The correction first: a week ago I wrote that Anthropic had compute commitments of 540 to 570 billion dollars, per Fortune's estimate. The prospectus itself did not contain the number then; there was a blank. Since 2 November it is filled in: 895.4 billion. And on the same day I read that, t3n headlines that the five big tech companies carry 3.2 trillion dollars of hidden AI debt. Both numbers are right, I checked them. Neither is debt. And that is exactly where it gets interesting, because the difference between a commitment and a debt is not hair-splitting but the whole question of where the risk sits.
The correction first
Last week's article celebrated a series: in 2023 a dollar of compute brought 0.63 dollars of revenue, in 2025 it was 1.82, in the second quarter of 2026 then 2.08. Less compute cost per revenue, quarter after quarter. The new prospectus adds the third quarter, and the series bends: 16.96 billion of revenue against 9.86 billion of compute and infrastructure cost, that is 1.72. The share of compute in revenue rises again, from 48 to 58 percent. I described the curve as a trend. One quarter later it shows that it is not one, but a number that swings with purchasing. That belongs here, visibly, as I promised.
The third quarter: the curve bends
| Q2 2026 | Q3 2026 | |
|---|---|---|
| Revenue | 11.6 bn | 17.0 bn |
| Compute and infrastructure | 5.6 bn (48%) | 9.9 bn (58%) |
| Revenue per compute dollar | 2.08 | 1.72 |
| Operating result | +974 m | −567 m |
| Adjusted operating result | 2.2 bn (19%) | 2.5 bn (15%) |
| Growth on previous quarter | 144% | 47% |
The explanation is in the prospectus itself, and for me it is the most important sentence of this whole follow-up: demand grew faster than the firmly contracted capacity, so Anthropic bought roughly five billion dollars of short-term compute that "currently costs significantly more" than the multi-year agreements. On top came a non-cash charge of two billion from an employee donation matching programme, and about 600 million of revenue given up through price cuts and discounts. Adjusted for the special items, profit in dollars actually rose, but the margin fell. For nine months the figures are now 33.2 billion of revenue and 1.26 billion operating loss. The net loss of 104 billion is again almost entirely the accounting entry from the revaluation of convertible notes that I explained last week; not a dollar of it has flowed.
$895 billion, taken apart
Now the big number. 895.4 billion dollars of purchase and lease commitments as of 30 September, "substantially all" for compute and infrastructure. That is not an annual amount and not a debt. It is the sum of all contracts over their term, and the term runs to 2045. This is how the prospectus splits it:
| Item | bn USD |
|---|---|
| Total commitments (as of 30 Sept, incl. those entered after) | 895.4 |
| of which due in the next 12 months | 69.8 |
| of which by end of 2030 | 447.5 |
| of which 2031 to 2045 | remainder, about 448 |
| of which with termination rights or open conditions | 213.8 |
| Additionally, no fixed amount, estimated (mainly Google TPUs 2028 to 2031) | 150 to 175 |
| Actual financial debt | 6.2 |
| Cash and short-term investments | 131.9 |
| Liquidity incl. credit line | 146.6 |
The largest items are named: at least 121.3 billion to Google through July 2033, at least 110 billion to Amazon through April 2036, at least 44.9 billion to Microsoft through June 2033, and about 161 billion of rent under five-year leases with Broadcom and financing entities for its own chips. Google and Amazon have true-up clauses if less is consumed than agreed. The contracts are, with few exceptions, non-cancellable.
And this is the point where "debt" is the wrong word and "harmless" is the other wrong word. Financial debt at Anthropic is 6.2 billion, almost nothing at this scale. But 69.8 billion falls due in the next twelve months, and that is roughly a full year of revenue at today's pace. It gets paid from the cash pile, not from operations: operating cash flow for the first nine months was 6.6 billion, financing brought in 110.7 billion. The prospectus itself says the funds are sufficient for at least twelve months; after that it depends on revenue growth, operating cash flow and further capital. That is not a warning from me, that is the text.
The asymmetry
What really occupies me in this prospectus sits in the risk factors, in a single thought: many of the contracts are minimum purchases, payable regardless of utilisation. Revenue, on the other hand, is mostly consumption-based, and a large part of it without a long-term contract. Fixed on one side, variable on the other. As long as usage grows, that is good business, because fixed capacity is cheaper than short-term top-up, see the third quarter. As soon as usage stalls or prices per unit of compute fall, the fixed payments keep running, and the variable income does not.
