Power for compute: the grid as AI's binding constraint

In aggregate, the electricity demand of data centres is manageable. The International Energy Agency puts global data-centre consumption at about 415 terawatt-hours in 2024, roughly 1.5 percent of world electricity, and projects about 945 terawatt-hours by 2030 in its base case, just under 3 percent. That growth of around 15 percent a year is more than four times faster than every other sector combined, yet data centres still account for less than 10 percent of global demand growth to 2030. The United States adds about 240 terawatt-hours, an increase of roughly 130 percent, and China about 175; together they supply nearly 80 percent of the growth (IEA 2025).

The aggregate number conceals the three properties that make the load difficult. It is concentrated, because data centres cluster where land, fibre, and existing grid capacity coincide. It is fast, because a campus can be operational in two to three years while the energy system serving it needs far longer. And it is firm, because servers run near continuously and cannot tolerate the intermittency that grids absorb from other customers. The constraint is therefore not energy in the abstract but delivered firm power at a named location by a named date.

Four layers must align to deliver it: interconnection and tariff rules that let a load connect and be priced fairly; grid equipment, from transformers to substations; firm generation that is actually available when the load peaks; and the customer's willingness to flex when the system is stressed. When all four move together, capacity is energised on schedule. When any one lags, the others sit idle. In the United States the stakes are already visible: Lawrence Berkeley National Laboratory estimates that data centres consumed about 4.4 percent of national electricity in 2023 and could reach roughly 7 to 12 percent by 2028 (Shehabi et al. 2024). Capital is not scarce, since the four largest US hyperscalers are guiding to roughly 700 billion dollars or more of 2026 capital expenditure, nearly double the prior year (Alphabet et al. 2026). Time is.

Power systems were designed for predictable growth of around a percent a year, and for two decades advanced economies delivered almost none. Their planning machinery reflects that history: assets are studied one at a time, in sequence, with network upgrades allocated to whoever triggers them, and regulators approve spending only against demonstrated need. The result is a process that protects incumbent customers from stranded costs at the price of speed. Berkeley Lab's Queued Up series shows the consequence on the supply side: the typical project built in 2024 took about 55 months from interconnection request to operation, against 36 months in 2015 and 22 in 2008, and of the capacity that requested interconnection between 2000 and 2019, only 13 percent had reached operation by the end of 2024 while 77 percent had withdrawn (Rand et al. 2025).

That machinery is now being applied to a customer it was not built for. A hyperscale campus requests hundreds of megawatts to a gigawatt at a single site, may file requests with several utilities at once, and expects energisation in under three years. Each request is a step-change in a forecast that planners treated as a trend. Utilities respond with deposits, minimum-take contracts, and verification requirements, but each response adds process to a system whose binding problem is already process.

Cost-of-service regulation sharpens the mismatch. A transmission upgrade or power plant is financed over fifteen to forty years, while the servers it feeds depreciate in five or six. If the customer's demand disappears early, the stranded asset falls on the remaining ratepayers. This is the legitimate reason regulators demand long contracts and collateral, and it explains why political resistance to data-centre load tracks ratepayer bills rather than technology. The slowness is partly a defensible allocation of risk, not merely bureaucratic inertia.

The delivery problem is a mismatch of clocks. The IEA notes that a data centre can be operational in two to three years. Every layer that has to supply it runs on a longer one. New generation spends years in interconnection studies. Large gas turbines, the default source of new firm capacity, carry lead times of three years or more at Siemens Energy, whose gas-turbine backlog stood at 69 gigawatts at the end of June 2026 (Utility Dive 2026a). GE Vernova reported 116 gigawatts of gas-power equipment backlog and slot reservations at the end of the second quarter, of which 53 gigawatts was firm equipment backlog and 63 gigawatts paid slot reservations, and targets at least 125 gigawatts under contract by year-end (GE Vernova 2026).

