The trillion-dollar reallocation.
The AI economy is usually told in adjectives. This report tells it in receipts: what five companies swore to the SEC they spent, what the government's own laboratory measured the buildout drinking from the grid, who signed for nuclear reactors, and whose jobs moved out of the buildings the money moved into. Every figure below traces to a source you can pull yourself.
authored from the space's records · retrieved 2026-08-26 · rendered by the GRIDINT builder
What we did, and what this can and cannot say.
We assembled four record sets and read them together. Capital expenditure comes from SEC XBRL company facts — annual figures from 10-K filings, each carrying its accession number; two companies (Amazon, Nvidia) file the concept under their own tag, which we name rather than blend. Power comes from the Department of Energy's LBNL 2024 study and the IEA's 2026 update — measured figures and projections kept typographically separate throughout. Workforce cuts come from a curated ledger where no event enters without a filing-class receipt; it is deliberately small, and its rejection queue is part of the record. Industry exposure comes from a tracking engine's model outputs, and is labeled as model output everywhere it appears. What this report cannot say: intent, causation between any single cut and any single dollar of spend, or anything about companies outside the tracked sets. The analysis is AI-assembled from these records; the claims are human-written; every figure traces to a row.
Five things the records establish.
- The spend inflection is filed fact, not narrative. Every major's capex line bends after late 2022 — Microsoft 4.8×, Nvidia 6× by their latest fiscal years. → exhibit A2
- Amazon is the quiet giant. Under its own filing tag, its FY2025 capital spend — $131.8B — is the largest single figure in this report. → A2
- The grid is the binding constraint being bought off. 4.4% of US electricity measured in 2023; a government floor of 6.7% by 2028; and every major buyer signing for nuclear supply. → B4 · B6
- The cuts and the spend are one decision. Meta's −8,000 and its $135B land in the same season; across the ledger’s fifteen receipted events (edition 2), −115,875 heads sit beside record capital formation. → C3
- Payoff is real but uneven. By the engine's read, 19 of 31 industries can demonstrate results; three remain pure spend — and the most exposed verticals are not the ones spending most. → D2
The spend, under oath.
Guidance is a promise; a filing is a sworn statement. Strip away every press release and analyst deck, and the AI buildout is still unmistakable in the one document a company cannot inflate: the 10-K. Microsoft's capital expenditure quadrupled in four filed years. Amazon's passed $131 billion. Nobody is narrating this chart — the companies filed it.
"The inflection is not an opinion. It has an accession number."
What the money becomes: buildings that drink.
Capital expenditure turns into data centers, and data centers turn into demand on the grid. In 2023 the Department of Energy's own laboratory measured American data centers taking 4.4% of the nation's electricity. Its projection for 2028 is a range — 6.7% at the floor, 12% at the ceiling — and this report draws it as a range, because pretending a projection is a line is how charts lie. The floor is the headline; the ceiling is dashed.
The other door of the same buildings.
While the capex lines climbed, fifteen tracked companies have announced 115,875 job cuts (the ledger, edition 2) — and the strongest single fact in this report is a pairing: Meta announced 8,000 cuts and $135 billion of capital spending in the same season. That is not a company tightening its belt. That is a company changing what it buys with its money — silicon instead of salaries. The ledger below is deliberately small: no event enters without a receipt, and each row wears its filing class.
Where it lands, industry by industry.
Spend is not payoff. GRIDINT's tracking engine scores thirty-one industries on whether AI results are demonstrated, merely emerging, or still just spend — and scores each vertical's fragility from the companies it tracks. Its current read: nineteen industries can already point at demonstrated payoff; three are spending with nothing yet to show. These are the engine's model outputs, presented as exactly that — the method is on file, and no score below pretends to be an observed fact.
What follows from this.
For operators: the reallocation is structural — capital is moving from payroll to compute at filing-grade scale, and planning that treats it as a cycle will misread it. For the energy question: the floor case alone (6.7% of US electricity by 2028) makes power procurement a board-level topic for any data-dependent business, and the nuclear ledger shows the largest buyers acting on exactly that reading. For analysts: the honest seams in this data — tag differences, small receipted ledgers, model-output labels — are where the real signal lives; sources that blend them are hiding the interesting part. These are implications, labeled as analysis — the records above are the part that is sworn.
Four record sets — filed capex, the power studies, the deal ledger, the cuts ledger — land in a grid space as records with payload, source and retrieval date, each at its own address. The address chips on the exhibits above are real: A2, B4, C3, D2 are where these cards live on the sheet.
Actuals never share a line style with projections. Ranges are bands, floors are headlined. Estimates say whose they are. Absences are stated, never implied clean. A row without a source does not exist.
This report was assembled from those records by the GRIDINT builder. The claim sentences are a person's; the figures are the board's. This edition is dated and stands as published — it updates only as a new dated edition with an extended dataset, never silently.
Nothing here is investment advice, and the engine's scores are a model's view with its method on file. What a company filed is its own sworn statement; what anything means is left with the reader — with receipts.
This page came off a board. Walk into it.
The demo space is the real product holding these records — writing turned off, everything else on. Open the cards, read the receipts, export the lot as plain JSON.