The binding constraint on artificial intelligence is no longer chips, capital, or talent. It is electricity. US datacenters consumed roughly 176 terawatt-hours in 2023, about 4.4% of national electricity demand [VERIFY: LBNL 2024 US Data Center Energy Usage Report], and Lawrence Berkeley National Laboratory projects that figure reaching 325 to 580 TWh by 2028, as much as 12% of US consumption [VERIFY: LBNL 2024 report]. The grid supplying that load is adding capacity at a fraction of the required pace, with roughly 2,600 gigawatts of proposed generation and storage stalled in interconnection queues [VERIFY: LBNL Queued Up 2024]. That mismatch, demand compounding at datacenter speed against supply moving at utility speed, is the most investable imbalance in the market today.
The demand shock, in terawatt-hours
The demand side of the ledger is compounding. The International Energy Agency estimates global datacenter consumption at roughly 415 TWh in 2024, about 1.5% of world electricity, and its base case has that doubling to around 945 TWh by 2030 [VERIFY: IEA Energy and AI report, April 2025], with AI-accelerated servers accounting for the majority of the growth [VERIFY: IEA]. The physics behind the curve is rack density: a legacy CPU rack draws 8 to 17 kilowatts, while a modern GPU training rack draws 80 to 130 kilowatts or more [VERIFY: NVIDIA GB200 spec sheets; Uptime Institute]. The funding this build-out, Microsoft, Alphabet, Amazon, and Meta, guided to combined 2025 capital expenditure above $300 billion, most of it aimed at AI datacenters [VERIFY: company earnings calls].
This piece is the physical-world sequel to a thesis Ezana Echo has already mapped: the consolidation analysis "They're Not Buying the Model" argued that 2026's $100-billion-plus tech M&A wave is a land-grab for proprietary data pipelines and workflow positions rather than for models. Every one of those pipelines terminates in a building that draws megawatts. The same logic that made data the scarce input upstream makes firm power the scarce input downstream, and the scarcity premium is migrating from software multiples to generation assets.
Against that demand curve, supply is arithmetic. US utility-scale capacity additions ran about 49 GW in 2024 and roughly 63 GW planned for 2025 [VERIFY: EIA generator inventory], strong years by historical standards, yet most of it intermittent solar and storage rather than the around-the-clock firm power an requires. A gigawatt-class training campus needs its power at a 99.999% duty cycle; a solar farm delivers a 25% capacity factor. The gap between the bars and the line in the figure below is the whole story: demand measured in terawatt-hours consumed, supply measured in gigawatts that mostly cannot run at night.
FIG. 1 · DEMAND COMPOUNDS, SUPPLY CRAWLS
US DATACENTER POWER DEMAND (TWh, BARS, LEFT AXIS) VS. US UTILITY-SCALE CAPACITY ADDITIONS (GW, LINE, RIGHT AXIS) · 2024 ONWARD BLENDS ACTUALS, EIA PLANNED ADDITIONS, AND A MIDPOINT PATH OF THE LBNL 2028 SCENARIO RANGE.
Bar = datacenter demand (TWh, left axis); line = grid capacity additions (GW, right axis).
← swipe to explore →The interconnection queue is the real bottleneck
The US does not lack proposed generation; it lacks the ability to connect it. Lawrence Berkeley National Laboratory's "Queued Up" study counted roughly 2,600 GW of generation and storage in active interconnection queues at the end of 2023, about 95% of it solar, wind, and batteries [VERIFY: LBNL Queued Up 2024]. The median project now waits around five years from interconnection request to commercial operation, versus under two years for projects built in 2008 [VERIFY: LBNL Queued Up], and historically only 13 to 14% of queued projects ever reach commercial operation [VERIFY: LBNL completion-rate analysis]. Queue position, not on paper, is what determines when electrons actually flow, and the queue has become the single longest lead-time item in the entire AI supply chain.
