Shells rise in 18 to 24 months; grid power averages four years. The biggest AI builders have already contracted the firm megawatts everyone else is still queuing for.
Nvidia is in talks to take a minority stake in Lancium, the Texas power developer behind OpenAI and Oracle's Stargate campus in Abilene, The Information reported this week. Anthropic has held early-stage conversations with the firm as well, according to follow-on reporting. Nothing is signed, and the talks may go nowhere. The notable part is who is talking to whom: the company that sells the accelerators is weighing equity in the company that delivers the electrons.
Equity in a power developer may confer what a power purchase agreement typically does not: strategic influence over how projects get developed and which customers they prioritize. A minority stake would carry no dispatch rights, but it ties the investor to the developer's whole pipeline rather than to a single contracted block of power. That alignment is worth pursuing because the contracts that define today's market were signed one and two years ago, and there is little sign the holders are willing to give those positions up.
The scarce asset in AI infrastructure is scheduled, contracted, deliverable power: a megawatt tied to a site, a signature, and a date. GPUs can be bought and models licensed; power on those terms is secured only years in advance.
Amazon holds 1,920 MW of nuclear output from Talen Energy's Susquehanna plant under a power purchase agreement running through 2042, restructured as a front-of-the-meter deal in June 2025 after federal regulators rejected the original behind-the-meter arrangement. Microsoft signed a 20-year agreement with Constellation in September 2024 that funds the restart of the 835 MW Three Mile Island Unit 1, now the Crane Clean Energy Center, targeted for 2028. Meta followed in June 2025 with a 20-year deal for the 1,121 MW output of Constellation's Clinton plant in Illinois, starting June 2027. Google's arrangement with Kairos Power runs through the Tennessee Valley Authority: a 50 MW molten-salt reactor at Oak Ridge scheduled to begin operating in 2030, the first tranche of 500 MW planned by 2035.
Together, Amazon, Microsoft, and Meta have roughly 3.9 GW of firm nuclear output committed on terms of 17 to 20 years. Those commitments sit at different distances from an operating electron, and the distinctions matter. Clinton runs today; Meta's contract keeps it running. Susquehanna is delivering now and ramps to full volume by 2032. Crane is a restart that has yet to produce a watt, and the Kairos reactor has not been built. Announced, contracted, connected, and operating capacity are four different things, and buyers have grown increasingly careful to distinguish among them.
A data-center shell goes up in 18 to 24 months. Getting utility power to it averages about four years nationally, according to JLL's 2026 Global Data Center Outlook, which now ranks "speed to power" as the top site-selection criterion. In Northern Virginia, JLL puts a 100 MW connection at roughly seven years. Candidate sites in PJM territory face average queue waits of 3.4 years.
Equipment stretches the schedule further. Large power transformers average around 128 weeks to deliver, and generator step-up units around 144; Hitachi Energy has quoted waits beyond 30 months. GE Vernova's electrification backlog grew from $25 billion to $42.4 billion in a year, and its CEO said in March 2025 he expected the company to be "largely sold out through the end of '28" across gas turbines, transformers, and switchgear.
SCN documented the collision in April, when Nvidia's projected order volume hit the grid's capacity wall, and again in the analysis of AI training power demand outrunning grid build times. The megawatt now behaves the way the HBM market did earlier this year: the binding constraint is allocation, not supply. New capacity takes years to add, so what exists gets rationed among the buyers who committed earliest.
The announced pipeline dwarfs what can be built on schedule. Bernstein Research estimates that 35 to 40 percent of announced global data-center capacity is at risk of delay or cancellation through 2027, per ANI. Goldman Sachs has put the figure more bluntly: only about half the capacity scheduled two years out materializes on time. SemiAnalysis disputes part of that math, arguing that much of the "canceled" pipeline consists of speculative interconnection filings that never represented real projects. Either reading lands in the same place. Whether the announced pipeline is delayed or was never real, buyers end up competing for the same finite pool of projects that can be energized on schedule, and the market increasingly values a megawatt with a credible delivery date above any volume of announced capacity.
Developers are responding by building the power first. Cleanview's February report identified 46 US data centers, roughly 56 GW combined, planning behind-the-meter generation, about 30 percent of the country's planned capacity and up from effectively zero a year earlier. Nearly three quarters of that on-site equipment burns natural gas, a useful corrective to the impression that the moat is a nuclear story. Nuclear is one source of firm power, with the longest schedule of any of them. Gas arrives sooner, which is why Meta's Alberta campus is pairing with a planned 932 MW gas-fired plant in a province whose grid cannot host several hyperscale projects at once. Even compute contracts have adopted the denomination: OpenAI's agreement with Cerebras is written as 750 megawatts of Cerebras wafer-scale systems, a deal Cerebras now values above $20 billion, and Anthropic's 3.5 GW commitment for Google TPU capacity followed the same logic of buying capacity years before it is needed.
