This is the seventh part in a series on how AI is crossing out of software and into the physical world. This time, Jan Bosch reaches the constraint underneath all of it.
There’s a pleasing symmetry in the way the electricity demand of artificial intelligence usually gets discussed. While AI is an enormous new load on the electricity system, it’s also a powerful tool for running that system better. The two roughly cancel, and we can all move on. I’ve come to think the symmetry is false, and that the reason for that is the most useful thing in this part of the series. The two halves run on completely different clocks.
Start with the load, which is real but frequently misdescribed. The IEA’s “Energy and AI” report puts global data center electricity consumption at about 415 TWh in 2024, roughly 1.5 percent of world consumption, and projects it to more than double to around 945 TWh by 2030 – more than Japan uses today. That’s a large number in absolute terms and a modest one in proportional terms: about a tenth of global electricity demand growth to 2030, though more than 20 percent of demand growth in advanced economies. So, this isn’t a civilizational energy crisis; it’s something more specific and, for anyone trying to build, considerably more awkward. Because the binding constraint isn’t generation; it’s interconnection.
Lawrence Berkeley’s “Queued up” report counted roughly 1,312 GW of generation and 749 GW of storage sitting in US interconnection queues at the end of 2025. There’s no shortage of people who want to build power plants. Unfortunately, the median time from interconnection request to commercial operation for projects completed in 2025 was over five years. And of all the interconnection requests submitted between 2000 and 2020, only 13 percent had reached commercial operation by the end of 2025. Seventy-five percent were withdrawn. Hold on to that withdrawal rate. We’ll need it twice.
If you want to see where this ends up, the Netherlands got there first. In April, a Dutch court ruled against the data center developer Goodman and in favor of the grid operator Tennet, confirming that the electricity grid around the Vijfhuizen substation in Haarlemmermeer – next to Schiphol, in the most connected corner of a rich, well-run country – is genuinely full. Goodman wanted a 70 MW connection and asked the judge to impose a 500,000-euro daily penalty until it got one. The court declined. Capacity in that region may not free up until something like 2035.
Read that again, because it isn’t a price signal. A developer with money, permits and urgency was told to wait a decade, by a judge, because the wires aren’t there. This is what a hard physical constraint looks like when it finally arrives in a market that assumed capacity was purchasable.
Which brings us to the clocks. An AI capacity plan runs on an eighteen-month cycle. Chips are ordered, a site is chosen, a building goes up and the whole thing is obsolete in three years. Grid infrastructure runs on ten to fifteen years: planning, consultation, permitting, land, transformers with multi-year lead times, construction. These two cycles can’t be reconciled by trying harder; they’re different kinds of activity. So, the AI industry stopped waiting and did something structurally new: It became an energy developer.
A recent survey of behind-the-meter projects identified 59 of them with roughly 90 GW of announced capacity – about a quarter of all planned US data center capacity – and 92 percent of that was announced since January 2025. The generation mix tells you how much of a hurry everyone is in: aeroderivative turbines originally designed for aircraft and naval use, refurbished industrial turbines, reciprocating engines and mobile gas generators mounted on semitrailers. This isn’t an energy transition; it’s a field hospital.
And now use that withdrawal rate again, because it applies here, too. Of the 90 GW announced, only around 2 GW is actually operating. The announcement is cheap, the turbine isn’t, and a great many of these projects will quietly not happen – exactly as three quarters of interconnection requests have quietly not happened for two decades. Anyone building a business on the assumption that this capacity arrives should discount it heavily.
Now the other half, and the reason why I think the symmetry is false in an interesting way rather than a disappointing one.
AI applied to the grid isn’t poetic justice; it matters because software is the only component of the energy system that operates on the same time constant as the demand growth. You can’t build a transmission line in eighteen months. You can, in eighteen months, forecast wind output better, predict which transformer is about to fail, schedule storage more intelligently and recover curtailed renewable generation that’s currently thrown away. The IEA estimates that widespread adoption of existing AI applications could deliver emission reductions equivalent to roughly 5 percent of energy-related emissions by 2035.
