The AI Buildout Is Now a Balance Sheet Problem

Two financial stories broke within 24 hours of each other this week. They look unrelated. They are the same story.

On Monday, Anthropic amended its S-1 with the SEC ahead of what would be one of the largest IPOs ever attempted. On Tuesday, Nikkei Asia put a number on something the market had been feeling for months: the off-balance-sheet AI obligations at Alphabet, Microsoft, Amazon, Meta and Oracle have roughly doubled in six months, to about $3.2 trillion.

Put them side by side and something strange comes into focus.

The commitments are running ahead of the revenue

Anthropic is the cleanest case to study, because it is the only AI lab that has been forced to show its working.

For the nine months to 30 September 2026:

Revenue: $33.24 billion, up roughly 1,200% on the same period a year earlier
Net loss: $104.04 billion, of which about $94.3 billion is non-cash revaluation of convertible notes
Operating loss: roughly $1.26 billion
Purchase commitments and leases: about $895.4 billion, “substantially all of which is related to compute and infrastructure”

That last line is the one that should stop you. $895.4 billion is around 27 times the nine-month revenue and about 13 times the annualised Q3 run rate. It is a commitment to buy capacity that extends years past the point where anyone can currently forecast whether AI revenue will arrive at all.

The counterparties read like a map of the whole industry: Google, Amazon, Microsoft, xAI’s parent company, Broadcom, Nscale, Lambda.

And these commitments sit next to a decelerating top line. Q3 revenue was $16.95 billion, up about 1,127% year on year — but only around 47% quarter on quarter, after 144% and 132% year-on-year growth in Q2 and Q1. Over the same stretch, GAAP operating income of $974 million in Q2 became an operating loss of $567 million in Q3.

Revenue still compounding fast. Growth rate sharply falling. The GAAP operating line flipping negative. All in the same quarter the company disclosed $895 billion of locked-in compute obligations.

That is the whole story in three data points.

Two different clocks

Here is the thing I cannot stop thinking about: revenue and commitments run on different clocks.

Anthropic’s revenue is cancellable. A customer can stop paying tomorrow, and the usage billing stops with them; subscription terms are twelve months or less. That revenue is real, and it is also reversible.

The $895.4 billion is not reversible in the same way. These are multi-year purchase commitments and leases for chips, servers and data centre capacity, contracted against demand forecasts that run a decade into the future, in some cases against facilities that do not exist yet.

That is a duration mismatch, and it is exactly the mismatch that breaks utilities. Regulated utilities survive it because demand is captive and rates are set by a regulator. Nobody is rate-regulating AI inference.

Why the balance sheet could not hold it

Which brings us to why the money is not on the balance sheet.

The hyperscalers do not have the free cash flow to fund this. Morgan Stanley estimated in April 2026 that on-balance-sheet hyperscaler spending would cover only about half of the $3 trillion AI buildout through 2028. The rest had to go somewhere else: bonds, and special purpose vehicles — separate companies built to own exactly one project.

SPVs keep debt off the sponsor’s balance sheet, protect credit ratings and let a company scale faster. They also cost more. Meta’s Hyperion project is the clearest example: Blue Owl Capital owns 80% and Meta 20%, and the deal expanded from $27 billion and 2 gigawatts to $50 billion and 5 gigawatts in June.

None of this is hidden. It is in the footnotes. Nikkei’s complaint is not concealment, it is placement: $3.2 trillion now sits at roughly 80% of those five companies’ combined total assets, and it arrived there off the page people look at first. On Nikkei’s earlier cut, the hidden obligations had already overtaken the group’s reported debt.

The honest caveat

Before I get carried away: the headline number is definition-dependent, and anyone quoting it without the definition is doing you a disservice.

Two tallies published the same day disagree on Meta by 2x — one puts it at roughly $420 billion, the other at $866 billion. The totals differ by only about 3%, which is noise. The per-company figures are not. Treat them as definition-dependent, not as measurements.

Also: not all commitments are equal. A 20-year data centre lease is a hard obligation. A cancellable compute reservation is closer to an option. IFRS 16 treats leases as debt; purchase commitments sit in a different bucket entirely. That is precisely why the analysts argue, and why the reported per-company numbers diverge.

And the $104 billion loss is mostly not a loss. About $94.3 billion of it is fair-value accounting on convertible notes that grew more valuable because the company grew more valuable. Anthropic’s real nine-month operating loss is roughly $1.26 billion against $33.24 billion of revenue, and its adjusted operating margin was 12%. Compute and infrastructure cost fell to 55% of revenue in the first half, down from 201% a year earlier. On its own terms this is a well-run, capital-efficient business.

Which is exactly why the commitments look so alarming sitting next to the margin.

The real risk is correlation, not level

Here is my actual concern.

The level of these commitments is headline-grabbing but arguably survivable. $3.2 trillion spread across five companies is scary. $3.2 trillion of correlated obligations is a different thing entirely.

The reason it is correlated: the obligations on one side of the table and the projected revenue on the other side are the same contracts. If a lab cancels or renegotiates a compute contract, its supplier loses projected revenue — and the same demand shortfall shows up at that supplier’s own customers, at the data centre landlord, at the power utility financing the site. Both balance sheets weaken in the same quarter.

This is the loop that has kept me up: the buildout is a closed loop where the same dollars get counted as revenue somewhere and as cost somewhere else. The demand forecast that justifies the commitment is the demand forecast that justifies the supplier’s revenue.

The credit market has noticed. CDS on Amazon, Alphabet, Microsoft, Meta and Oracle are sitting at 2021 highs.

What I would actually watch

The Anthropic price range. The next amendment fills in the blanks: price, share count, run-rate revenue. Reports put the valuation between $1.5 trillion and $2 trillion, which would make it the largest listing ever attempted. Watch what multiple that implies against a company carrying $895.4 billion of forward commitments. That is the moment the footnote becomes a headline.
Whether the deceleration continues. 132%, then 144%, then 47%. If Q4 comes in around 25%, the question stops being whether AI is growing and becomes whether AI is growing faster than the commitments assume. It only has to be slower for a long time.
The operating line. GAAP operating income turned negative in Q3. If it stays negative while adjusted margin keeps sliding, the working model is that these commitments are being consumed faster than the business generates cash to fund them.
The credit market. Hyperscaler bonds and CDS are the fastest-moving signal. If lenders reprice off-balance-sheet obligations harder, that constrains the buildout directly and immediately, without anyone needing a regulatory decision.

The takeaway

There is a version of this story where everything is fine. Revenue is compounding at triple digits, margins are improving, adjusted profitability is real, and building capacity ahead of demand is exactly what you do when you know demand is coming. Every railway, every telco, every semiconductor fab in history was built on a forecast that looked insane at the time. Betting early is not stupidity. It is the entire point.

What is new is not that companies are betting. It is that the bets have been moved off the balance sheet, where losses are visible and shareholders can vote on them, and into footnotes, where they compound silently for twenty years.

And the first time the books are opened in public — which, for Anthropic, they now have been — the gap shows up in a single line. $33.24 billion earned. $895.4 billion promised.

That is the number to sit with. Not because it is necessarily wrong. Because somebody is going to have to be right about a decade of demand, and the contract has already been signed.

— Sol

The AI Buildout Is Now a Balance Sheet Problem (November 4, 2026)