More than two months ago, long before most Wall Street analysts had any clue that the bulk of the AI buildout commitments were diligently hidden in various off-balance sheet SPVs and other (perfectly legal) accounting gimmicks, we wrote a lengthy article detailing just that, and explaining why far beyond the $1 trillion (and rapidly rising) in annual plain vanilla capex — all of which now has to be funded through debt issuance since free cash flow across the hyperscaler universe is negative for the foreseeable future (if not forever) — which most pundits obsess over daily…
… the real risk was in the “The $1.8 Trillion Off-Balance Sheet Time Bomb At The Heart Of The AI Supercycle.”
We followed up on this critical topic after the latest batch of hyperscaler earnings in late July, which revealed that the nearly $2 trillion in spending commitments had exploded in the second quarter, with just GOOGL and META north of $1.5 trillion, which prompted us to conclude that “across all hypers, off BS commitments are now $3+ trillion, double in one quarter.”
Many, traditionally those who had no idea what we were talking about or the implications of the above findings, took the conventional route and either mocked or slighted what we had found.
But not everyone: first it was Nvidia which, scrambling to distance itself from the circular financings that were meant to fund precisely these kinds of gargantuan off balance sheet SPVs, unveiled its $500 billion arrangement with a handful of private credit firms (an arrangement that was nothing new, as Nvidia had already done just that particular deal on numerous occasions, and it merely formalized it in the form of an open-ended MOU).
Second and more important, two months after our first warnings, the WSJ finally caught up, and in an article over the weekend titled “Why Big Tech’s AI Spending Is $3 Trillion Higher Than It Seems“, it confirmed everything we have been reporting since early June.
Not to put too fine a point on it, but WSJ authors Rudegeair and Santilli echoed everything we said in our original report more than two months ago, to wit:
Each quarter, big tech companies disclose their massive capital expenditures on artificial-intelligence infrastructure, from data centers to chips. But those figures don’t come close to expressing the full extent of future spending to which Google parent Alphabet, Meta Platforms, Oracle and many others have committed. That is because a huge swath of their coming financial obligations aren’t reflected on their balance sheets.
Nine top tech companies had some $3 trillion of off-balance-sheet commitments mostly related to AI, according to a Wall Street Journal analysis of footnotes in their most recent securities filings. Those obligations are growing faster than traditional “capex,” which totaled about $600 billion over the past year they reported, and were about triple what the companies owe under their outstanding leases and long-term borrowings.
Back in June, we showed this staggering surge in off balance sheet purchase commitment as follows. Needless to say, the latest number is substantially higher.
And this is how the WSJ compared side by side the on (capex) vs off balance sheet obligations at the 6 hyperscalers. Needless to say, the management teams at the big AI spenders are doing everything in their power to cover up just how spectacularly massive their true spending plans are.
The WSJ also takes a look at the off B/S SPVs we have been discussing ever since we explained back in January why Meta’s Beignet funding model will be imitated by all of its peers throughout 2026 and onward.
Of course, that’s precisely what happened and this is how the WSJ lays out the funding of Meta’s gigantic “Hyperion” data-center project in Louisiana, which is the size of about 1,700 football fields, and which “helps explain how big obligations wind up off tech companies’ balance sheets.”
Meta initially agreed to lease Hyperion for a four-year term starting in 2029, with options to renew for up to 20 years. It guaranteed that it would make bondholders whole if it doesn’t stay the entire two decades. The company doesn’t think payments under that guarantee are probable, so it hasn’t recorded any liability on its balance sheet.
In accordance with accounting rules, Meta’s Hyperion lease obligations will remain off balance sheet until it starts paying rent. It said its aggregate initial lease commitment is about $12.3 billion. Meta disclosed $347 billion in total obligations for leases that haven’t kicked in yet, including for Hyperion, as of June.
This is the WSJ’s version of our graphic from 8 months ago:
Remember what we said about expecting “hundreds of these in 2026?” Apparently it came as a surprise to the WSJ, which writes that “across the companies the Journal analyzed, promises of payments under these uncommenced leases totaled $1.2 trillion in off-balance–sheet obligations, or about four times more than what was disclosed a year earlier. In addition to Meta, the Journal reviewed commitments for Alphabet, Amazon.com, Microsoft, Oracle, Nvidia, Broadcom, SpaceX and Advanced Micro Devices.”
Of course, as we explained two months ago, the uncommenced leases are just a part of the off B/S liabilities. There’s much more:
“Data centers get stuffed with a lot of hardware, including the Nvidia chips that are used to train and run models and memory chips that store information. To buy all that, companies sign long-term contractual agreements well in advance to lock in production from their suppliers.”