The scramble for position extends well past anything that will generate power this decade. On July 7, GridMarket and Deployable Energy announced a target pipeline of more than 3 GW of Unity microreactors through 2035. No customers, sites, or power prices were named, and the $22.5 billion project value attached to the pipeline is a company estimate. Deployable's Unity reactor did clear a real milestone on June 30, a zero-power fueled criticality experiment at Idaho National Laboratory. Criticality means a sustained chain reaction under DOE authorization. It does not mean commercial electricity; no power-conversion system was demonstrated.
The federal government is financing the same reflex at far larger scale. In June, the Department of Energy issued a conditional commitment of up to $17.5 billion in loans to buy long-lead components for up to 10 AP1000 reactors across five projects, including Fermi America's Texas campus. The covered items are reactor pressure vessels, steam generators, and structural modules; the loans finance components, not construction. Reserving forgings before projects clear licensing follows the same logic as the hyperscaler contracts: pay early for a position in a queue that money alone cannot shorten later.
Many of those forgings come from Korea. Doosan Enerbility has begun long-lead forging work for Fermi America's proposed AP1000 units and is building a dedicated SMR production line at its Changwon plant, and TerraPower has named HD Hyundai Heavy Industries a preferred manufacturer for its Natrium program. A US operator can control the site and the reactor design while a Korean press controls the production slot for the largest components. The power moat has upstream dependencies of its own. A private grid is not sovereign if its forgings, fuel, or transformers remain scarce.
The costs of this competition are landing on shared infrastructure. PJM's capacity auction for the 2027/28 delivery year cleared at the FERC-approved price cap of $333.44 per megawatt-day in December, the third consecutive record, for total capacity costs of about $16.4 billion. PJM attributes nearly 5,100 MW of the 5,250 MW growth in its forecast peak load to data centers. Those capacity costs flow into retail electricity bills across 13 states and the District of Columbia.
Regulators are formalizing the boundary between large loads and everyone else. AEP Ohio's data-center tariff, approved by state regulators effective July 2025, requires customers above 25 MW to commit to 85 percent take-or-pay contracts for a minimum of 12 years. At least 36 utilities now maintain large-load tariffs, according to the Edison Electric Institute. Companies that signed capacity before these regimes took effect sit ahead of them, an advantage the newer tariffs do not reach.
Securing a megawatt is half the job. The other half is extracting more compute from it.
Meta's 83,000-GPU AI supercomputer remains the cleanest published example. The company provisioned its GB200 accelerators at 80 percent of thermal design power because aggregate throughput per megawatt mattered more than peak performance per chip, then used better power measurement to raise the limit and recover about 2 percent of throughput. As SCN reported in May, the cluster, switchboards, cooling plant, and workload scheduler operate as one machine. Two companies connected to the same number of megawatts can produce different quantities of useful tokens from them, which means the moat extends past generation into load shaping, scheduling, and the willingness to trade chip-level speed for facility-level output.
Electricity is a commodity. Scheduled, contracted, deliverable power is not, and the numbers behind that difference are now legible: a four-year national average to connect, seven years for 100 MW in Northern Virginia, 128-week transformer queues, equipment makers expecting to be largely sold out through 2028, and reactor forgings reserved years before licensing.
None of this makes the moat permanent. Transmission does get built, transformer makers are adding production lines, and the turbines and reactors ordered in this cycle will eventually connect; each addition narrows the schedule gap. The open question is whether those improvements arrive before the next wave of AI demand widens the gap again. Nor is buying expensive power later a clean escape for a laggard: merchant electricity can be had at a price, but interconnection timing and transmission rights gate delivery regardless of willingness to pay. An operator that missed the contracting window waits in the same queue as everyone else.
For now, the operators ahead in this market hold power with contract signatures, interconnection positions, and delivery dates: Amazon's 2042 horizon at Susquehanna, Meta's June 2027 start at Clinton, Microsoft's 2028 restart at Crane. An announcement, however large, buys none of that. If the Lancium talks turn into a deal, the investment would reflect the same conclusion the contracts already point to: the scarce asset is scheduled, contracted, deliverable power, and the surest remaining way to acquire it is upstream.
The SCN Staff is a small AI editorial squad working under human direction. Each agent owns one job.
Scout does the research. It runs down primary sources and checks what's already been published, on SCN and everywhere else, before a story gets written. If a claim can't be traced back to a real document, Scout flags it.
Forge writes. It takes what Scout found and turns it into a draft, argument and sentences and all. Every SCN piece starts here, then gets sharpened.
Cipher handles search: the titles, descriptions, and keyphrase work that decides whether a good article ever gets found. Least glamorous job on the squad. Also one that matters more than it looks.
Pixel makes the visuals. Images, charts, the occasional diagram, all built to SCN's brand instead of pulled from a stock library. When something's easier to see than to read, it goes to Pixel.
Editorial judgment and the final call stay with the humans. So does the fact-checking.
