Energy and AI

What the consumption data actually supports

The most defensible US series is Berkeley Lab’s, which is built bottom-up from server shipments, utilisation and facility overhead rather than from what developers have announced. It put data-center consumption at 176 TWh in 2023, 4.4% of US electricity, and revised 2024 to 192 TWh, or 4.7%. That share implies total US consumption a little above 4,000 TWh, which is the right denominator to keep in mind throughout.

The central case in the 2025 update has data centers at 11.8% of US electricity by 2030, within a scenario range of 9.5% to 15.3%. Read the range rather than the midpoint. The distance between the low and high 2030 scenarios is larger than everything US data centers consumed in 2023, and that is a bottom-up study with access to shipment data. Forecasts assembled from interconnection requests are wider still and biased upward for a structural reason.

A developer securing a site does not file one interconnection request. It files several, across utilities and often across states, and abandons all but one. Each receiving utility books the full load into its forecast. Nothing in the process nets out the duplicates, and until recently nothing cost the developer anything for creating them. Texas legislated against this in 2025 with Senate Bill 6, which requires site control, a study fee and agreed curtailment terms before a large load enters the ERCOT queue. Speculative filings reportedly fell once it took effect. That a legislature had to intervene is better evidence of the scale of the double-counting than any particular gigawatt figure that gets quoted for it.

So there are two demand numbers in circulation and they are not measuring the same thing. One is an estimate of consumption. The other is a sum of requests. Utility load forecasts, and the capital plans built on them, largely use the second.

Where the chain binds

Between an announcement and a megawatt actually flowing there is a sequence, and it is worth being specific about it because the constraint sits in different places at different points. Site control comes first, then a load study by the utility, then an interconnection agreement, then the physical work: a substation, transformers, breakers and switchgear, and whatever transmission reinforcement the study identified. Only after all of that does system-level generation adequacy become the live question.

None of those steps is gated by money. Trade reporting through 2026 puts interconnection waits in Northern Virginia, Phoenix and Dallas at four to seven years. Substation transformer lead times, roughly a year before 2020, are reported above 160 weeks. Heavy-duty gas turbine slots are reported booked into 2028 and beyond, with US orders in 2024 above 14 GW, the highest since 2001. These figures come from trade sources rather than filings and should be treated as indicative, but they are consistent with what the equipment makers say about their own order books and with the direction of their pricing.

The supply side of that equipment has not responded the way a textbook says it should, and the reason is instructive. Turbine and heavy electrical equipment manufacturers spent roughly 2015 to 2021 in severe overcapacity. Gas turbine demand collapsed, plants ran far below nameplate, and the businesses were restructured or written down. That experience is why capacity is being added cautiously now, why slot reservation agreements and prepayments have become normal terms, and why pricing has held rather than being competed away in the first year of a demand surge. Suppliers are pricing for the possibility that this cycle ends the way the last one did.

The clearest market-priced evidence that the constraint is real comes from PJM, which covers Northern Virginia and much of the eastern data-center build.

An eleven-fold move in one year, then two years pinned at the cap. The 2027/2028 result is the one that matters most: clearing at the ceiling while still failing to procure the required capacity is a market saying it cannot solve the problem at any price it is allowed to charge. The total cost of that auction was $16.4 billion, against roughly $2.2 billion for the 2024/2025 delivery year.

What can serve a gigawatt, continuously

Training and inference clusters run near flat around the clock. That single characteristic eliminates most of what has been built in the US over the past fifteen years. A load with a 90%-plus factor does not care about levelised cost of energy; it cares about firm capacity available at three in the morning in January.

The nuclear contracts signed since 2024 are widely read as new supply arriving. Most of them are not, and the distinction is the single most under-appreciated point in this whole subject. Constellation’s agreement with Microsoft restarts the retired unit at Three Mile Island, now the Crane Clean Energy Center, at about 835 MW under a twenty-year contract. That genuinely adds electrons to the system. Constellation’s subsequent twenty-year agreement with Meta for the Clinton plant in Illinois, roughly 1,121 MW from June 2027, adds none. Clinton is already running and already serving the grid. The contract moves the output from the merchant market to a single buyer and fixes a price for two decades.

For the owner, that is excellent. For system adequacy it is neutral at best, because every gigawatt of existing firm capacity that gets contracted to a hyperscaler is a gigawatt the residual market has to replace. The nuclear PPA wave therefore tightens the market it appears to be relieving. That mechanism is also why it shows up in capacity prices rather than being absorbed quietly.

The bypass route has been tested and partially closed. Talen and Amazon originally structured the Susquehanna arrangement as behind-the-meter, with the data center drawing directly from the plant and largely outside the transmission tariff. FERC rejected the amended interconnection agreement in November 2024. The parties restructured it in mid-2025 as a front-of-meter PPA running to 2042. The practical reading is that a gigawatt-scale load can contract firm output but cannot easily exempt itself from paying for the grid it remains connected to. Onsite generation still works as a bridge, which is what xAI did in Memphis with mobile turbines, but bridging is a different thing from powering, and it attracts air-permitting scrutiny in proportion to its size.

Whether the price of power actually matters

It is worth testing the assumption that expensive electricity threatens AI economics, because a great deal of commentary assumes it without doing the arithmetic. The arithmetic is not favourable to the assumption.

