AI capex beyond the hyperscalers

Capex is what a company reports spending on property and equipment, on its own accounting basis: CoreWeave’s deployed assets, TSMC’s consolidated capex, Micron’s net capex. Each reading names that basis. Chipmaker figures are whole-company spend, the cleanest public read on AI chip capacity.

They are commitments, and the ledger labels them that way. OpenAI’s USD 38bn and USD 100bn AWS deals and Anthropic’s USD 30bn Azure deal are future cloud spend. They show where the capex lands: in the supplier’s data centers.

Add within one value type, one currency and one period. Across them the rows overlap: annual totals contain quarters, budgets contain their components, and a lab’s cloud contract pays for racks a neocloud reports as capex. Filter by value type and currency first, then total.

At hyperscalerscapex.com. Microsoft, Alphabet, Amazon, Meta and Oracle each have one summary row here that links to the full quarterly series. This ledger covers the rest of the AI capital stack: specialist clouds, labs, sovereign programmes, chipmakers and telcos.

Some deals are announced in megawatts or GPUs with no dollar figure. Anthropic’s Colossus 1 agreement is more than 300MW and over 220,000 GPUs; Stargate UAE is a 1GW cluster. The ledger records the capacity as disclosed and leaves the amount blank.

Use its permanent detail page and the original source. CSV and JSON exports retain value types, currencies, notes and provenance. The methodology explains accounting differences, overlapping announcements and historical forecast vintages.