The Queue, Not the Chip: Why Grid Connection Has Become the First Constraint on the Artificial Intelligence Build-Out

Semiconductors are treated as the scarce input of the artificial intelligence build-out and capital as the enabler. The International Energy Agency describes the obstacles differently, placing grid connections held up by planning and permitting first, supply chains for gas turbines and transformers next, and advanced chips among the tightening supply chains rather than at the head of them. On the Agency's own quantification, around 20 percent of planned data centre projects are at risk of delay because of strain on electricity grids.

The prevailing account of the artificial intelligence build-out treats semiconductors as the scarce input and capital as the enabler. Neither proposition is supported by the institutions that model the electricity system directly. In Key Questions on Energy and AI, published in April 2026, the International Energy Agency describes the obstacles to data centre delivery in a consistent order: planning and regulatory systems that hold up grid connections, then supply chains for energy technologies such as gas turbines and transformers, then advanced chips and other information technology components. Semiconductor supply appears in that account as one tightening supply chain among several rather than as the first constraint. The Agency's special report Energy and AI quantifies the leading constraint, finding that around 20 percent of planned data centre projects could be at risk of delay because of strain on electricity grids, with connection queues for both supply and demand projects described as long and complex. That finding sits against record capital availability. Hyperscaler capital expenditure is expected to exceed 600 billion dollars in 2026; market estimates compiled from company guidance range from roughly 600 to 690 billion dollars for the five largest spenders depending on the perimeter used, against a 2025 base near 400 billion dollars. The IEA records the same direction, noting data centre investment above 400 billion dollars in 2025 and a further increase of about 75 percent in 2026. Capital is not the binding constraint, and on the Agency's own account neither is silicon. The constraint is an electricity system whose construction cycle is measured in years while compute capacity is contracted in quarters. Two earlier editions bear directly on this argument and are distinct from it. Weekly Insight #9 examined the global shortfall in network investment, that is, how much capital the grid requires and why that capital costs more in emerging economies. The constraint examined here is different in kind: it concerns the right to connect to capacity that in many cases already exists, and the rules that allocate it. Weekly Insight #11 examined the water dimension of the same infrastructure; the electrical constraint analysed here is separate from it and prior to it.

From terawatt-hours to gigawatts: the shape of the new load

Gartner projects that global data centre electricity consumption will rise from 447 terawatt-hours in 2025 to 565 terawatt-hours in 2026, an increase of 26 percent in a single year. Within that total, consumption by AI-optimised servers rises from 95 to 175 terawatt-hours, growth of 84 percent, lifting the artificial intelligence share of data centre consumption to 31 percent in 2026; on the same projection, AI servers overtake the entire conventional server fleet in consumption during 2027. For system planners the more consequential measure is capacity rather than energy, and Gartner puts data centre power demand at 105 gigawatts in 2025 and 133 gigawatts in 2026. The IEA series points in the same direction, although the two use different definitions and coverage and should be read as parallel estimates rather than as a single reconciled series. On the Agency's figures, data centre electricity use was approximately 415 terawatt-hours in 2024, about 1.5 percent of global electricity, rising to roughly 945 terawatt-hours by 2030, slightly below 3 percent. The implied annual growth of about 15 percent is more than four times the combined growth of all other sectors. The divergence was already visible in 2025, when AI-focused data centre consumption grew by around 50 percent, total data centre consumption by 17 percent, and global electricity demand by 3 percent. Distribution matters as much as scale. On IEA figures for 2024, the United States accounted for 45 percent of global data centre electricity use, China for 25 percent and Europe for 15 percent. Goldman Sachs Research places United States data centre power demand at 31 gigawatts in 2025 and 41 gigawatts in 2026, raising the sector's share of summer peak demand from 4.1 to 5.3 percent. Rystad Energy and Carbon Brief expect Chinese data centre consumption to move from 1.2 percent of national electricity in 2024 to 2.3 percent by 2030, close to 289 terawatt-hours, growing at about 19 percent a year with installed capacity above 60 gigawatts. Wood Mackenzie expects Japanese data centre consumption to more than triple, from 19 terawatt-hours in 2024 to between 57 and 66 terawatt-hours by 2034, equivalent to 15 to 18 million households. IEEFA projects Indian data centre capacity to rise from 1.4 to 9 gigawatts by 2030, taking the sector from under 1 percent of national electricity to around 3 percent.

