The Number That Should Worry Every Investor
Three point two trillion dollars.
>That is the sum of off-balance-sheet AI obligations sitting in the SEC footnotes of just five U.S. tech giants — Alphabet, Microsoft, Amazon, Meta, and Oracle — as of September 2026. Nikkei Asia’s November 4 report puts the figure at roughly triple the combined on-balance-sheet debt those same companies report, up 90 percent in six months. The earlier mid-2026 tally for the same cohort was $1.65 trillion. The jump is not a rounding error.
This is not a loan scandal. It is not accounting fraud. It is something more structural and, in some ways, more troubling: the quiet engineering of a financial commitment so large that even the people buying the stock may not fully grasp what has been signed away.
What “Hidden Debt” Actually Means
Let me be precise, because the terminology matters. The $3.2 trillion is not conventional debt. It does not appear on the front page of the balance sheet. It lives in the footnotes — specifically, in two buckets: future purchase commitments and uncommenced leases.
Future purchase commitments are legally binding agreements to buy chips, servers, and data-center hardware at set volumes and prices over multi-year periods. Uncommenced leases are contracts to lease data-center space and equipment that have been signed but not yet activated. Both are standard accounting treatment. Under current rules, companies do not have to book these as liabilities until the underlying asset or service goes live.
That is technically correct. It is also exactly the mechanism that lets the numbers grow this fast without triggering the alarm bells that come with a rising debt-to-equity ratio on page one.
The Run Rate Is the Story
Nikkei first cut this same five-company cohort at $1.65 trillion in mid-2026. Six months later, the figure is $3.2 trillion. A 90 percent jump in half a year signals that the hyperscalers are committing to spend at a pace that far exceeds the revenue most of them can realistically generate from AI services in the near term.
A Wall Street Journal analysis, cited by multiple outlets including AOL and Fortune, puts the aggregate across nine companies — adding Nvidia, Broadcom, AMD, and SpaceX — closer to $3 trillion based on filings through June 2026. A separate tally from Global Markets Investor in October pushed the combined figure for a broader set of names to $4.1 trillion, driven by Q3 disclosures from Meta (+$195 billion), Alphabet (+$164 billion), Amazon (+$131 billion), Microsoft (+$35 billion), and SpaceX (+$51 billion).
The exact number depends on which companies you include and which footnotes you read. But the direction is unambiguous and consistent across every source: the obligations are climbing exponentially. If you are following the broader pattern of inflated AI expectations, this is the other side of the same coin.
How the Math Gets Hidden
The most instructive single detail I found was in the Substack analysis by Nitish Asthana: Alphabet holds roughly $100.2 billion in on-balance-sheet long-term debt — a number investors watch closely. Its off-balance-sheet purchase commitments for AI infrastructure alone sit at approximately $811 billion, with another $91 billion in uncommenced leases. That is a 10-to-1 ratio of footnote commitment to headline debt for one company.
Microsoft reports about $40 billion in long-term debt. Its purchase commitments are $329 billion; uncommenced leases are $229 billion. Meta is now carrying more than $349 billion in purchase commitments and $347 billion in leases that have not started.
The structural trick, as Asthana and others note, is the “neocloud” arrangement. Instead of building a data center and booking the capex immediately, a hyperscaler signs a take-or-pay lease with a specialized data-center operator. The operator borrows against the contract and builds the facility. The tech company gets guaranteed capacity spread over ten to fifteen years. The debt sits on the operator’s balance sheet, not the tech giant’s. The lease commitment lands in a footnote — as an “operational expense” — rather than in the liabilities column.
This is legal. It is disclosed. But when you have a thousand dollars of footnote obligations for every dollar of reported debt, the word “disclosed” is doing a lot of heavy lifting.
The Revenue Problem Nobody Wants to Say Out Loud
Here is the tension that makes this a real story rather than an accounting curiosity: these commitments must be paid whether or not the AI revenue materializes to cover them.
A multi-year GPU purchase contract does not pause because the model does not monetize. A data-center lease does not get cancelled because the inference demand plateaued. The hyperscalers have structured themselves to spend whatever it takes to secure capacity before their competitors do. The problem is that capacity is only an asset if someone pays to use it.
Bloomberg’s analysis in early November projected that if all current AI infrastructure plans are realized, total investment could exceed $10 trillion. Even at a fraction of that, the capex intensity dwarfs any previous technology cycle. The dot-com era looks restrained by comparison.
And while the big five dominate the headlines, the obligation is spreading downstream. Nvidia, Broadcom, AMD, and even SpaceX are accumulating their own off-balance-sheet commitments as they build the physical substrate — chips, rockets for satellite ground links, fabrication capacity — that the AI boom requires. The exposure is not concentrated in one set of income statements.
What It Means for the Philippines
I manage an ICT division in a Philippine government institution. Our procurement cycles already feel squeezed by the global chip shortage and the soaring cost of servers. What happens when the five largest technology companies in the world have collectively locked up trillions in future capacity and hardware?
The answer is not abstract. It means the hyperscalers that Philippine businesses and agencies depend on for cloud services — AWS, Azure, Google Cloud — have priced in their own infrastructure scarcity. It means the bargaining power in contract negotiations tilts further toward the providers. It means that if you are planning a digital transformation project with multi-year cloud spend, the pricing assumptions you made last year are already wrong. I have seen budget proposals from other government offices that treated cloud as a variable cost. It is not. Under these take-or-pay structures, it is becoming a fixed commitment dressed in variable clothing.
This is the local angle that gets lost in the aggregate headlines. The $3.2 trillion is not someone else’s problem. It is the foundation on which every cloud bill, every SaaS renewal, and every government digital service contract for the next decade will be built. If you are evaluating whether to self-host your AI infrastructure rather than rent it, this is exactly why that conversation is becoming urgent.
The Accountability Gap
The Bloomberg Opinion column by Bill Dudley, former president of the Federal Reserve Bank of New York, put it plainly: the AI investment boom demands stronger financial oversight. The current disclosure framework was designed for an era when purchase commitments were measured in millions, not trillions. It does not require companies to book obligations that may not begin for years. It does not force them to disclose the revenue assumptions behind the spending. This gap is exactly the kind of structural blind spot I have been writing about when it comes to leadership and accountability in the AI era.
Investors who read only the balance sheet see a manageable debt load. Investors who read the footnotes see a different picture: companies that have pre-committed the equivalent of their entire asset base to a build-out whose returns are entirely hypothetical. The disconnect between those two views is widening by the quarter, and the market has not yet decided which view is correct.
The right question is not whether the accounting is legal. It is whether the accounting is honest about the risk.
So Where Does That Leave Us?
The AI boom is real. The infrastructure being built is real. The chips, the data centers, the power contracts — they are all physical and they are all happening. But so is the obligation.
For someone watching from the sidelines, the $3.2 trillion figure is a reminder that every gold rush creates its own geology of risk. The companies selling the picks and shovels in this rush have signed contracts to buy so many picks and shovels that they may never be able to pay for them if the miners do not strike gold.
That is not a reason to stop believing in AI. It is a reason to look past the hype and understand the balance sheet — all of it, not just the first page.
I’d love to hear how your organizations are thinking about multi-year cloud and AI infrastructure commitments. Drop a comment or reach out. These numbers are too big to ignore.