A viral post says America's tech giants are hiding $1.65 trillion of AI debt where nobody looks. We rebuilt the tally from SEC filings: the total brackets Nikkei's figure, and both company numbers Nikkei names match to the decimal. The part that breaks is the word "debt", and fixing it makes the story bigger.
"$1.65 trillion in debt that doesn't show up on any balance sheet... More than the $1.35 trillion they actually report. Almost all of it is AI. Nobody reads the footnotes."
Kohei Yamada of Nihon Keizai Shimbun published the analysis on 20 July 2026 in Japanese and on 21 July in English, as "five US tech giants' hidden debts soar to $1.65tn on opaque AI funding"1. The five are Alphabet, Microsoft, Amazon, Meta and Oracle. The $1.65 trillion is Nikkei's own tally, roughly eight times the level of about four years earlier, built from two footnote categories in SEC filings: leases signed for data centres that have not started operating, and purchase commitments for GPUs, servers and power that have not been delivered.
We rebuilt the same categories from the latest filings ourselves. The reconstruction lands between $1.52 and $1.67 trillion, which brackets Nikkei's number2. And in the two cases where Nikkei published a company figure, it reconciles exactly: Meta's $420bn is $182.88bn of leases not yet commenced plus $237.67bn of commitments, $420.55bn to the decimal3. Oracle's $273.3bn is $260bn plus $13.309bn, straight from its 10-k for the year to 31 May 20264.
Billions of dollars, from each company's latest filing: 10-qs to 31 March 2026 for four of them, Oracle's 10-K to 31 May 2026.
| company | leases not yet commenced | purchase obligations | footnote total | |
|---|---|---|---|---|
| meta | 182.9 | 237.7 | 420.6 | Matches Nikkei exactly |
| oracle | 260.0 | 13.3 | 273.3 | Matches Nikkei exactly |
| microsoft | 196.6 | ~110 | ~306.6 | Po from FY2025 10-K |
| alphabet | 75.6 | 232.7 to 332.4 | ~308 to 408 | Range |
| amazon | 106.4 | 103.8 | 210.1 | Incl. Non-AI retail items |
| five companies | ~1,520 to 1,670 | Brackets Nikkei's 1,650 |
Microsoft's leases not yet commenced, "primarily for datacenters", more than doubled in nine months, from $92.7bn at 30 June 2025 to $196.6bn at 31 March 2026. The totals are not purely AI: Amazon's commitments include retail content and energy, Meta's include reality labs hardware. The "almost all AI" attribution is Nikkei's reading, and the categories and growth make it plausible, but no company publishes the split.
The comparison the post should have made. Grey is what each company has actually borrowed; orange is what sits in the footnotes. Approximate, sourced per company above.
None of this is illegal and most of it is ordinary accounting. Under us GAAP, a lease that is signed but has not commenced creates no balance-sheet liability until the data centre starts operating, and a purchase contract for undelivered GPUs is a disclosure, not a debt. The footnote is exactly where the rules say these belong.
The migration is also already visible, in Oracle's own numbers. In its fiscal 2026 alone, recognised lease liabilities jumped from $16.4bn to $37.9bn as facilities came online, and $260bn more is contracted to commence between fiscal 2027 and 20294. Oracle's own counterweight is $638bn of remaining performance obligations, contracted future revenue, with heavy OpenAI concentration4. Four years ago Oracle's footnote stack was $8.9bn. It is now 30.7 times that.
Michael Burry amplified the Nikkei piece on x on 21 July, quoting its first line7. Worth knowing when you read him: he holds Oracle put options, a position recently halved.
The on-balance-sheet story broke last week: Alphabet doubled its long-term debt in six months, took quarterly free cash flow negative for the first time since its 2004 listing, and guided to as much as $205bn of capex this year. That was the visible half.
The footnote layer is the committed half: roughly $1.5 to 1.7 trillion of contracted spending that lands on these balance sheets as the data centres switch on and the chips arrive. If AI revenue ramps, it is absorbed. If it does not, the biggest companies in the market get more leveraged on a schedule they have already signed. That is a conditional. The post selling you doom is also selling you a trading group.
The European angle is scale. The loudest sovereign-cloud headline in Europe, schwarz's Lübbenau, is eur 11bn over an unstated lifecycle. The American footnote stack is about a hundred and forty times that, and it is contractually committed. The compute gap is not closing on current money. It is widening in the fine print.