Where the system binds: connection queues, transformers and turbines

IEA Electricity 2026 reports more than 2,500 gigawatts of projects held in grid connection queues worldwide. That total covers renewables, storage and large loads including data centres, and no authoritative source isolates the data centre share within it; it nonetheless remains the clearest single measure of how far connection requests have outrun connection capacity. Individual pipelines make the same point at national scale. ERCOT in Texas has received more than 438 gigawatts of large-load interconnection requests, approximately 90 percent of them from data centres, roughly five times the system's peak demand. Time is the variable that binds. The IEA finds that transmission line construction takes four to eight years in advanced economies, while grid connection waiting times within the European Union range from two to ten years depending on the member state, approaching a decade in some critical zones. Delivery times for essential grid equipment such as transformers and cables have doubled over the past three years. Generation equipment is tighter still. On figures compiled by Bloomberg, Utility Dive and IEEFA, GE Vernova, Siemens Energy and Mitsubishi Power together account for around 75 percent of large-frame gas turbine output, order books extend approximately five years, and delivery positions are sold beyond 2030, reaching eight years on some production lines. Announced expansions across the three manufacturers raise combined output by only 20 to 25 percent, and the IEA notes that turbine delivery delays can push commissioning past 2030. Scarcity has repriced the asset: IRENA finds that the shortage has doubled the capital cost of a new combined cycle plant in the United States to about 2,400 dollars per kilowatt, pushing the levelised cost of gas generation toward 100 dollars per megawatt-hour in markets with high gas costs such as Italy, Germany and Japan, while it remains in the 50 to 60 dollar range where gas is inexpensive. The geography of the constraint is not American. Amsterdam will not assess new data centre applications until 2035 on grounds of grid congestion, and projects above 70 megawatts of IT capacity or 10 hectares are prohibited across most of the Netherlands. Cushman and Wakefield finds that roughly 75 percent of South Korea's operating data centre capacity and 68 percent of new projects are concentrated in the Greater Seoul area, converting a commercial siting preference into a transmission problem. In Malaysia's Johor corridor the limitation is explicitly connection rather than generation: of 51 approved projects, 17 are in operation and 11 under construction, with capacity moving from 1,025 megawatts at the end of 2025 toward more than 2,000 megawatts by the end of 2026 against a pipeline of 3,500 megawatts.

The constraint already carries a price

The clearest monetisation of scarcity is the PJM Interconnection capacity market in the United States. Clearing prices moved from 28.92 dollars per megawatt-day for the 2024/25 delivery year to 329.17 dollars for 2026/27, and two auctions have been held since. The 2027/28 auction, cleared in December 2025, settled at 333.44 dollars per megawatt-day, the ceiling then set by the Federal Energy Regulatory Commission, procuring 134,479 megawatts at a cost of about 16.4 billion dollars and falling roughly 6,500 megawatts short of the reliability requirement. The 2028/29 auction, with results published on 14 July 2026, cleared at 325 dollars per megawatt-day, which is again the regulatory ceiling itself and the third consecutive auction to clear at the cap; the cap fell by 2.5 percent between the two auctions, so the lower headline price reflects the administrative ceiling rather than any easing in the market. The auction procured 138,318 megawatts for approximately 16.4 billion dollars, 6,831 megawatts below the reliability requirement, and attracted only about 525 megawatts of new resources, of which 208 megawatts were uprates to existing plant. PJM's own simulation, as reported by Utility Dive, indicates the auction would have cleared at 555 dollars per megawatt-day without the cap, 71 percent above it. The shortfall therefore widened while the price fell, and for the first time in PJM's history the region as a whole failed to meet its reliability requirement in two consecutive auctions. Attribution here is documented rather than inferred. Monitoring Analytics, PJM's independent market monitor, calculates that data centre load accounted for 29.4 billion dollars of the 63.6 billion dollars of capacity cost across the last four auctions, or 46 percent, and that 63 percent of the price increase in the 2025/26 auction was data centre driven, equal to 9.3 billion dollars borne by consumers. PJM estimates that record capacity prices will raise some consumer bills by approximately 1.5 to 5 percent. In Washington DC, Pepco residential bills rose by an average of 21 dollars a month from June 2025, roughly half of that attributable to capacity prices, and the Natural Resources Defense Council projects cumulative additional costs of 100 to 163 billion dollars through 2033. Cost pass-through has produced a second constraint that is political rather than physical. Data Center Watch recorded more than 75 projects worth about 130 billion dollars blocked in the early months of 2026, a record period, with 833 opposition groups identified across 49 states. The documented driver of those blockages is local opposition centred on electricity prices, water use and noise rather than grid capacity itself, which makes social licence a distinct bottleneck alongside connection capacity rather than evidence of it.