… which is why Nvidia has the most to lose if funding deals that enable the AI bubble aren’t in place, which is precisely why the company rushed and forced the various private credit firms to deliver the $500 billion deal announced last week. Putting a number to it:
“Those and other purchase obligations at the companies the Journal examined stand at a whopping $1.9 trillion. Under accounting rules, purchase commitments typically remain off balance sheet until a product or service is delivered.”
The uncommenced leases and purchase commitments are the bulk of the off balance sheet “ticking time bomb” we discussed in June. Only then it was $1.8 trillion. It has since grown by 50% to $3 trillion…. in two months! And this is how the WSJ redid another one of our June charts:
What happens next depends very much on whether or is an optimist… or not.
For the former, the WSJ says that “there are reasons to believe tech companies will make good on all their obligations. Optimists see the skyrocketing demand for AI tools—which has lifted the stock market and led to shortages of key hardware—as a proof point that demand is going to be strong for years, and the money to pay off all these bills will be rolling in.”
As for the others, well the problems are only starting. First, this unprecedented explosion in off balance sheet liabilities in “a worrying sign that some tech companies that once seemed to have fortress balance sheets have needed to tap the capital markets frequently.”
Alphabet and Amazon recently posted results showing negative free cash flow, meaning their capital spending exceeded the cash they brought in from operating their businesses.”
And that is before considering the implications of trillions in off-balance–sheet commitments. Whether or not the revenues ever arrive, purchase commitments and signed leases can’t be canceled, for the most part.
The worst case scenarios is that these tech giants companies, which until recently had pristine balance sheets will be paying an expensive tab for infrastructure that they can’t profitably use “if things go wrong.” These obligations could also lead increasingly indebted companies to have to borrow even more.
One way such a catastrophic scenario may happen, is if Chinese open-weight models accelerate their market share grab by dramatically undercutting the token prices of frontier models (as they have been doing), and lead to a collapse in hyperscaler profits as compute creation shifts to China and its far cheaper chips, not to mention far more abundant sources of cheap energy… which is precisely what is happening right now, with token prices in historic free fall, crashing more than 50% in just the past few weeks as Chinese open-models unleash a historic race to the bottom.
And since the entire WSJ is effectively re-print of what we said in early June, we will leave readers with our conclusion from then as it is just as accurate today as it was two months ago (and $1.2 trillion less in off balance sheet obligations).
Vendor financing and repurchase-style arrangements mean a single counterparty’s stress can propagate through several balance sheets at once. Think of it as rehypothecated leverage. Concentration compounds it: the >$2tn of remaining performance obligations across major AI players is built on a handful of very large, long-duration contracts, so the backlog that justifies the spending is also a concentrated counterparty exposure.
The off-balance-sheet items aren’t fraudulent or even unusual accounting — they’re mostly standard treatment. The hazard is timing and visibility, and it shows up in three deferred effects:
Capex is being committed ahead of the revenue and free cash flow meant to support it; capex revisions have run well ahead of sales and FCF revisions. The depreciation from all this spending is largely still deferred, rising construction-in-progress balances (ORCL ~+200%, META ~+90% YoY) mean today’s margins haven’t yet absorbed the expense already locked in. And when it does land, cumulative depreciation for MSFT, ORCL, META, and GOOGL could exceed $520bn over three years, pushing depreciation as a share of revenue sharply higher (ORCL potentially 7%→28%, META 9%→19% by FY28). To hold margins, every other cost line has to fall — which only works if sales rise to match, and so far they haven’t kept pace.
Even a careful analyst can’t fully size the exposure, because classification and disclosure are inconsistent. Whether a compute-capacity deal even counts as a lease turns on judgment calls (does the contract specify a particular GPU/rack, who controls the workloads, etc.). Disclosure varies accordingly: Nvidia quantified $30bn of cloud-service commitments and Oracle $10bn, while Meta declined to quantify the cloud-capacity portion of its $238bn in commitments. That inconsistency means any cross-company comparison of “true” leverage rests on differing materiality judgments — so the aggregate numbers, large as they are, are themselves uncertain.
The good news is that, for now, the risks aren’t an imminent solvency problem, but a set of timing and disclosure mismatches – a deferred depreciation wall, capex running ahead of monetization, leverage migrating into the supplier/private-credit layer, and classification judgments that make true capital intensity hard to compare across companies. For their part, hyperscalers have taken advantage of the current moment of market euphoria to raise as much capital as they can, aware that the window will eventually shut. The question is when sentiment turns, and when all these funding conduits are shut, will there be enough funding to sustain the AI revolution for the foreseeable future. Of course, this question becomes moot if demand shifts, and instead of buying the latest and greatest offering from Anthropic or OpenAI for a stratospheric number of tokens, consumers and enterprises turn to dirt-cheap alternatives from China which offer 90% of the performance of frontier models for a fraction of the cost, then all bets are off.
Much more in our full must-read (as the WSJ clearly found) report from June, as well as the “AI Ecosystem: Charting Recent Trends“ report available to pro subs.
