Opinion · AI and the State of Humanity, Part 2 of 3  ·  12 November 2026

The Demand Is Real. The Cash Is Borrowed.

On negative free cash flow, record borrowing, a $105 billion guarantee in a footnote, and who ends up holding the risk if the sums do not add up.
By Alan Wright  ·  The Haunted Lighthouse Limited  ·  Peel, Isle of Man

Last week was about who holds the lever. This week is about who pays for it.

The story the industry tells about money is simple. Demand for AI is enormous, it outstrips supply, and the companies building the infrastructure are rich enough to pay for it themselves. The first two parts of that story are broadly supported by the evidence. The third part stopped being true this year, and the filings say so in plain numbers for anyone willing to read past the press release headline.

This piece does not predict a crash. I do not know whether one is coming, and neither does anyone selling you a newsletter that says otherwise. What it does is set the claim alongside the cash flow statement, and ask who ends up holding the risk if the sums do not add up.

The rule for the series still stands: every figure is checked against a primary source where one exists, and anything that is not gets labelled.

The claim

Microsoft's chief executive told investors in October that customer demand "continues to exceed available supply". Amazon's chief executive said the company is installing capacity about as fast as it can and that customers are "gobbling it up". Alphabet's cloud business more than doubled its revenue year on year.

None of that is theatre. Google Cloud revenue rose 103 per cent to $30.7 billion in the quarter to September. Azure grew 48 per cent. AWS grew 39 per cent, its fastest rate in nearly five years. If you are looking for evidence that nobody wants this technology, you will not find it in these numbers.

The theatre is somewhere else. It is in the assumption, repeated in almost every earnings call, that the spending is a sign of strength rather than a bet that has to be financed.

What the cash flow statements say

Free cash flow is what is left of the cash a business generates after it has paid for its buildings and equipment. It is the money that funds dividends, buybacks and, until recently, the AI buildout itself.

Here is the quarter to 30 September 2026, from the companies' own results releases:

Capex in the quarter Free cash flow 2026 capex guidance
Alphabet $51.4bn negative $9.1bn $195bn to $205bn
Amazon not like for like (see note) negative $42.4bn over 12 months about $220bn
Meta $34.9bn negative $2.1bn $135bn to $145bn
Microsoft $45.2bn (incl. finance leases) positive $18.6bn about $170bn

Three of the four burned more cash than they generated. Reuters described Meta's quarter as its first cash burn in more than a decade. Alphabet has now been negative for two quarters running. Amazon's trailing twelve months went from an inflow of $14.8 billion a year earlier to an outflow of $42.4 billion, which Amazon itself attributes primarily to investment in AI.

Two fair caveats. Alphabet's figure includes a $5.2 billion payment on a 2018 European Commission fine and a build-up of inventory for selling its TPU chips to customers, so not all of that outflow is data centres. And the four companies do not measure capex the same way: Meta includes finance lease principal, Microsoft quotes a figure including finance leases, Amazon reports net of incentives. The guidance figures add up to roughly $720 billion to $740 billion for 2026, but that total is my arithmetic across four different definitions, and should be read as an order of magnitude rather than a precise sum.

Microsoft is the exception, and it is worth looking at why.

Capex that falls without spending falling

In April, Microsoft expected to spend about $190 billion on capital in calendar 2026. In July that became about $175 billion. In October it became about $170 billion.

Read only the headlines and that looks like discipline. Read the calls and it is something else.

The July reduction came from an accounting change. Microsoft extended the estimated useful life of its data centres from 15 to 25 years, and as a result more of its future data centre leases will be classed as operating leases rather than finance leases. Its chief financial officer, Amy Hood, explained the effect directly: finance leases are counted in capital expenditure, operating leases are not. She also said that, outside this change, Microsoft's investment expectations for the year were unchanged.

The October reduction, about $5 billion, she tied to third-party capacity arriving later than expected, with spending expected to accelerate once it lands. That is timing, not restraint.

None of this is improper. Lease accounting is governed by standards, and Microsoft disclosed the change openly. But it shows something the Theatre Pulldown framework is built to catch: the headline number moved, and the activity underneath it did not. Microsoft's July filing showed $329.1 billion of data centre leases that had signed but not yet started, running from fiscal 2027 to 2033. Those obligations are real whether they sit inside the capex figure or outside it.

The bond market fills the gap

When the cash runs short, the money has to come from somewhere. This year it came from the debt and equity markets, in sizes nobody has seen before.

According to LSEG data reported by Reuters on 6 November, US investment-grade corporate bond issuance reached $1.88 trillion in 2026 through 4 November, beating the previous annual record of $1.86 trillion set in 2020. Amazon, Alphabet, Meta, Microsoft and Oracle accounted for more than $370.7 billion of that, roughly a fifth.

Some of the individual transactions:

Again, the caveat. These are among the strongest balance sheets in corporate history, the order books were full, and borrowing to invest when you can borrow cheaply is not reckless in itself. Microsoft's leverage after the deal remains modest by any normal standard.

The point is narrower. The story that the buildout is self-funded is no longer the story the cash flow statements tell. One bank cited by Reuters expects bonds to fund about 35 per cent of hyperscaler capital expenditure in 2027. The report did not name the bank, so treat that figure as indicative.

Guarantees: the off-balance-sheet fact

The most interesting money in this story does not appear on anyone's balance sheet at all.

On 17 August Nvidia filed an 8-K disclosing residual value guarantees, capped at $105 billion in aggregate, supporting leases at a data centre campus in Pike County, Ohio. SoftBank's SB Energy will build, own and operate the site. OpenAI is the tenant, on a 20-year lease. Nvidia will be the exclusive supplier of the AI compute inside it, and is separately putting $1.5 billion of equity into SB Energy.

