Anthropic had approximately $451.2 billion of signed leases that had not yet started at 30 September, against $443 million spent on property and equipment in the third quarter: about $1,018 of unstarted leases for every $1 of quarterly capex, by our arithmetic from its amended Form S-1 filed with the SEC on 2 November and its first S-1 of 26 October.
At Meta and Microsoft the same ratio is about 10.5 and 9.3, and at Alphabet and Amazon about 2.5 and 2.1, by our arithmetic on their quarterly filings. Global Data Center Hub spotted the gap first, putting the first S-1’s $482 million of capex next to $321.4 billion of signed leases in a post on 3 November.
The first S-1, filed on 26 October, covered the six months to 30 June. Its cash flow statement shows $482 million of purchases of property and equipment for the period, and its lease note puts leases not yet commenced at approximately $321.4 billion at 30 June, up from approximately $50.2 billion at 31 December 2025.
The S-1/A of 2 November covers nine months. It shows $925 million of purchases of property and equipment for the nine months to 30 September, and puts leases not yet commenced at approximately $451.2 billion. The note says these obligations are “primarily related to compute and infrastructure capacity” and that, generally, they will commence between 2026 and 2028, with terms from less than one year to 20 years.
By our arithmetic, that leaves $443 million of capex in the July-to-September quarter, while the unstarted-lease line grew by $129.8 billion in the same three months, about 293 times the quarter’s capex. The line is about nine times its size at the end of 2025.
Leases not yet commenced at 30 September 2026 per $1 of July-to-September 2026 capex, for Anthropic, Meta, Microsoft, Alphabet and Amazon. Anthropic from its S-1/A and S-1 (third-quarter capex = nine-month $925 million minus first-half $482 million); the others from their 10-Qs and our Big Tech AI capex tracker. All ratios are our arithmetic.
Business Pill · Capex and Depreciation
An evergreen Business Engineer explainer on capital spending and depreciation. It explains the general idea only and does not discuss Anthropic or its filings.
The key insight: As we read it, Anthropic’s amended filing shows a compute build that sits mostly in signed leases rather than in its own equipment: approximately $451.2 billion of leases not yet commenced at 30 September against $443 million of third-quarter capex, about $1,018 per $1, by our arithmetic. At Meta, Microsoft, Alphabet and Amazon the same ratio runs from about 2 to about 10.
The Same Ratio at the Hyperscalers
The big cloud builders also carry large unstarted leases, but they spend far more on their own equipment. Their 10-Qs for the quarter to 30 September put leases not yet commenced at approximately $355.73 billion at Meta, $364.6 billion at Microsoft, $130.5 billion at Alphabet and $164.92 billion at Amazon, as compiled in Anthropic’s $895B of Commitments Rival Meta’s, Filings Show.
Their third-quarter capex was $33.9 billion at Meta, $39.4 billion at Microsoft, $51.4 billion at Alphabet and $77.3 billion at Amazon, according to our Big Tech AI capex tracker built from their SEC filings. That gives about 10.5, 9.3, 2.5 and 2.1 dollars of unstarted leases per dollar of quarterly capex, by our arithmetic, against about 1,018 at Anthropic.
Measured against operating cash, Anthropic’s capex was about 16.8% of the $2,640 million its operations generated in the third quarter, by our arithmetic from the two filings. In the same calendar quarter the five largest builders together spent 116% of their operating cash flow on capex, and Apple 8.0%, as reported in Apple Returned $28B as Capex Took 8% of Cash; AI Giants 116%.
Why the Capex Line Is Small
The filing describes the model behind the gap. Anthropic is “shifting from sourcing compute solely through fixed-term cloud agreements to a hybrid model that also includes leasing chips directly and hosting them in dedicated facilities operated on our behalf”, which it calls Anthropic Direct.
It says Anthropic Direct “represents a more flexible use of capital by allowing us to separate our commitments for data center capacity from our commitments for the compute infrastructure deployed inside them”, and that the approach “enables us to secure data center capacity early while deferring larger accelerator investments until our specific compute needs are more certain”.
