Five years for a wire, six for a turbine

Every argument about artificial intelligence eventually arrives at a substation. The capital markets have spent this year learning to finance data centres — a Brookings paper published on Wednesday put the American build-out at $10.3tn by 2032 — and financing can be arranged in a fortnight. Electricity cannot. Underneath the term sheets sits a queue, a factory order book and a capacity auction, all three now measured in years rather than quarters.

Start with the queue, because it is the best-documented part. Berkeley Lab compiles interconnection data from more than fifty transmission operators covering about 98 per cent of American generating capacity. Its 2026 edition, published in June with data through the end of 2025, counts roughly 8,200 active projects seeking connection: 1,312 gigawatts of generation and about 749 gigawatts of storage. That total fell 10 per cent on the year — not because projects were built, but because developers withdrew.

The composition is the story. Active natural gas capacity rose 86 per cent to 253 gigawatts. Solar fell 19 per cent to 773, wind fell 19 per cent to 220, storage fell 16 per cent to 749. The queue is turning into a gas queue, because what a data centre buys is not energy but firmness — a guaranteed number of megawatts at three in the morning in February — which intermittent generation cannot sell without a great deal of storage attached.

The waiting time has not improved. Berkeley Lab puts the median duration from interconnection request to commercial operation at more than five years for projects completed in 2025, roughly double the figure for plants built between 2000 and 2007. The attrition is more striking still: of all the capacity that entered the queue between 2000 and 2020, only 13 per cent had reached commercial operation by the end of 2025. Seventy-five per cent was withdrawn. Ten per cent is still waiting. Another 549 gigawatts holds a draft or executed interconnection agreement and has not been built — which is to say the paperwork is not the binding constraint either.

Suppose a developer clears the queue. It then needs a machine, and three companies build heavy-duty gas turbines at scale. One files quarterly with the Securities and Exchange Commission: GE Vernova's second-quarter results show gas equipment backlog rising from 44 to 53 gigawatts and slot reservation agreements — payments made to hold a place in the production schedule, before an order exists — rising from 56 to 63. That is 116 gigawatts of contracted and reserved capacity.

Set that against output. GE Vernova says it is on track for 20 gigawatts of annual gas turbine production in the third quarter of 2026, 24 gigawatts in 2028 and 30 gigawatts in 2030. At today's rate the existing book is about six years of factory time; even on the 2030 target it is four. Orders rose 88 per cent organically; data-centre orders alone passed $5bn for the year to date, more than double the whole of 2025. Siemens Energy tells the same story from Munich: record quarterly orders of EUR17.9bn, a backlog of EUR162bn and a book-to-bill ratio of 1.57 — which is to say the queue lengthened again during the quarter.

Nuclear is the announced answer, and it is worth being precise about what has been committed. Carnegie's John Pendleton and Mackenzie Schuessler counted the hyperscaler deals in June and found about 13 gigawatts across power purchase agreements and direct partnerships. Roughly half is restarts, uprates and life extensions of existing reactors — Three Mile Island, Clinton, Susquehanna — which add electrons without adding a plant. The other half is small modular reactors whose demonstration units are scheduled for 2030 at the earliest and whose full portfolios land between 2034 and 2039. Their estimate is that the entire programme, if it all arrives, would produce about 102 terawatt-hours a year: less than a fifth of projected demand growth through 2035.

Faced with a five-year wait, the rational response is to leave the queue. That is what behind-the-meter generation is: a gas plant, a fuel cell stack or a nuclear offtake wired directly to the campus. The best-known cases — the retired Homer City coal site in Pennsylvania being rebuilt as a multi-gigawatt gas plant for a planned campus, and Amazon's offtake from Talen's Susquehanna reactors — are widely reported, though the aggregate pipeline figures circulating in industry research could not be verified at source by this desk. The laboratories have started shopping for smaller sites precisely because power, not land, is the scarce input.

What does the shared grid have to show for the money? PJM, which runs the largest wholesale market in America, held its capacity auction for the 2028/2029 delivery year in July. The result was a clearing price of $325 per megawatt-day — the FERC-approved cap, for the entire footprint — and a total cost of $16.4bn. It procured 138,318 megawatts and finished 6,831 megawatts short of its own reliability requirement, the second consecutive year of a shortfall.

One number in that release deserves to be read twice. Of everything the auction cleared, new generation and generation uprates accounted for 525 megawatts. A $16.4bn capacity auction, held at the price cap, in the most contested power market in the country, called forth half a gigawatt of new supply. Capacity prices are supposed to be an investment signal. The signal is being sent at maximum volume and the response is a rounding error, because the constraint is not the price. It is the turbine, the transformer and the queue.

Which leaves the question of who pays for the wires, and two answers are being worked out at once. In June the Federal Energy Regulatory Commission issued show-cause orders to each of the six regional grid operators, directing them to justify their existing tariffs or file changes creating expedited study processes for large loads — generally demand above 20 megawatts — and for generation electrically proximate to them. At state level the mechanism is the large-load tariff, and it is spreading fast: the Sabin Center's June survey counted 23 states with at least one approved and seven more pending as of May.

The terms are the interesting part. AEP Ohio's tariff covers customers above 25 megawatts, requires twelve-year minimum contracts and obliges them to pay for 85 per cent of contracted capacity whether or not they use it. Dominion's GS-5 tariff in Virginia, effective from January 2027, runs fourteen years, charges 85 per cent of contracted transmission demand and 60 per cent of generation demand, and demands $1.5m per megawatt of collateral. The effect in Ohio, the Sabin Center notes, was that AEP's large-load forecast halved. Some of the demand in the queue was never demand. It was optionality, held for free, and it disappeared the moment it acquired a price.

Set against all this, the aggregate numbers look sedate. The Energy Information Administration's September outlook forecasts American electricity sales of 4,135 billion kilowatt-hours in 2026, up about 2 per cent, and record generation of 4,368 billion. Commercial sector sales, the category containing data centres, grow 3.3 per cent and account for 63 per cent of the total increase. The EIA also notes, drily, a pause in new data centre projects in Texas.

Two per cent national growth and a five-year queue are not contradictory; they are the same fact seen from different ends. Demand grows slowly in aggregate because the increment is concentrated in a handful of nodes, and at those nodes it is enormous, lumpy and arrives faster than steel can be poured. The industry has spent three years treating compute as the scarce resource and capital as the enabling one. Both are now abundant relative to the thing that is not: a signed date on which the power will actually flow.