Follow the loop. Nvidia guarantees a minimum value on the facilities. That guarantee helps SB Energy borrow to build them. OpenAI leases them and fills them with Nvidia hardware. The guarantor's revenue depends on the tenant it is guaranteeing, and the collateral's value depends on continued demand for the guarantor's own product.

Nvidia's CFO commentary for the quarter to July puts its total maximum gross exposure under guarantees at $108.5 billion, against $56.6 billion of cash, cash equivalents and marketable debt securities.

That comparison needs handling carefully, and most of the coverage has not. Maximum gross exposure is not expected loss. The guarantees become effective as each lease starts, with the first expected in fiscal 2029. Nvidia says its exposure declines as OpenAI pays its rent. Nvidia pays only the gap between a guaranteed minimum value and whatever SB Energy recovers by re-letting or selling the facilities, and the $105 billion cap is reached only if recovery is close to nothing. The guaranteed minimum values themselves have not been published.

So the honest summary is this: the largest single piece of credit support in the AI buildout is disclosed in a footnote, its realistic downside depends on numbers nobody outside the deal can see, and its trigger is the credit of a single company. Those are exactly the conditions under which an auditor should ask to see the schedule.

Backlog is not cash

The answer to all of this, from the companies, is backlog. Alphabet's cloud backlog stood at about $561 billion at the end of September. Microsoft and Amazon point to similar contracted commitments.

Backlog is real and it matters. But Alphabet itself expects to recognise only a little over half of that backlog as revenue within the next 24 months. The rest sits further out, where contracts can be renegotiated, delayed or not renewed. And Alphabet's finance chief described a newer pattern on the call: a TPU deal goes into backlog when it is signed, Google builds the systems into inventory, and the cash arrives only on delivery. The money goes out before it comes in.

There is also the question of who the backlog is owed by. A meaningful share of the demand at the top of this market comes from a small number of AI labs, and those labs are themselves financed by the hyperscalers, by Nvidia, and by investors betting on an IPO. OpenAI told investors in October that its annualised revenue was approaching $50 billion. It has also said it is targeting around $600 billion of compute spending through 2030. Annualised revenue is the latest month multiplied by twelve; it measures pace, not money in the bank.

I have not been able to find a reliable public breakdown of how much of the hyperscalers' backlog traces back to a handful of AI labs. That absence is itself worth noting.

Who holds the tail risk

If this ends well, the shareholders of five or six very large companies do very well. If it ends badly, the losses do not stay with them.

Investment-grade bond indices are bought mechanically by passive funds. When a handful of issuers sell a fifth of a record year's supply, every pension scheme, insurer and savings product that tracks those indices ends up holding more of that debt, whether anyone chose to or not. Project finance backed by guarantees moves risk to lenders, and from lenders to whoever holds their paper.

For a small jurisdiction this is not abstract. The Isle of Man's financial sector includes life companies, pension providers and fund administrators who hold global investment-grade bonds as a matter of course. I am not claiming any particular firm is over-exposed; I have no evidence of that. The question is simpler: does anyone here know how much AI infrastructure credit sits inside the island's savings, and would they find out before or after the market did?

That is the sovereignty argument from Part 1 again, in financial form. Dependency is not only about whose software runs your services. It is also about whose debt sits in your pension.

What this is not

It is not a prediction that the AI boom is a bubble about to burst. The revenue growth is real, the companies are profitable, and negative free cash flow during a heavy investment phase is a choice, not a crisis. Railways, telecoms and electricity grids were all built this way, and some of those builds paid off.

It is also worth remembering that some of those builds wiped out the people who financed them while leaving behind infrastructure everyone else used for a century. Both things can be true.

What it is, is a gap between the claim and the evidence. The claim is a self-funded buildout by companies too rich to need anyone else's money. The evidence is record borrowing, accounting choices that move headline numbers, guarantees that sit in footnotes, and backlog that converts slowly.

Five dull things I would ask for

  1. One capex definition. A common measure across the major providers, including finance and operating leases, so that a figure can fall only when spending does.
  2. Signed-but-not-started obligations in the headline. If a company has hundreds of billions of leases waiting to commence, that number belongs next to the capex guidance, not in a note to the accounts.
  3. Guarantee schedules, not just caps. Publish the guaranteed minimum values, so the market can price the realistic exposure rather than the theoretical maximum.
  4. Backlog by counterparty concentration. Not names, if that is commercially sensitive, but how much of the backlog comes from the top five customers.
  5. Look-through for savers. Pension and fund providers, including those on this island, telling members how much of their bond allocation sits with a handful of AI-infrastructure issuers.

None of this is radical. Most of it is ordinary disclosure discipline applied to an unusual amount of money.

Next week

Part 3, on 19 November, is the environment. It picks up exactly where this piece ends. The Department of Energy has said the generation planned for that Ohio campus is at least 9.2 gigawatts of natural gas out of 10. The money is only half of what the grid has to carry.

Can we afford the AI or not?

Part 3: The Environment, Thursday 19 November.


Sources

This is an opinion piece based on public filings and published reporting. Figures are checked against primary sources where one exists, and anything that is not is labelled in the text.

Questions about this analysis, or interested in working with The Haunted Lighthouse?
consultancy@haunted.lighthouse.co.im

The Sovereign Auditor covers digital sovereignty, cybersecurity governance, and data protection policy, with particular focus on Isle of Man jurisdiction and Crown Dependency issues.

Support independent analysis. Subscribe directly, or scan on your phone.

Payments via PayPal. Credentials delivered by email. No Substack. No Stripe. No middlemen.