Nine technology companies owe about $3 trillion in AI commitments that haven’t been booked as debt. The headline wants to know whether off-balance-sheet AI spending is a bubble. A smaller buyer has a better question, which is what that fixed bill does to the vendor’s behaviour at your renewal.
Ask for a price cap while they still want your signature.
Key points
- What counts as off-balance-sheet AI spending? Promises to pay that haven’t shown up as debt yet.
- Is this a bubble story or a contract story? A bubble story ends with someone else’s losses. A contract story ends with your invoice.
- What would a renewal cap save a 100-person firm? None of the $3 trillion appears on your invoice, which lists seats, usage and overage.
- Do two independent tallies agree? Roughly. Axios reported on August 27 that a Morgan Stanley analysis of company filings found about $3 trillion of AI-related infrastructure commitments across seven companies, plus roughly $770 billion of debt and lease obligations already on balance sheets.
What counts as off-balance-sheet AI spending?
Promises to pay that haven’t shown up as debt yet. A Wall Street Journal analysis of securities filings counts about $3 trillion of them. According to investingLive’s summary of the WSJ piece, nine firms (Alphabet, Meta, Amazon, Microsoft, Oracle, Nvidia, Broadcom, SpaceX and AMD) carry it.
Roughly $1.2 trillion is lease obligations for data centres that haven’t started, about four times the level a year earlier. About $1.9 trillion is chip and hardware purchase commitments.
Two data points show the speed. Alphabet alone stood at $811 billion as of June, more than double the $332 billion reported three months earlier. The summary sets the whole $3 trillion against roughly $600 billion of conventional capital spending the nine firms reported over the past year. It also says Alphabet and Amazon have both posted negative free cash flow recently, with no figures given.
Every dollar is owed whether or not customers show up.
Is this a bubble story or a contract story?
A bubble story ends with someone else’s losses. A contract story ends with your invoice. The easy take is that $3 trillion is a warning for investors. The more useful one for you is that these vendors have fixed costs that don’t flex with demand, and fixed costs change how a seller treats a buyer.
Think of a hotel with a mortgage. A half-empty hotel discounts rooms, offers long-stay deals and pushes loyalty cards. A full hotel raises rates. Which one an AI vendor turns out to be depends on whether demand grows into the capacity, and nobody can say that today. That part is our hypothesis, not a forecast, and we aren’t trying to predict which.
Watch behaviour, and take the filings as background. The tell will be in small things, like a promotional discount that stops, a free tier that shrinks or a new minimum seat count. None of those looks like a price rise, and each is one.
Neither outcome is stable, though. A buyer who signs today’s price for three years is betting the first scenario lasts. A buyer on a monthly plan is betting the second never arrives. A contract term turns a bet into a known cost.
The sceptic says take-or-pay commitments are normal in infrastructure and that utilities, airlines and telecoms have run on them for decades. Buyers of electricity don’t read the utility’s lease schedule. Fair, and for a small firm buying a few seats it may be all the attention this deserves. Where we differ is speed. Lease obligations of $1.2 trillion at four times last year’s level mark an industry whose cost base is changing quickly, and a utility’s rates are set by a regulator while an AI vendor’s are set by whatever the quarter looks like.
Our June look at OpenAI building its own chip covers one way a supplier tries to lower that fixed cost, and our piece on token price cuts and the Uber budget explains why a cheaper token doesn’t shrink a total bill.
What would a renewal cap save a 100-person firm?
None of the $3 trillion appears on your invoice, which lists seats, usage and overage. The lease and the chip order are buried inside the price per token and the price per seat. That’s why the number matters to you only indirectly, and why the tools to manage it are contractual.
A worked illustration, with assumptions that are ours. A 100-person firm pays $40 per seat per month for an AI assistant, which is $4,000 a month or $48,000 a year. The vendor raises the list price by 10% at renewal. That’s $4,800 more a year, with no new feature to show for it. If your contract capped the increase at 3%, the same renewal costs you $1,440 more, and you keep $3,360. The numbers show the mechanism. They aren’t a prediction of any vendor’s behaviour.
Four terms are worth asking for, in this order. A renewal price cap, stated as a percentage or tied to a published index, in writing. A term you can live with, since twelve months with a renewal option usually beats three years at a small discount when the market is moving this fast. Data export in a usable format, so that moving vendors costs weeks and not months. And a second vendor you’ve already tested on one real task, which is your negotiating position. The cap matters most, because it’s the only one of the four that puts a number on the thing you’re afraid of.
See our breakdown of what AI really costs a Canadian business for the budget lines that sit around these terms, and cost per successful task for the number to track alongside them.
None of this makes the vendors bad actors. A seller with large fixed costs and a rival at its heels will sign nearly any buyer in the short run. That’s the window. Prices set in the window carry through the term, and you want them written in before the incentive changes.
Do two independent tallies agree?
Roughly. Axios reported on August 27 that a Morgan Stanley analysis of company filings found about $3 trillion of AI-related infrastructure commitments across seven companies, plus roughly $770 billion of debt and lease obligations already on balance sheets. Its split was $1.1 trillion of unstarted data centre leases and $1.7 trillion of chip, memory and networking purchase commitments.
The two tallies differ in who they count and how they slice it, so we treat both as estimates. They land on the same order of magnitude, and two counts that arrive there independently are more persuasive than either alone. Morgan Stanley’s researchers also noted that suppliers and developers can borrow against these commitments, which makes the total debt load hard to see. Axios adds that it’s unclear over what period the money will be spent, or whether all of it will be, since contracts can be renegotiated.
The WSJ article is paywalled, so our figures come from investingLive’s summary of it and from the later Axios piece. Neither shows the full list of contracts. Off-balance-sheet commitments aren’t the same as debt, and many are ordinary purchase contracts that firms with large profits can carry comfortably. Negative free cash flow at two firms in a heavy build year isn’t distress. We’re describing a risk to prices and terms, not a risk of failure, and we’d back off if AI prices stayed flat or fell through the next two renewal cycles. Then a cap costs you nothing but the time it took to ask.
Keep it short. State the percentage you want as the cap, name the term you’d accept and ask for export in writing. Vendors that want the renewal will say yes to most of it, and one that won’t has told you where you stand. Either way you’ll know before the commitments behind the price come due, not after.
Send that email before the vendor sends the renewal notice.
Frequently asked questions
What does off-balance-sheet AI spending mean?
It means contractual promises to pay for things like data centre leases and chip orders that haven’t been recorded as debt. The WSJ analysis counted about $3 trillion of them across nine tech firms, based on securities filings.
Will this make my AI subscription more expensive?
We can’t say. Fixed commitments give vendors a reason to fill capacity, which can mean discounts or price rises depending on demand. The safe move is a contract with a renewal cap and an exit.
How should a small business protect itself?
Ask for a written renewal price cap, a term of twelve months or less, usable data export, and keep a second vendor tested on one real task.
Archive entry dated August 18, 2026, written and fact-checked on October 8, 2026. Sources are linked on the claims they support.