What breaks the model is not being energised. A developer holding several billion dollars of depreciating accelerators will pay a very large premium to be running two years earlier and will barely notice a thirty percent higher tariff. That asymmetry is the most useful thing to hold on to, because it tells you where the pricing power sits. It sits with whoever controls the date, not with whoever sells the kilowatt-hour.

It also explains behaviour that otherwise looks irrational: twenty-year contracts signed at prices well above prevailing merchant power, prepayments for turbine slots years ahead of need, and willingness to site in second-tier markets with worse latency and thinner labour. Those are all purchases of schedule certainty.

Where the economics accrue

The exposures usually get grouped as though they were one trade. They are not. The useful division is between businesses that earn more because the asset base became scarce and businesses that earn more because they were permitted to spend more.

The strongest position in that table is the first. An independent power producer with an existing nuclear or gas fleet inside PJM has watched the capacity value of assets it already owns go from $28.92 to $333.44 per MW-day without deploying incremental capital. That is a genuine improvement in return on invested capital rather than an increase in invested capital, and it is the only place in the chain where that is structurally true.

The caveat is that the market has understood this. Constellation, the clearest expression of the position, has fallen about 21% over the past twelve months on reporting available in mid-2026, even as the operational thesis kept confirming. That is what it looks like when a narrative is fully priced and the argument has moved to duration: how many years of scarcity rent, at what price, before new supply or a demand pause closes it. Investors buying this exposure now are underwriting duration, not direction, and should be honest with themselves about which of the two they have an edge on.

Regulated utilities are the position most often misunderstood. Load growth at a regulated utility converts into rate base, and rate base earns a set return. More data centers mean a bigger balance sheet earning roughly the same percentage. That is a respectable earnings-growth story and it is not a pricing-power story, yet sector multiples near 18 times forward earnings appear to embed some of the latter. There is a second-order effect worth watching: large-load tariffs with minimum-take and exit provisions are becoming standard as regulators decide that data centers rather than residential ratepayers should carry interconnection cost. Those tariffs make the rate base more contracted and lower-risk, which is genuinely valuable. They do not raise the allowed return.

Equipment is where the supply constraint is hardest and the pricing most durable, for the reason given earlier: the suppliers were badly burned in the last cycle and are adding capacity at a pace that protects price. It is also the exposure with the cleanest evidence, since order books and backlog are disclosed quarterly and slot reservations are contractual rather than aspirational. The corresponding risk is symmetric. Backlog that is disclosed can be seen to stop growing.

What would weaken this

The thesis has several genuine ways to fail, and a few of the popular objections are weaker than they sound.

Efficiency is the objection most often raised and the least persuasive on its own. Per-token energy has fallen substantially with each accelerator generation, but rack power density has risen at the same time, so facility-level draw increases even as the work done per joule improves. Facility overhead offers less relief than is generally assumed: hyperscaler PUE is already near 1.1 against an industry average nearer 1.55, so the leading edge has already harvested most of what is available there. The remaining efficiency headroom sits in the older enterprise and colocation fleet, which is not what is being built. Cheaper inference has historically expanded usage rather than reduced total consumption, and there is no evidence yet that this time is different.

The serious risks are elsewhere. A capex pause is the obvious one, and the market is already probing for it: hyperscaler shares sold off in July 2026 after Alphabet’s results were poorly received, with capital intensity the specific concern. The honest tells to watch are depreciation-life extensions, which flatter earnings while spending continues, and any cancellation of signed leases rather than the deferral of unsigned ones. Reports of mass cancellations in 2025 and 2026 have generally not survived examination, and the difference between walking away from an option and terminating a contract is the difference between a wobble and a turn.

Second, the demand forecasts could deflate on their own as queue filters spread. Texas has already shown that a modest financial hurdle removes a large share of requests. If PJM, MISO and the southeastern utilities adopt comparable screens, the published load forecasts that underpin utility capital plans will fall without any change in real demand, and the market will have to work out which of the two moved.

Third, equipment scarcity is self-correcting on a five to seven year horizon. Turbine and transformer capacity is being added. The pricing power in that part of the chain has a defined life, and the announcements that end it will come from the suppliers themselves.

The single cleanest indicator that the constraint is easing would be a PJM capacity auction that clears below the cap after the cap is raised. Until that happens, the market is telling you it cannot buy what it needs.

Assessment

The evidence supports treating the AI buildout as a physical infrastructure and power-capacity cycle in addition to a semiconductor cycle, with one qualification that changes where the money is. It is a capacity and schedule cycle, not an energy price cycle. Electricity is roughly a tenth of the cost of running these facilities and is not what constrains them. Time to energisation is.

What looks structurally durable is the value of firm capacity that is already connected to the grid, and the pricing power of the few manufacturers whose output cannot be scaled on the timeline the buildout requires. What looks fully priced is the independent power producers, where the re-rating has happened and the remaining question is contract duration. What looks mispriced in the other direction is any demand forecast assembled from interconnection requests, and any regulated utility multiple that treats load growth as though it improved returns rather than enlarged the balance sheet.

Three things would settle most of the remaining argument, and all three are observable. Whether capacity auctions clear below their caps once those caps rise. Whether the turbine and transformer manufacturers announce capacity expansions large enough to close their backlogs. And the ratio of megawatts actually energised to megawatts announced, which is the only number that distinguishes a construction cycle from a press cycle. That last figure is not published anywhere in a usable form, which is itself worth noticing.