Grid hardware is no faster. Wood Mackenzie's second-quarter 2025 survey put standard large power transformers at around 128 weeks and generator step-up units at around 144 weeks, with a supply deficit near 30 percent for power transformers (Wood Mackenzie 2025). Expansion is under way, with Siemens Energy planning roughly 50 percent more transformer and gas-insulated switchgear capacity by 2030, but new plants sell into existing backlogs and add little before the end of the decade (Utility Dive 2026a). Even a fully approved project can wait on a single piece of steel and copper.

Constellation's restart of Crane, the Three Mile Island Unit 1 reactor contracted to Microsoft under a twenty-year agreement signed in 2024, shows the pattern in its purest form. The company still targets a 2027 restart, yet a PJM analysis reported in 2026 suggested that grid connection could slip toward 2031 (ANS 2026). The reactor is a physical asset of roughly 835 megawatts waiting on a paper right. The cheapest new firm megawatt is often an old one, and it is still hostage to interconnection.

The clearest price signal comes from PJM, the largest US capacity market. Its auction for the 2027/28 delivery year cleared in December 2025 at the 333.44 dollar-per-megawatt-day cap, the third record in a row, and procured about 6.6 gigawatts less than its reserve-margin target; PJM estimated the uncapped price near 530 dollars (Utility Dive 2025). The 2028/29 auction in July 2026 cleared at the 325-dollar cap in every locational area, left the market about 6.8 gigawatts short of its reliability requirement at a reserve margin near 14.4 percent against a target above 20 percent, and would have cleared at roughly 555 dollars without the cap. Only about 525 megawatts of new generation and uprates cleared, and PJM's own forecast peak for the year rose by around 2,000 megawatts, largely from large loads (PJM 2026). A price can say that capacity is scarce, but it cannot conjure turbines and transformers on a shorter clock than their lead times.

Texas shows the same dynamic from the demand side. The ERCOT large-load queue reached roughly 474 gigawatts by August 2026, more than five times the system's record peak and about 90 percent data centres. On 3 August Governor Greg Abbott called for an audit of every data centre in the queue, covering self-supply, water use, and public subsidy, and ERCOT delayed the first batch of its new large-load transmission study (Utility Dive 2026b). Under Senate Bill 6 of 2025, large loads of 75 megawatts or more connecting from 2026 must be capable of curtailment in grid emergencies (Texas Legislature 2025). The regime is moving from first-come, first-served toward verified, flexible, and increasingly self-supplied demand.

Federal regulators have reached the same conclusion. On 18 June 2026 the Federal Energy Regulatory Commission issued show-cause orders to six grid operators, PJM, MISO, SPP, CAISO, ISO New England, and the New York ISO, instead of a one-size-fits-all rule. Each must justify its tariff or reform five areas: study and transmission-service processes, cost shifting, co-location and behind-the-meter generation, a new transmission service for flexible large loads, and the study of generation serving co-located loads. The broader rulemaking, docket RM26-4, remains open (FERC 2026). The significance is the premise: a regulator that once treated load as a passive customer now accepts that large loads need their own interconnection pathway.

With the grid slow, buyers with balance sheets are paying to jump the queue, and the cheapest route is gas. GE Vernova reports data-centre orders above 5 billion dollars in the first half of 2026, more than double its full-year 2025 total, and its output of gas turbines is scheduled to reach an annual run rate of 20 gigawatts in the third quarter of 2026, 24 in 2028, and 30 by 2030 (GE Vernova 2026). Siemens Energy sees an addressable market of up to 120 gigawatts a year, roughly half in the United States (Utility Dive 2026a). Slot reservations are options, not orders, and only part of the backlog serves data centres, but the direction is unmistakable: firm capacity is being bought years ahead of need.

On-site generation goes further and bypasses the interconnection layer entirely. Behind-the-meter plants and co-located campuses trade a grid queue for turbine slots, fuel supply, and air permits, and for the risk that the plant is built at a scale that the grid-connected market later makes redundant. That is why co-location terms sit in the middle of the federal agenda and why PJM is designing expedited tracks for load paired with new generation, alongside a connect-and-manage option in which loads accept curtailment risk in exchange for faster access (PJM 2026).

