The order books nobody reads: $1 trillion of signed AI demand, straight from the filings

Every AI bubble argument you have read this month — for or against — was probably built from price charts, capex headlines and vibes. There is a class of evidence that settles most of it, and almost nobody quotes it: the order book disclosures buried in SEC filings.

Remaining performance obligations. Revenue backlog. Commitments not yet recognized. Take-or-pay minimums. Different names, same substance: money customers have signed binding contracts to spend, that has not yet appeared in anyone's revenue line. Opinions about AI demand are free. Obligations are filed. Here is the filed record, company by company, pulled through Aether.

If you are the kind of reader who wants a price target at the end, wrong account. If you want to know whether the demand under the buildout is real and contracted, keep going.

Oracle: $638 billion, up 363%

Oracle $ORCL disclosed RPO of $638 billion as of May 31, 2026, up 363% year over year and up $85 billion in a single quarter (June 10, 2026 release). Three years of hypergrowth would still leave most of it unrecognized. Two structural details matter: most of the recent increase came from large AI contracts, and $75 billion of those contracts is customer-prepaid or customer-supplied hardware — buyers are so committed they are financing the seller's buildout. We covered the funding-model tension in our companion piece on Oracle; the demand side of its ledger is not in dispute.

Microsoft: $678 billion, filed this week

Microsoft $MSFT put an absolute number on its order book two days ago: commercial remaining performance obligation of $678 billion as of June 30, 2026, up 84 percent year over year — another $51 billion added in the single quarter since March's $627 billion (Q4 FY2026 release, July 29). Alongside it: Azure growth accelerating to 43 percent, fiscal-year Azure revenue past $100 billion for the first time, and Microsoft 365 Copilot past 30 million paid seats — our full results breakdown. Enterprises are not renting AI capacity month to month. They are signing multi-year obligations for capacity that partly does not exist yet.

Micron: the strongest form of order book that exists

An RPO can, in principle, disappoint — recognition depends on delivery and usage. Micron $MU disclosed something stricter on June 24: 16 Strategic Customer Agreements structured as take-or-pay, with binding volume commitments running to calendar 2030, covering roughly 20% of its DRAM volume and a third of its NAND volume. Fourteen of the sixteen carry cumulative minimum revenue of about $100 billion, backed by a projected $22 billion in customer cash deposits and financial commitments. Take-or-pay with deposits is the hardest demand evidence a filing can contain: the customer pays whether or not they take the memory. Full breakdown in our Micron deep dive.

Amazon: the arc nobody charts

Amazon $AMZN discloses, in every 10-K and 10-Q, commitments in customer contracts (primarily AWS) for future services not yet recognized. The filed arc: $80.4 billion (December 2021), $110.4 billion (December 2022), $157.7 billion (March 2024), approximately $195 billion (June 2025) — with a weighted-average contract life around four years. That is AWS demand more than doubling in contract form across the exact period the industry was supposedly overbuilding. It rhymes with the growth reacceleration in the segment tables: 17%, 20%, 24%, 28% over the last four quarters. Amazon reports Q2 2026 after today's close; the fresh commitments figure arrives with the 10-Q, and we will extend the arc against the document when it lands.

Alphabet: the backlog that keeps re-rating

Alphabet $GOOGL calls it revenue backlog, primarily Google Cloud. Filed arc: $64.3 billion (December 2022), $78.8 billion (June 2024), $108.2 billion (June 2025). Then came the quarter just reported: Cloud revenue up 82% with operating income tripling, and a $49.6 billion equity raise whose stated purpose is "capital expenditures to scale AI infrastructure and global compute" (Q2 2026 8-K, July 22 — our full results breakdown). Posts circulating after the earnings call put the new backlog figure above $500 billion; that number is reported commentary until the Q2 10-Q is filed, and we will update this piece against the document when it lands. That discipline — filed versus said — is the entire method.

Read the fine print like an analyst, not a fan

Order books are the best evidence in this debate, not perfect evidence. Three honest caveats, from the definitions in the filings themselves. RPO and backlog are recognized only as sellers deliver and customers consume, so timing can disappoint even when the commitment is real. Definitions differ: Alphabet excludes cancellable contracts and contracts under a year; Amazon counts only contracts with original terms over one year, and usage can extend beyond the term. Concentration cuts both ways: much of Oracle's increase comes from a small number of very large AI customers, which is exactly why the take-or-pay and prepayment structures matter — they convert relationship risk into contract law.

The tally

Add up only what is filed: Oracle's $638 billion, Microsoft's $678 billion, Amazon's ~$195 billion as of mid-2025, Alphabet's $108 billion as of the same date and climbing, Micron's $100 billion floor. Signed AI-era demand now clears $1.7 trillion without counting a single announcement, MOU or press-conference number — no NVIDIA–SK initiative, no Samsung–Broadcom MOU, none of the letters of intent that dominate the timeline.

And this week filed the other side of the same ledger. The evening Microsoft disclosed its $678 billion order book, Meta $META reported $31.1 billion of quarterly capex that squeezed its free cash flow to $784 million (our breakdown). The sellers' backlogs and the buyers' compressed margins are the same contracts, viewed from opposite ends.

The bubble question, properly phrased, is not "is the spending real" — the spending is contracted. It is "can delivery keep pace with what has been signed." That is a question filings will answer quarter by quarter, in the same footnotes almost nobody reads.

Headlines move prices; filings size the exposure. Small first step: pick one company you own and find its RPO or backlog footnote in the latest 10-Q — one search, ninety seconds — or check any figure in this piece against the source documents on Aether, our SEC search engine.

Sources: ORCL Q4 FY2026 release, June 10, 2026 (acc 0001193125-26-265848); MSFT Q4 FY2026 8-K, July 29, 2026 (acc 0001193125-26-323632) and Q3 FY2026 8-K, April 2026 (acc 0001193125-26-191457); META Q2 2026 8-K, July 29, 2026 (acc 0001628280-26-050596); MU Q3 FY2026 release and call, June 24, 2026 (acc 0000723125-26-000013); AMZN 10-K FY2022 (acc 0001018724-23-000004), 10-Q March 2024 (acc 0001018724-24-000083), 10-Q June 2025 (acc 0001018724-25-000086); GOOGL 10-K FY2022 (acc 0001652044-23-000016), 10-Q June 2024 (acc 0001652044-24-000079), 10-Q June 2025 (acc 0001652044-25-000062), Q2 2026 8-K (acc 0001652044-26-000066). All retrieved and cited via Aether.

Research, not investment advice.