The response: allocation rules, flexibility and firm supply

The industry's response appeared first in siting. Bloom Energy's 2026 Power Report describes a market in which site selection now follows available generation capacity and energisation schedules rather than customer proximity or fibre, with developers optimising for the earliest energisation rather than the lowest cost. Allocation rules are changing in parallel. The European Commission published its Electrification Action Plan, COM/2026/595, on 17 July 2026 alongside the European Grids Package. The IEA's assessment of Europe sets out why allocation has become the operative question: connection capacity is scarce in most member states, and data centres compete for the same capacity as housing, industrial electrification, electric vehicle charging and heating, which makes this a matter of public allocation rather than of one technology sector. European guidance moves connection from first-come-first-served to first-ready-first-served, adding maturity criteria, milestones, penalties and queue clean-up. In Ireland, the Commission for Regulation of Utilities lifted the Dublin connection moratorium in December 2025 and replaced it with a requirement that new data centres install on-site generation or storage sufficient to meet their full demand and export to the grid when required, the first binding regulatory application of the bring-your-own-power principle. Singapore now rations capacity through allocation, closing applications on 31 March 2026 for a call covering at least 200 megawatts. China has been relocating compute since 2022 under East Data West Computing, with eight computing hubs in the west and ten data clusters in the east. In the Gulf, power supply architecture is designed into projects from the outset: the first 200 megawatt tranche of the 1 gigawatt Stargate cluster in the United Arab Emirates is targeted for the third quarter of 2026, and Saudi Arabia's Humain programme sits within national targets of 1.9 gigawatts by 2030 and 6.6 gigawatts by 2034, although announced gigawatts should not be read as delivered capacity. The available solution set indicates that the constraint is institutional before it is physical. The Nicholas Institute at Duke University calculates that the existing United States grid could absorb at least 76 gigawatts of new load, equal to 10 percent of national peak demand, in exchange for curtailment of only 0.25 percent a year. The IEA estimates that grid-enhancing technologies could release up to 175 gigawatts of transmission capacity without building a single new line, and finds in Electricity 2026 that complementary solutions could connect 1,200 to 1,600 gigawatts of advanced-stage projects currently held in queues. PJM is developing frameworks that allow data centres to connect on condition that they operate flexibly at peak. Behind-the-meter generation is presented as an alternative to waiting in the connection queue, but IEA satellite tracking shows that many of these projects in the United States remain at an early stage because of the technical and financial hurdles involved. On firm supply, large technology companies have contracted more than 10 gigawatts of prospective new nuclear capacity over the past year, including the 835 megawatt restart of Three Mile Island Unit 1 under a twenty-year power purchase agreement worth about 16 billion dollars and targeted for 2027 to 2028, Google's 500 megawatt agreement with Kairos Power, the first corporate small modular reactor fleet agreement in the United States, Amazon's 700 million dollar investment in X-energy for up to twelve Xe-100 reactors, and Meta's stated target of up to 6.6 gigawatts through Vistra, Oklo and TerraPower. IRENA's cost data also challenges the assumption that round-the-clock supply must come from gas or nuclear, placing firm solar plus storage at 54 to 74 dollars per megawatt-hour against 60 to 95 dollars for new combined cycle gas including fuel and emissions. The pace of the artificial intelligence build-out will be set by connection queues, transformer lead times, turbine delivery positions and the rules that allocate them, rather than by chip supply or capital.

Sources cited in text: International Energy Agency — Key Questions on Energy and AI (April 2026), Energy and AI (World Energy Outlook Special Report), Electricity 2026, and the Agency's commentary on overcoming energy constraints for Europe's data centre goals; Gartner press release of 10 June 2026; Goldman Sachs Research; Rystad Energy and Carbon Brief; Wood Mackenzie; IEEFA; Cushman and Wakefield; ERCOT large-load interconnection reporting; PJM Interconnection auction reports for the 2027/28 and 2028/29 delivery years, with auction analysis reported by Utility Dive; Monitoring Analytics, PJM Independent Market Monitor; Natural Resources Defense Council; Data Center Watch; European Commission — Electrification Action Plan (COM/2026/595) and European Grids Package; Commission for Regulation of Utilities (Ireland); Singapore data centre capacity allocation call; Nicholas Institute for Energy, Environment and Sustainability, Duke University; IRENA — Renewable Power Generation Costs 2025 and round-the-clock renewables analysis; Bloom Energy 2026 Power Report; gas turbine market share, order book and delivery figures as compiled by Bloomberg, Utility Dive and IEEFA. Figures are cited from institutional sources; copyrighted charts and tables are not reproduced. Analysis by UzEnergyNews. Figures reflect public sources as of 22 July 2026.