Can Our Energy Infrastructure Keep Up?

Every January, the organization responsible for the reliability of the North American power grid publishes a document nobody outside the utility industry reads for pleasure: a ten-year demand forecast. It is written, deliberately, to be dull. This January's wasn't. NERC's Long-Term Reliability Assessment projected summer peak demand to grow by 224 gigawatts over the coming decade — a number roughly 69 percent larger than the one it had published just a year earlier, and 24 percent above today's actual peak. The stated cause, in the assessment's own words, is new data centers for artificial intelligence. Thirteen of the twenty-three regions NERC assesses now face resource-adequacy challenges within the decade [1][2].

Five investigations in this Convergence have followed the Age of AI through claims, workforces, hiring, and concrete. This one lands on the input none of that buildout can simply buy its way past. Chips can be flown in overnight; a data center goes up in about two years. Power plants, turbines, and transmission lines move on five-to-fifteen-year clocks — and after two essentially flat decades of American electricity demand, they're being asked to sprint [4]. The question underneath everything else in this Convergence comes down to one plain sentence: does the power actually arrive in time?

The American grid spent roughly a generation learning how not to grow. From the late 2000s on, national electricity demand stayed essentially flat as efficiency gains absorbed most economic growth, and an entire industry that had been organized around expansion quietly reorganized itself around maintenance instead [4]. Utilities trimmed planning staff. Regulators tuned rate cases for a world of stasis. The interconnection process, which was never built for speed in the first place, calcified further. None of that was mismanagement. It was a rational response to twenty years of consistent evidence.

The Age of AI is now inverting that environment faster than institutions built for the old one can turn. History offers a specific, useful caution here, one our reactors investigation documents in full: the last time the computing industry was widely predicted to devour the grid, in the early 2000s, the forecasts missed by roughly an order of magnitude, undone by efficiency gains nobody had priced in at the time [7]. The grid today faces a genuinely two-sided risk on badly mismatched clocks — build for demand that never shows up, and ratepayers carry stranded plants for decades; fail to build for demand that does show up, and reliability fails in precisely the regions where the growth concentrates. Everything in the section below is the system trying to price that asymmetry in real time.

The national laboratory that tracks the grid's interconnection queue counts roughly 8,200 projects currently waiting to connect: about 1,300 gigawatts of proposed generation plus another 750 or so of storage, several multiples of what NERC says the grid actually needs [3]. But the median project that finally connected last year had waited more than five years from its original request to actual operation, and 549 gigawatts hold signed interconnection agreements while still not delivering a single watt of power [3]. This is not a shortage of ambition. It's a bottleneck of process and physical wire.

The composition of that queue is shifting underneath the numbers, too — proposed gas capacity surged 86 percent year over year to 253 gigawatts while proposed solar, wind, and storage all declined, the market's own answer to the question of what "firm power" means right now [3]. Except that market has its own bottleneck: the world's gas-turbine manufacturers are effectively sold out. One manufacturer alone reports an 80-gigawatt backlog stretching into 2029, with its reservations expected to be fully booked through 2030, and analysts estimate the industry running at close to 90 percent of manufacturing capacity globally [5]. Nuclear power, meanwhile, has become this era's showcase — America's first-ever restart of a retired reactor, its output contracted for twenty years to a single software buyer, is targeting 2027 [6] — but a restart is a one-time prize, while new reactors still live on a decade-long clock of their own. Every realistic path to firm power this decade — gas, nuclear, transmission-connected renewables — is limited by a queue, a factory, or a calendar. Demand moves on two-year data-center timelines. Right now, supply simply cannot.

Competing hypothesis NERC's own assessment contains an unusual admission: the same data centers driving its record forecast also make that forecast harder to trust, because a single campus can request interconnection in several utility territories simultaneously, and speculative projects sit in the same queue as real ones [8]. Nobody outside the utilities themselves knows the actual deflation factor — how many of those 224 gigawatts are the same load counted twice, or capacity that will never actually be built. One camp reads the number as understated, given how consistently adoption has surprised upward so far in this Convergence. The other camp hears the early-2000s echo and expects a good portion of that phantom load to evaporate the moment it meets a binding contract [7].

