You read the capex headlines. One tech giant commits tens of billions a quarter to AI infrastructure. Another matches it a week later. Those numbers are large, and they sit on the balance sheet where any analyst can find them.
There is a bigger number, and it sits somewhere quieter.
A Wall Street Journal analysis of the most recent securities filings from nine top tech companies points to roughly $3 trillion of off-balance-sheet commitments, mostly tied to AI. In the WSJ's words, "Nine top tech companies had some $3 trillion of off-balance-sheet commitments mostly related to AI." That total is several times the roughly $600 billion in traditional capital spending those same companies reported over the trailing year. And it is disclosed in the footnotes of the filings, not on the face of the balance sheet.
If you sit on a board, own a P&L, or read a competitor's financials to size up their AI bets, that distinction is worth a few minutes of your attention.
What did the Wall Street Journal actually find?
The WSJ added up disclosure footnotes across the latest filings of Alphabet, Amazon, Meta, Microsoft, Oracle, Nvidia, Broadcom, SpaceX, and AMD. The total came to about $3 trillion in commitments that are contractual and mostly AI-related, yet not booked as liabilities. No single filing states that figure. It is the WSJ's own computation across the nine.
In contractual, mostly AI-related commitments across nine top tech companies, disclosed in footnotes rather than booked as liabilities.
Source: Wall Street Journal, 2026Two things stand out about the size of it. The commitments are growing faster than traditional capex. And they run about triple what the same companies owe under their outstanding leases and long-term borrowings combined.
So the reported debt tells a partial story. The bigger story is one page deeper, in language most readers skip.
The bigger story is one page deeper, in language most readers skip.
Where does the $3 trillion hide?
Per the WSJ, the total splits two ways. About $1.2 trillion is leases that have not yet commenced, mainly data centers that were signed but have not started. About $1.9 trillion is purchase commitments for chips, energy, and data-center infrastructure. Both categories are climbing faster than the capex line most people watch.
Here is why each one lives off the balance sheet for now.
A lease that has not commenced is a signed agreement for space or facilities that have not started operating. Under current accounting, the obligation gets recognized when the lease begins, not when the ink dries. So a company can commit to years of data-center payments today and record the liability later, once the site comes online.
A purchase commitment works the same way. It is a promise to buy: gigawatts of power, batches of chips, construction capacity, often under multi-year supply agreements. The dollars are real and the contracts are binding. The balance sheet stays quiet until the goods and services are delivered.
Neither of these is a trick. Both follow the disclosure rules. The practical point for a leader is simpler: the headline capex figure understates how much a company has actually promised to spend on AI.
What does one company's filing reveal?
Alphabet's Q2 2026 Form 10-Q puts these categories in one place, so it is a clean example of the method the WSJ used. As of June 30, 2026, Alphabet reported $811.0 billion in purchase commitments and other contractual obligations, of which $200.7 billion was short-term, mainly for technical infrastructure and inventory.
The filing states it plainly: "As of June 30, 2026, we had material purchase commitments and other contractual obligations totaling $811.0 billion, of which $200.7 billion was short-term."
The lease disclosure is just as direct. The same 10-Q reports that Alphabet had "entered into leases, primarily related to data centers, that have not yet commenced with future lease payments of $85.2 billion that are not yet recorded. These leases will commence between 2026 and 2031 with non-cancelable lease terms between one and 26 years."
Read that timeline again. Payments running as far out as the early 2030s. Terms as long as 26 years. And $85.2 billion of it not yet on the balance sheet as of mid-2026.
The filing goes further into structures a casual reader would never see:
- Future funding commitments to variable interest entities of $21.9 billion, up from $1.1 billion at the end of 2025. Of that, $20.0 billion is a capital-funding commitment to a private company through 2030, accounted for as an equity derivative.
- Infrastructure backstops provided through financial guarantees and credit derivatives, with maximum potential future payments of $7.6 billion and $43.8 billion.
- An agreement to provide an estimated $24.1 billion of additional future backstops to support data-center and energy build-out.