That is the same structure I described last week as a bet, only with numbers. The bet is not whether AI works. The bet is whether usage grows as fast as the contracts assume, and whether the price per unit of compute does not fall faster than the cost of the contracted capacity.
$3.2 trillion: hidden is the wrong word
And now t3n. The number comes from Nikkei Asia, confirmed in scale by Goldman Sachs: Alphabet, Microsoft, Amazon, Meta and Oracle together carry about 3.2 trillion dollars of obligations that are not on the balance sheet, up 90 percent since March. Per Nikkei, 60 percent purchase commitments, 40 percent leases not yet commenced; the sum equals 80 percent of the five companies' assets. Alphabet alone 1.05 trillion, plus 160 percent, with Nikkei itself noting that Alphabet changed its calculation method and the comparison therefore limps.
Goldman names different components: 1.7 trillion in lease commitments, three quarters of which have not commenced, and 1.9 trillion in other purchase commitments, 685 billion of which fall due by the end of 2027. Add those and you get 3.6, not 3.2. So I do not add them. The scopes are not the same, and part of the leases is already on the balance sheet. The order of magnitude "around three trillion" is confirmed several times over. The decimal is not.
But "hidden" is wrong. These numbers sit in the notes to the quarterly reports; that is where Nikkei and Goldman got them. They are just not on the balance sheet, and there is a reason for that which is not trickery: a lease goes on the balance sheet when the data centre is handed over, not when the contract is signed. A purchase commitment is a promise to pay in future, not a loan taken. Economically, both can weigh exactly as heavily as a loan if the income fails to arrive. On the balance sheet and for default risk, it is not the same thing. Whoever calls the 3.2 trillion "debt" turns a supply chain into a credit crisis. Whoever waves it away overlooks that the payments come whether the usage comes or not.
What the credit market says
Here it gets more concrete than any headline. The price of insurance against default, the CDS spread, distinguishes very precisely between the five: Microsoft around 59 basis points, Alphabet around 92, Oracle around 294, as of late October per Axios. Oracle thus pays five times what Microsoft pays for the same insurance. At the same time all of them are raising money: Microsoft 25 billion in the bond market, its first US issue since 2017, Oracle 10 billion of bonds plus almost 10 billion of shares on the same day, Alphabet 6 billion euros. Goldman's analyst puts it plainly: debt financing will play "a key role" in the AI build-out, and the ratings are so high that even several downgrades would not restrict market access.
That is the difference the 3.2 trillion blurs. Microsoft can service its commitments from cloud, Office and Windows if the AI revenue fails to arrive. Oracle builds capacity with borrowed money for a few large customers and has to hope they pay. A specialised GPU provider like CoreWeave hangs even more directly on utilisation, hardware value and refinancing. Same number, three kinds of risk. The market prices that. The headline does not.
The risk no headline mentions
And now what I miss in all the coverage, and what I know from the workshop: hardware gets better faster than the contracts expire. A data centre leased today for ten years will, in five years, be computing on chips two or three generations old. Each new generation delivers the same compute with less power at a lower price, and thereby lowers the market price for compute overall. The old facilities do not become worthless; for inference, smaller models, training environments they remain usable. But the price at which they can still be rented out is set by the new generation, not by the old contract.
An invented example, only to show the scale: ten billion of investment, four years of hardware life, makes 2.5 billion of depreciation per year. Add financing at six percent, 0.6 billion, and power, staff, maintenance, say 0.7. The data centre has to bring in 3.8 billion a year just to break even, before tax, with no profit. If the market price for compute falls by a third because the next chip generation has arrived, 1.3 billion a year is missing at the same utilisation. The commitment keeps running. That is the risk behind the 895 billion and behind the 3.2 trillion: not the sum, but the term against the pace of the technology.
Why this occupies me
Not as an investor, I am not one and advise none. As someone who buys compute and builds a training environment. For us, each of these commitments is good news: whoever has ordered 895 billion of capacity has to fill it, and that means the price per token keeps falling; the 600 million of discounts in the third quarter are the beginning of that. But our customers' question remains last week's: what does it cost to solve a task completely, securely, integrated, checked, maintainable? And something else comes in that has directly to do with the commitments: the better the small, specialised models get, the more inference runs locally, the less of the giant central capacity is needed at the prices it is being calculated with today. We build in exactly that direction. Not because we bet against the big players, but because it is cheaper and safer for our customers. That it also changes the big players' arithmetic is a side effect I did not order, but do observe.
