The buildout runs on electrons, and the electrons now have receipts.
What data centers actually draw from the grid, what the government's own lab projects, and the nuclear supply deals the AI buyers have signed — actuals separated from projections, projections drawn as bands, ceilings named as ceilings.
Public-institution data plus a curated deal ledger — source class on every row.
The share curve — actual, then a band.
One measured point and one projected range. 4.4% of US electricity in 2023 is the Department of Energy lab's measured figure; 6.7% to 12% by 2028 is its projection, and it is drawn as a band because that is what a projection is. The floor is the solid line — the honest headline. The ceiling is dashed — real, possible, and not promised.
The global doubling.
The IEA's 2026 update: roughly 485 TWh of global data-center consumption in 2025, projected toward ~950 TWh by 2030, with AI-focused facilities tripling inside that. For scale — the 2030 figure approaches the annual electricity consumption of Japan.
The nuclear ledger — deal by deal.
Every major AI buyer has signed for nuclear supply: a restart, an SMR order-book, twenty-year purchase agreements, a multi-vendor procurement. Each row carries its announcement date and its source class. Where a figure is an announced strategy rather than a contract, the row says so.
How this was made.
LBNL/DOE and IEA for the demand picture; a curated deal ledger for supply — announcements are not API objects, so they are curated with the same discipline as the layoffs ledger: no row without a source.
Actuals and projections never share a line style. Bands for ranges, floors headlined, ceilings dashed and named. Press-reported terms labeled press-reported.
The EIA's public bulk files — free, no key — join the harvest cadence to bring the actuals layer onto the board as it updates.
The capex wave on the filings page is this same story's other ledger: the money and the megawatts are one buildout, receipted twice.