The nuclear premium: PPAs, restarts, and co-location
Hyperscalers responded by going around the queue and straight to existing firm capacity, and nuclear is the asset they want. Microsoft signed a 20-year power purchase agreement with Constellation to restart the 835 MW Three Mile Island Unit 1 as the Crane Clean Energy Center, with the return-to-service target pulled forward to 2027 [VERIFY: Constellation announcements]. Amazon agreed to take up to 960 MW co-located at Talen's Susquehanna plant, in a structure reworked after FERC rejected the original interconnection amendment [VERIFY: Talen-AWS filings]. Meta signed a 20-year deal for the full 1,121 MW output of Constellation's Clinton plant from 2027 [VERIFY: Constellation-Meta announcement]. These contracts are running in reverse: instead of the utility locking in the customer, the customer is locking up the plant, for decades, at premium prices.
The energy industry cannot be the reason that America loses the AI race. Restarting a retired reactor for a single buyer would have been unthinkable five years ago; today it is the market clearing.
— Paraphrasing Joe Dominguez, Constellation CEO, at the Crane announcement, September 2024 [VERIFY: as-reported remarks]
The market signal confirming the scarcity is the price of capacity itself. PJM, the grid operator spanning 13 states from Virginia's datacenter corridor to Chicago, saw its 2025/26 capacity auction clear at $269.92 per megawatt-day, up from $28.92 the year before, a roughly ninefold jump [VERIFY: PJM Base Residual Auction results], and the following auction cleared at $329.17, pinned at the price cap [VERIFY: PJM 2026/27 BRA]. Joe Dominguez has built Constellation's entire equity story on this repricing: the largest US nuclear fleet, roughly 32 GW of capacity running at nuclear-fleet capacity factors near 93% [VERIFY: company filings; EIA fleet data], suddenly holds the scarcest commodity in the AI economy, always-on carbon-free power that already exists.
Three ways the build-out resolves
How the crunch resolves depends on two clocks: how fast reform accelerates the queue, and how fast hyperscalers route around it. FERC Order 2023 moved interconnection to first-ready-first-served cluster studies with real deposits [VERIFY: FERC Order 2023], and the ADVANCE Act plus a string of pro-nuclear executive actions have compressed NRC licensing timelines [VERIFY: ADVANCE Act, 2024], a genuine for both the queue and the reactor pipeline. The scenarios below trace the three paths the next four years can take, and what each one does to power prices and the trade.
FIG. 2 · THREE PATHS FOR THE POWER CRUNCH
BASE · ALT · BEAR: HOW GRID SUPPLY AND AI DEMAND RECONCILE THROUGH 2030, AND THE PRICE SIGNAL EACH PATH SENDS.
Three build-out scenarios; the kill switch is the indicator that flips the base case.
The uranium leg of the trade
If existing reactors are the scarce asset, their fuel is the scarce input behind it. Uranium spot prices rose from roughly $30 per pound in early 2021 to a peak near $106 in early 2024 before settling into a $65-80 band through 2025 and 2026 [VERIFY: UxC/Cameco price history], a repricing with the classic anatomy of a : a decade of underinvestment after Fukushima, inventories drawn down, then a structural demand shock arriving faster than mines can respond. Cameco, which restarted McArthur River in 2022 after a four-year suspension [VERIFY: Cameco], is the Western fuel cycle's anchor asset, and the US ban on Russian enriched uranium, signed in May 2024 with waivers running only through 2027 [VERIFY: Public Law 118-62], forces the re-shoring of a fuel chain in which Russia had supplied roughly a fifth to a quarter of US enrichment [VERIFY: EIA uranium reports].