What I watch
Free cash flow after investment. What is left after building. At Anthropic: 6.6 billion operating against 69.8 billion due next year. Growth has to close that gap, or the cash pile does.
New debt and CDS spreads. Not the sum of the commitments, but what the market charges for insurance. Oracle at 294 is the signal, not the 3.2 trillion.
Price per unit of compute. If it falls faster than the cost of contracted capacity, the asymmetry tips. You see that in the discounts before you see it in a balance sheet.
Terms against chip generations. 447 billion by 2030, 448 after. Which chips will be computing in those contracts in 2035, nobody knows today.
Independent revenue. How much demand comes from customers outside the loop in which investors supply capital that flows back as compute? No prospectus answers that question. It is the only one that counts.
What remains
I said last week that the number that matters is in no prospectus. That still holds. What the new prospectus contains is the size of the bet, and it is now quantified: 895 billion firm, up to 175 on top, 6 billion of real debt, 132 billion in cash, a year of runway by its own account. At the big five the bet is around three trillion, spread across companies that carry it very differently well. That is neither a crisis nor an all-clear. It is a body of contracts that assumes usage that is not yet there.
Commitments are not debt, and none of it is hidden. But a commitment payable regardless of usage, against revenue that only comes with usage, is a bet on pace. The 895 billion and the 3.2 trillion are the stakes. The odds are in the CDS spread, not in the headline. And whether the bet pays off is decided not by what AI can do, but by whether the industry finances more capacity than it can sell at future prices. Best knowledge, as of today. Last week's correction is at the top. The next one comes as soon as it is needed.
Sources and status
Anthropic, Form S-1/A, filed 2 November 2026. sec.gov. Q3 revenue 16.956 bn, compute and infrastructure 9.863 bn (+382%), operating loss 567 m, adjusted 2.463 bn (15%), nine months 33.242 bn revenue, 1.260 bn operating loss, 104.0 bn net loss (94.3 bn of it revaluation), commitments 895.4 bn with breakdown, financial debt 6.2 bn, cash 95.6 plus 36.3 bn investments, liquidity 146.6 bn, operating cash flow 6.569 bn, financing 110.735 bn. Traced via Filing Insights (links every figure to its passage in the document), Unite.AI (detail table of commitments, counterparties, 150 to 175 bn additional) and Data Studios (5 bn short-term top-up, 600 m price effect).
t3n, 4 November 2026. 3,2 Billionen Dollar: Versteckte KI-Schulden der grossen Tech-Konzerne in 6 Monaten fast verdoppelt, after Nikkei Asia: 90 percent since March, 60/40 purchase/lease, 80 percent of assets, Alphabet 1.05 trillion with changed method.
Goldman Sachs, quoted at Yahoo Finance, 4 November 2026. Big Tech has more than $3 trillion in AI commitments that don't show up on its balance sheets: 1.7 trillion leases (three quarters not commenced), 1.9 trillion purchase commitments (685 bn by end 2027), bond sales by Microsoft, Oracle, Alphabet, quote on the role of debt.
Axios, 3 November 2026. Bond investors have cooled to AI debt: 5-year CDS Microsoft 58.9, Alphabet 91.6, Oracle 294.4 basis points. Other sources put Oracle between 288 and 300 depending on the day.
My article from last week. The Anthropic prospectus: AI gets more efficient. But who knows the real costs?, now with an addendum.
Status and caveat. 5 November 2026. I am not a financial analyst and advise no one; this is reading, recalculating and opinion, and the opinion is marked as such. I have not read the full S-1/A line by line, but checked the quoted passages through three independent analyses that agree. Goldman's components add to 3.6 trillion, the headline says 3.2; I take the order of magnitude, not the decimal. The data centre example is invented. Commitments are neither expenses nor debt. Best knowledge and best information as of today; corrections are added here visibly, as the one to last week is above.
Written by me. The thoughts, the values, the learnings, the mistakes: all mine. Grammar and spelling are corrected by a twin model of my own, trained on my texts. Sometimes that turns into a mix. Of course I read everything again, but nobody is infallible. Whatever stays in is mine too.