The hyperscaler is becoming something close to a utility: contracting generation, funding transmission, and accepting operating constraints once reserved for regulated customers. This reallocates risk to those best able to bear it, but it also concentrates the choice of technology and location in a few firms whose planning horizon is far shorter than the assets they commission.

The strongest objection is that the constraint is overstated because the demand is. A 474-gigawatt queue against a record peak of less than a fifth of that size is not a forecast but a speculative options market, in which developers file at several sites, utilities, and states to hold positions. ERCOT itself cautioned that its own long-range forecast may be inflated, and the Texas audit exists because requests exceed any credible build-out (Utility Dive 2026b). Queue length is evidence of intent to claim power, not of power that will be used.

Demand is also uncertain on the fundamentals. The IEA's own range for 2035 runs from about 700 terawatt-hours in a Headwinds case to more than 1,700 in a Lift-Off case, a spread of nearly two and a half times that turns on efficiency gains and on whether AI revenues justify continued investment (IEA 2025). Efficiency can cut energy per unit of computation faster than expected, as the 2025 episode of a Chinese laboratory training frontier-class models at a fraction of the assumed cost suggested.

Flexibility is the more powerful rebuttal. Duke University's Nicholas Institute estimated that about 76 gigawatts of new load, equal to roughly 10 percent of US aggregate peak demand, could be connected across the largest balancing areas if it accepted curtailment for 0.25 percent of the year, rising to 98 gigawatts at 0.5 percent and 126 at 1 percent (Norris et al. 2025). The estimate is headroom, not a connection right, and it still requires some peaking capacity and loads that can in practice pause; model training may tolerate that, while latency-sensitive inference may not. But if flexible connection becomes the norm, it relaxes the constraint more quickly than any build programme.

These points narrow the thesis; they do not overturn it. Even a heavily discounted queue leaves a documented shortfall in the largest US market and equipment backlogs that are real contracts. What changes is the diagnosis of the scarce resource: it is firmness and timing of demand as much as of supply, and the actors who can offer flexibility and verified commitments will connect first.

The evidence supports three plausible paths, which can coexist across regions. In coordinated convergence, regulators create a distinct pathway for large loads, equipment makers hit their output targets, and flexible connection becomes standard, so scarcity eases by the end of the decade. In islanded build-out, hyperscalers self-supply behind the meter and grid-connected capacity stays capped and short, leaving ordinary ratepayers with a tighter system. In demand reset, efficiency or weaker AI revenue deflates the queue, stranding some turbine slots and wires built for load that does not arrive.

The paths are distinguished by observable signposts, not by argument. The most informative is conversion: GE Vernova converted 10 gigawatts of slot reservations into firm orders in the second quarter of 2026, and the pace at which the remaining 63 gigawatts of reservations convert will show whether buyers are committing or hedging (GE Vernova 2026). The others are whether grid operators' tariff filings under the FERC orders are approved, whether the PJM reliability backstop procurement attracts new capacity, whether the next auctions clear below the cap with real new supply, and whether hyperscaler capital guidance continues to rise.

The comparative evidence points to a selective conclusion. Compute scales where large loads arrive with capital, generation, and flexibility attached, and where regulators have given them a distinct and verified pathway; it stalls where each layer is negotiated separately in a queue designed for a different customer. Merchant markets with scarce capacity and politically exposed ratepayers face the sharpest friction, while jurisdictions that can pair load with new supply and contract flexibility can move faster.

For regulators, the task is a cost-reflective pathway that verifies demand, allocates network costs to those who cause them, and funds transmission and equipment capacity. For grid operators and utilities, it is staging requests, screening duplicates, and offering flexible-connection tariffs. For developers and hyperscalers, power is the critical-path procurement item: sites should be chosen for firm headroom, generation should be brought or contracted early, and curtailment should be treated as a bargaining asset. For investors and suppliers, scarce equipment slots and interconnection rights are worth more than speculative megawatts in a queue.

AI's physical footprint will be set less by how many chips can be made than by how many megawatts can be delivered, firm, to a specific substation by a specific date. The grid is not a barrier to the technology but its pace-setter, and the actors who learn to coordinate or circumvent its layers will decide where the build-out lands.