One company. One quarter. Several different ways of committing capital to AI, each disclosed in a different footnote, none of it visible from the capex headline alone.
Why should a board care about a footnote?
Because footnoted obligations can move onto the balance sheet. When leases commence or accounting standards tighten, today's disclosed commitment becomes tomorrow's recognized liability. Per the WSJ, that migration could change how much a company appears to owe materially, and it could happen quickly. The off-balance-sheet total already runs about triple outstanding leases and long-term borrowings.
That is a governance question before it is an accounting one. A director signs off on risk and disclosure quality. A CEO answers for the true scale of the company's commitments when an analyst or a lender asks. If the honest answer lives three footnotes deep, the board should know it is there and be able to speak to it.
That is a governance question before it is an accounting one.
None of this is unique to the nine companies the WSJ studied. Any organization signing multi-year AI infrastructure deals, whether a data-center lease, a large chip order, or a long-term cloud commitment, is creating obligations that the headline budget line may not fully show. The scale differs. The pattern is the same.
So the useful question for your own company is how much you have committed to spend, and where that commitment is recorded.
How do you read an AI commitment the way an investor does?
You look past the number in the press release and check the number behind it. In a filing, that means reading the sections titled purchase commitments, leases not yet commenced, and variable interest entities. In a vendor proposal, it means asking what the total multi-year obligation is, not just this year's invoice.
This is a proficiency skill, and it has a name. In the 7 Levels of AI Proficiency, Level 3 is called the Lieutenant, the Critical Thinker. That level is defined by critical evaluation: checking a claim against your own judgment instead of accepting what sounds good on the surface. The same habit that keeps a Level 3 Lieutenant from taking an AI's first answer at face value is the habit that keeps a leader from taking a capex headline at face value.
You do not need an accounting degree to build it. You need three questions and the willingness to ask them:
- What is the total commitment across all years, not just the current one?
- Where is that obligation recorded, on the balance sheet or in the footnotes?
- What would have to happen for it to move from one to the other?
Ask those three about a competitor's filing and you can tell a company that is genuinely all-in on AI from one that only says it is. Ask them about your own vendor contracts and you will size your real exposure before it surprises you. That is what reading like an investor looks like, and it is a learnable skill.
Where to start this week
Pick one AI commitment your organization has made this year. A cloud contract, a data-center lease, a large software or hardware order. Find the total multi-year obligation, not just this quarter's cost. Then find where it is recorded.
If that number surprises you, you have already done the exercise that this whole story is about. And if you want a way to measure how your team reads situations like this one, the 7 Levels of AI Proficiency gives you a shared standard for the critical-evaluation skill a Level 3 Lieutenant brings to every AI decision.
The headline is rarely the whole bill. The footnotes are where the full commitment is written down.
Related reading: Level 3: The Lieutenant (Critical Thinker).
Sources
- Why Big Tech's AI Spending Is $3 Trillion Higher Than It Seems (Wall Street Journal analysis, syndicated via MSN)
- Alphabet Inc. Form 10-Q, Q2 2026
Frequently Asked Questions
Which companies did the analysis cover?
Nine: Alphabet, Amazon, Meta, Microsoft, Oracle, Nvidia, Broadcom, SpaceX, and AMD. The roughly $3 trillion figure is the Wall Street Journal's sum of disclosure footnotes across their most recent securities filings. It is a WSJ computation, not a line item in any one filing.
Is putting these commitments in the footnotes allowed?
Yes. The obligations are disclosed, and Alphabet's Q2 2026 10-Q spells them out in detail. Under current accounting, a lease is recognized when it commences and a purchase commitment when goods are delivered. The concern is visibility and disclosure quality, not a hidden liability.
What would change if these moved onto the balance sheet?
Per the WSJ, reported obligations could change materially, and quickly, if leases commence or standards tighten. The off-balance-sheet total already runs about triple the companies' outstanding leases and long-term borrowings, so the recognized figure could climb well beyond what the current balance sheet shows.
Bring us the work that keeps coming back to you.
We will trace the workflow, identify where a system can create capacity, and decide what should remain with your people.