired of waiting. That is the point of the earlier editorial on Argentina’s AI free-economic-zone bet. Compute is mobile. Transformers are not. Nations that offer electrons, land, and a yes will take the campuses that Northern Virginia, Dublin, and Phoenix delay.
Not another AI safety workshop. Not a tax credit for a press release. The binding constraints are factories, steel, crews, and a rulebook that can say yes to a substation without handing the bill to a household that never asked for a training run.
Treat large power transformers and grain-oriented electrical steel as defense-industrial items. Expand domestic melt and rolling for GOES. Fund test bays and winding apprenticeships, not just ribbon cuttings. A plant that opens in 2028 does not cover a 2026 queue.
Standardize large-load studies the way FERC Order 2023 tried to standardize generation: deposits, site control, cluster studies, published timelines. Kill speculative megawatts early. Publish how much requested load is contracted versus brochureware.
Order turbines and keep existing dispatchable plants if NERC says the reserve margin is going red. Fund first-of-a-kind SMRs as first-of-a-kind, with 2030s in mind. Do not tell a county that a 2032 reactor solves a 2027 campus.
Upgrades, substations, and new generation triggered by a hyperscale campus should not hide in a residential rate case. Virginia is already moving large customers onto higher minimum bills. That principle should travel.
Air, closed-loop, reclaimed, or seawater. Put the water budget in the permit. A project that cannot state its gallons in a drought year is not shovel-ready.
Pair fast permitting with pre-built substations, fiber, and an identified generation stack—the thing Argentina is advertising and the Gulf is building. A zone without a transformer is a brochure.
Working URLs as of research on August 31, 2026. Forecasts are labeled as forecasts. Queue figures are requests unless noted.
This special report is educational commentary on energy infrastructure, industrial policy, and AI siting. It is not investment, legal, engineering, or environmental advice. Figures are taken from named public sources and are often ranges or forecasts. Interconnection-queue megawatts are not the same as built load. Transformer and turbine lead times vary by voltage class, customer, and contract. Readers should check the latest utility filings, FERC and NERC documents, and manufacturer disclosures before relying on any number for a siting, procurement, or capital decision.