Every one of those is a way of getting more out of assets that already exist. That isn’t a consolation prize. On a decade-long build cycle, it’s the only lever that moves.
To a startup, this provides an unusually clean selection rule: Make the thing that defers or avoids a physical build. Forecasting, flexibility and demand response, curtailment recovery, predictive maintenance on transformers, siting intelligence that tells a developer which substation has actual rather than theoretical headroom – these all share the property that the customer can measure the savings against a number they already report, and the payback arrives inside a budget cycle rather than a planning cycle. The interconnection queue itself is a product: In a world where 75 percent of requests are withdrawn, knowing which queue positions are real is valuable information that nobody currently sells well.
The trap is the opposite: anything whose value depends on new wires being built on schedule. That’s a bet on the slowest clock in the system.
The instinct is to say that utilities must build faster. But the deeper difficulty is that regulated planning assumes forecastable load, and AI load isn’t forecastable in the way the planning process needs. It’s enormous, lumpy, concentrated in a few locations – and speculative. Developers hold positions in multiple queues simultaneously, which means the queue isn’t a demand forecast at all; it’s a portfolio of options, most of which will be abandoned.
That puts a utility in a genuinely unpleasant position. Build for the load and it may not arrive, leaving ratepayers holding assets they didn’t ask for. Decline to build and you’re the reason the region lost the investment. There’s no version of this where the utility isn’t blamed, and the current regulatory frameworks give it very little help in distinguishing a real request from a free option. Flexible connection agreements and non-firm contracts are the most interesting response I’ve seen, precisely because they let a utility say yes without betting the balance sheet on a forecast it doesn’t believe.
For society, the framing that matters is one the debate has almost entirely missed. We argue about whether AI uses too much energy. The more pressing fact is that grid capacity has quietly become a rival good.
When a data center takes the last available capacity at a substation, it doesn’t take it from an abstraction; it takes it from the electrolyzer someone wanted to build, the steel plant’s new electric furnace, the port’s shore power project, the heat pumps in a housing development, the charging depot for a bus fleet. Every one of those was also going to electrify something that currently burns fuel. In the Netherlands, this isn’t a thought experiment; it’s a waiting list, and the people on it include exactly the industrial decarbonization projects that climate policy is meant to be accelerating.
And the allocation rule we’re using, almost everywhere, is the order in which applications were filed. Not value created, not emissions displaced, not jobs, not whether the applicant could have been flexible. First come, first served – a rule we’d find indefensible if we stated it out loud about any other scarce public resource.
Which is where I want to end, because there’s a genuinely useful piece of arithmetic that almost nobody in this argument seems to know.
Researchers at Duke University’s Nicholas Institute asked what would happen if new large loads were willing to be curtailed occasionally rather than demanding firm power every hour of the year. Their finding, in “Rethinking load growth,” is that the largest 22 US balancing authorities could absorb somewhere between 76 GW and 126 GW of new load on the existing system – if that load is willing to be curtailed. The price is smaller than you’d guess. At a curtailment rate of 0.25 percent of annual load, which works out at about 85 hours a year, you get the lower end of that range; at 1 percent, around 366 hours, you get the upper end. Eighty-five hours. Under four days a year, concentrated in the worst peaks, and seventy-six gigawatts of capacity appears in infrastructure that’s already built and already paid for.
Amory Lovins spent a career making the point that we’ve been asking the wrong question. Customers, he wrote in 1990, don’t want kilowatt-hours; they want “hot showers, cold beer, lit rooms and spinning shafts, which can come more cheaply if electricity is used more efficiently.” Likewise, a data center doesn’t want 70 megawatts; it wants inference delivered. Those aren’t the same requests, and the difference between them is worth as much as 126 gigawatts of headroom in a grid we’ve already built.
So, the next time you hear that the grid can’t take it, the honest follow-up question is: Can’t take what? Firm load, or your load? Because the grid isn’t full. It’s full of our assumption that every new connection must be uninterruptible, and that assumption is a choice – made, so far, almost entirely by the people who benefit from it.
