Oracle sent a force-majeure notice to a unit of Blue Owl Capital, the developer of its Project Jupiter campus in New Mexico, citing potential delays in securing enough power. The company insists the 2028 target holds. Take the notice at face value anyway: it is the first time this buildout has written the power constraint into a contract.
Last week this newsletter covered what host communities were asking for — benefits, rate protection, local legitimacy. This week the conversation moves one layer down the stack, from local politics to the physical ceiling those negotiations were always actually about.
The megawatt you can energize, and the date it arrives, now gates frontier capacity more tightly than any allocation of accelerators.
The megawatt you can energize, and the date it arrives, now gates frontier capacity more tightly than any allocation of accelerators.
TrendForce puts global data center power demand at 161 gigawatts this year, up 31 percent, and projects 490.7GW by 2030. Grid supply is estimated to top out at 222.6GW. That is a 268GW gap — roughly the output of 270 standard Korean nuclear reactors.
The United States alone is forecast to run a shortfall above 170GW by 2030.
The mismatch is a timescale problem before it is a money problem. A data center can be built in one to three years. Building and permitting ultra-high-voltage transmission takes five to fifteen.
More than 70 percent of US ultra-high-voltage transformers have already exceeded their 30-to-40-year service life. Transformer lead times that ran 24 to 30 months before the AI buildout now stretch to four or five years, and AEP Ohio has paused all new data center interconnections because the grid cannot absorb the load.
At the end of 2025, more than 2,000GW of generation and storage sat in the US interconnection queue — close to twice the entire installed national fleet — with a median wait past four years.
The industry's answer is to stop waiting. Oracle moved Project Jupiter off gas turbines and diesel generators onto Bloom Energy fuel cells as an onsite microgrid, sized at roughly 2.45GW. That is not a data center power requirement; it is an industrial energy system bolted to a computing system.
It also has its own dependencies. The natural gas pipeline feeding the site slipped to February 2027, and a separate air-quality permit for the fuel-cell system is still pending.
Others are buying generation outright. Meta has agreements with TerraPower, Oklo, Vistra and Constellation for up to 6.6GW of nuclear. Microsoft signed a 20-year deal to revive Three Mile Island. Amazon has secured 1,920MW of nuclear through 2042.
Siemens Energy and Hitachi are calling the result Supercycle 2.0, running through at least 2035.
Now the part that should discipline every projection in this space. On one tracker's US 2026 sample, two-thirds of announced capacity showed no construction progress, and nearly half of the 12GW planned for 2026 is already facing delays or cancellations.
Meanwhile US data center construction spending hit $49.5 billion through April 2026, nearly four times the pace of a year earlier. Capital is moving fast. Steel is not.
The unit economics explain the desperation. SemiAnalysis — a contested, single-source figure — puts a gigawatt of energized AI capacity at roughly $12 billion to $13 billion a year in revenue. On that math, 200MW delivered six months early is worth $1.2 billion to $1.3 billion, and a year of interconnection delay on a gigawatt campus is a multi-billion-dollar opportunity cost compounding against a depreciation clock already running on chips sitting in a warehouse.
Earlier estimates in the same vein: each month of earlier delivery is worth about $500,000 per megawatt of IT capacity. The gap between a six-month build and an eighteen-month one is roughly $3 million per megawatt.
The construction side is being attacked in parallel. On August 10, 2026, Alibaba said its CUBE 5.0 modular architecture cuts delivery of a large AI data center to 100 days, against 6 to 12 months for conventional domestic builds, 12 to 18 in the US, and up to 30 months in parts of Europe and Japan.
The design raises modularization from about 30 percent to 90 percent, splits the schedule into 30 days of factory production and site prep, 50 days of site assembly, and 20 days of commissioning, and has produced one-time test pass rates near 100 percent with container-level initial cost of $200 to $400 per kilowatt-hour against $300 to $600 for traditional builds.
Alibaba's global data center chief, Wang Zhaoyang, framed the stakes bluntly: "Today not a single card is idle."
Read that construction advance carefully, though. The 50-day site phase still requires piping and low-voltage cable trays — site work is compressed, not eliminated. And 100-day delivery solves the building, not the substation.
That is where the equipment layer becomes the story. HD Hyundai Electric's order backlog reached 12.3105 trillion won in the first half — nearly three times annual sales — at a 25.0 percent operating margin. LS ELECTRIC posted record quarterly revenue of 1.577 trillion won, raised its annual order target to 6.5 trillion won, is expanding in Utah, and has secured commercialization technology for 800V DC solid-state transformers and solid-state circuit breakers. Hyosung Heavy Industries passed a 17.5 trillion won backlog and formed a high-voltage circuit breaker joint venture with Quanta Services, the top US power EPC. Sanil Electric, which supplies fuel cells for onsite self-generation, runs a 37.8 percent operating margin. Taihan Cable's backlog passed 4 trillion won.
These are manufacturing margins, not utility margins. That is the tell.
There is an uncomfortable supply-chain fact sitting under all of it. China remains the world's largest producer of the electrical gear — transformers, switchgear, substations — needed to connect data centers to the grid. The country best positioned to compress the build and the country supplying the long-lead electrical layer are the same one, which is awkward for a policy apparatus that has spent three years treating chips and model safety as the strategic surface.
NYU's DRI put it plainly in June: the next AI bottleneck may not be a model, a chip, or a regulation — it may be a substation. Power is the first AI policy, and the interconnection queue, transformer supply chain and local permitting process are strategic bottlenecks, not routine utility matters.
Scale makes the politics unavoidable. Global data center electricity demand will exceed 1,000 terawatt-hours this year — more than Japan consumes in total — and the IEA projects roughly 945 TWh by 2030, more than double 2024's 415 TWh. US data centers could reach 12 percent of national electricity by 2028, up from 4.4 percent in 2023. A single training facility can pull 100 to 500MW continuously, which is a small city.
Which is why the host-community bargain from last week is not a public relations layer on top of the power problem. It is the same problem. Goldman Sachs has warned data center demand will add 0.1 percent to core inflation in both 2026 and 2027. One Carnegie Mellon analysis estimates an 8 percent rise in the average US electricity bill by 2030, exceeding 25 percent in the highest-demand markets.
Every one of those dollars is a permit hearing.
Watch what changes operationally. Site selection logic shifts from near customers to near power, near cooling, near cheap land. Training runs get queued rather than scaled. Regional price divergence widens.