Competing hypothesis There's a second, live dispute over who pays and what actually gets built in the meantime. Whether AI demand ends up financing a cleaner, firmer grid — the pattern set by the nuclear restart above — or a new fleet of gas plants plus higher household rates is genuinely unresolved, and it's regional: the same argument our reactors investigation traces through individual state rate cases [6]. A newer variable cuts across both camps entirely: NERC's own analysis finds that data centers willing to curtail their own power use during peak demand measurably improve reliability across an entire region [1]. Whether hyperscalers accept that kind of flexibility at scale may end up mattering more than any single power plant getting built. The debate has moved past whether the Age of AI strains the grid — the referee has already said it does [1] — and settled on whether that strain gets negotiated, built through, or partly imagined.

The interconnection bottleneck documented above is not one system's problem — it's four systems moving at different speeds. Utility regulation and interconnection review run on five-year-plus clocks [3]. Turbine manufacturing runs on a factory-capacity clock, sold out into 2029–2030 [5]. Corporate capital — the same $700 billion buildout traced in the previous investigation in this Convergence — moves in a single budget cycle [1][3]. And the data centers themselves are built in about two years. Every plausible resolution here is really a question of which of these four clocks bends to match the others.

This is the direct continuation of the previous investigation's finding that the grid is the one constraint no capital program can simply purchase its way past. The nuclear restart named throughout this piece — first covered in a standalone investigation on this site — sits exactly at the intersection of the capital and regulatory systems: a software buyer's capital commitment underwriting a regulator-approved plant restart [6].

Nobody disputes the grid faces real strain — NERC said so itself. What's genuinely open is how that strain gets resolved: through negotiation between utilities and their largest customers, through a straightforward if slow buildout, or because the forecast driving all of it turns out to be counting demand that was never really there. None of these are mutually exclusive, and none are predictions.

The infrastructure can keep up — but not on its current clocks. The generation exists on paper several times over [3]; what doesn't exist yet is a five-year interconnection process compatible with two-year data-center construction, spare turbine manufacturing capacity before 2030 [5], or a regulatory framework that can decide quickly who actually pays. Every plausible path through this decade turns out to be a story about institutions adjusting their speed, not a story about physics. Public, trackable indicators worth watching over the coming months and years:

• Regulatory filings — Whether the next NERC assessment grows the 224-gigawatt forecast again or begins to deflate as speculative loads wash out [1][8].
• Construction & manufacturing — Crane restart milestones against the 2027 target [6]; gas-turbine order books and any manufacturing-capacity expansions [5].
• Corporate disclosures — Whether major hyperscalers sign publicly disclosed flexibility or curtailability agreements — the first hard evidence for the negotiated-grid path.
• Interconnection & rate cases — Whether interconnection reform actually shortens the median queue time [3]. State rate-case outcomes on who pays for data-center-driven grid upgrades — the clearest political test available.
• Measured demand — Measured US demand against both the 224-gigawatt path [1] and the IEA's base case [4] — the number that finally grades every forecast made across this entire Convergence, in either direction.

Six investigations ago, this Convergence asked what kind of thing the Age of AI actually is. The evidence answered piece by piece: a population-scale adoption event; a workforce reshaping itself from the entry level up; a hiring system re-pricing what it trusts; new work already visible in real hiring decisions and real employment counts; a construction program large enough to show up in the national accounts; and now, a demand shock arriving at the one system nobody can simply airfreight in. What happens next to any of it is being decided right now in interconnection queues, rate cases, and turbine order books — argued about in public, where the evidence lives for anyone willing to look. This Convergence is finished. The weekly work of finding out what matters next starts again.

This Convergence is done, but the publication isn't. Of everything the Age of AI touches — work, hiring, capital, concrete, power — what should we look into next, and what evidence have you seen that we haven't found yet? Tell us. We read every message, and this is genuinely how the next investigation gets chosen.