The small modular reactor names are the venture tranche of the same thesis. NuScale holds the first NRC-certified US SMR design [VERIFY: NRC records], and Oklo, led by co-founder Jacob DeWitte, is pursuing a combined license for its Aurora fast reactor after an initial application was denied in 2022 [VERIFY: NRC docket]; Sam Altman chaired Oklo's board until stepping down in April 2025 as OpenAI began exploring energy deals directly [VERIFY: Oklo filings]. Centrus, the only US-licensed producer of the high-assay fuel most SMR designs need, delivered its first HALEU to the Department of Energy in late 2023 [VERIFY: Centrus/DOE]. None of these companies has meaningful revenue from operating reactors yet; they trade on licensing milestones and letters of intent, which makes them optionality, not cash flow.
The utility capex supercycle
For regulated utilities, the crunch is not a crisis but an order book. Investor-owned utility capital spending tracked by the Edison Electric Institute ran near $185-200 billion in 2025, with five-year plans totaling well over a trillion dollars [VERIFY: EEI capex surveys], the largest peacetime infrastructure program in the sector's history. John Ketchum, NextEra's CEO, frames the demand backdrop as US power consumption growing roughly 55% over the next two decades, about six times the prior decade's pace [VERIFY: Ketchum public remarks]. Every dollar of that spend earns a regulated return, which is why growth utilities now guide to 8-10% annual rate-base growth [VERIFY: investor materials], bond-like businesses compounding at tech-adjacent rates because the runs through their balance sheets.
The physical constraints are where the cycle bites. GE Vernova's heavy-duty gas turbine slots are effectively sold out into 2028 [VERIFY: GE Vernova earnings calls], and large power transformer lead times have stretched to two to three years, roughly double pre-2020 norms [VERIFY: Wood Mackenzie/NREL]. Scarcity along the equipment chain means the companies that already own capacity, or already hold delivery slots, capture the pricing; latecomers pay up or wait. That is the same dynamic playing out at every layer of this story, from queue positions to reactor fuel to transformer steel.
How to frame the trade
The investable map sorts into four baskets. Unregulated nuclear-heavy generators (CEG, VST, TLN) sell existing firm power into a market that suddenly wants decades of it. Regulated growth utilities (NEE, D) convert the same demand into rate-base compounding. The fuel cycle (CCJ, LEU, and the URA ETF as the basket expression) is a play on the enrichment the Russian import ban created [VERIFY: Public Law 118-62]. And the SMR tranche (SMR, OKLO, with GE Vernova's turbine franchise as the picks-and-shovels adjacency) is licensing-milestone optionality. The figure below plots the generator set on the axes that matter: how hard the fleet already runs, how big it is, and what the market pays for it.
FIG. 3 · WHO OWNS THE FIRM POWER
US POWER PRODUCERS · X = FLEET CAPACITY FACTOR (%) · Y = GENERATING CAPACITY (GW) · BUBBLE = MARKET CAP ($B) · SHADED ZONE = NUCLEAR-CLASS UTILIZATION AT SCALE, WHERE THE HYPERSCALER PPAS ARE BEING SIGNED. URANIUM AND SMR NAMES ARE EXCLUDED: CAPACITY-FACTOR AXES DO NOT APPLY TO MINERS OR PRE-REVENUE DEVELOPERS.
Up and to the right at size = always-on fleets the AI buyers want under contract.
← swipe to explore →The base case is that the crunch persists through the decade: LBNL's midpoint path implies US datacenter demand near 480 TWh by 2028 [VERIFY: LBNL scenario range], queue reform helps at the margin, and firm-power owners keep pricing power while the capex supercycle compounds through regulated rate bases. The bear case for the trade is a demand air-pocket: if AI training spend consolidates the way the M&A wave suggests capital already is, some announced gigawatts never get built, and the names priced for scarcity, particularly the pre-revenue SMR tranche, reprice hardest. The honest posture is to own the assets that exist (fleets, fuel, delivery slots) and treat the futures (SMRs, announced campuses) as options. Ezana's screening tools can track the whole stack: build the four baskets, watch PJM auction prints and NRC milestones as the catalysts, and let the queue data, not the press releases, mark the pace of the build-out.
Comments
No one's written in yet. The first thought is always the loudest.