The filing also says the mix is changing: “we expect an increasing portion of our compute and infrastructure investment to take the form of capital expenditures and other commitments, including lease obligations, that require significant upfront payments well in advance of the related compute becoming operational”. The nine-month cash flow statement already shows a separate line of $1,955 million of prepayments for leased assets and equipment, which is not counted in the $925 million. Including it, capex and prepayments came to about 43.8% of nine-month operating cash flow of $6,569 million, by our arithmetic.
How the Numbers Fit, Without Double Counting
The $451.2 billion is one part of the approximately $895.4 billion of aggregate committed spend the S-1/A reports, not an addition to it. By our arithmetic, the filing’s pieces are the contractual obligations table of $319,993 million, approximately $17.4 billion of short-term commitments, the $451.2 billion of unstarted leases, approximately $100.2 billion of commitments entered into after 30 September and approximately $6.6 billion of later short-term agreements, which add up to about $895.4 billion.
The table itself lists only commenced leases: $4,224 million of operating leases and $1,950 million of finance leases, undiscounted. The $321.4 billion and $451.2 billion are two readings of the same line at different dates, so they should not be added together either.
The lease note adds a condition: some of the unstarted leases are “subject to either exercisable termination rights or contingent conditions prior to commencement, such as the counterparty’s ability to secure financing, facilities, or regulatory approvals, as applicable”. “If those conditions are not met, the arrangement may be terminated”.
What Global Data Center Hub Argued
The post opened: “Anthropic pays for most of its AI infrastructure through contracts”. It argued that the capacity “sits on other companies’ balance sheets” and that “The tenant’s promise to pay is what lets those companies finance it”.
It compared the model to a store chain that signs a 20-year lease before the mall is built, writing that “The store chain’s own balance sheet stays light”. Its figures came from the 26 October S-1; the amended filing raises both, to $925 million of nine-month capex and $451.2 billion of unstarted leases.
As we read the two filings, the gap widened in the third quarter: the unstarted-lease line grew by $129.8 billion while capex was $443 million, so the lease line grew about 293 times faster than spending on owned equipment, by our arithmetic.
As we read it, the filing’s own description of Anthropic Direct explains the shape: data center capacity is committed early through leases, while the larger accelerator spending is deferred. The $1,955 million of prepayments for leased assets and equipment sits outside the capex line, and with it the nine-month share of operating cash rises from about 14.1% to about 43.8%, by our arithmetic.
Anthropic PBC, Form S-1/A, 2 Nov 2026
“Because data centers require longer lead times but represent a smaller fraction of total compute costs, this approach enables us to secure data center capacity early while deferring larger accelerator investments until our specific compute needs are more certain”
Three Implications
FOR DATA CENTER LANDLORDS AND LENDERS The S-1/A puts approximately $451.2 billion of signed leases, primarily for compute and infrastructure capacity, ahead of commencement between 2026 and 2028, against $3,061 million of net property and equipment at 30 September.
FOR READING THE $895.4 BILLION The unstarted leases are a component of the approximately $895.4 billion aggregate, alongside the $319,993 million contractual obligations table and the commitments signed after 30 September, so the two figures are not added together, by our arithmetic.
FOR THE COMPARISON WITH BIG TECH Anthropic’s third-quarter capex was about 16.8% of its operating cash, by our arithmetic, while the five largest builders spent 116% of theirs, as reported in our Apple piece.
The framework reads: “the capex race is no longer a corporate spending cycle. It is a sovereign-scale financing problem dressed in tech-company clothing”.
As we read it, Anthropic’s filings show one form that financing takes: the buildings and much of the equipment are committed through leases that other parties finance, while Anthropic’s own capex stays in the hundreds of millions a quarter. The framework was published in May 2026 and does not discuss the S-1/A.
The ratio compares a stock of future lease payments with one quarter of spending; it is a measure of mix, not of cost or risk. The filings do not say how much of the $451.2 billion is subject to termination rights or contingent conditions. Anthropic’s third-quarter capex is our subtraction of two filings, and the hyperscaler capex figures come from our tracker rather than from each 10-Q line we re-read today. We did not contact Anthropic or Global Data Center Hub.
This Week In The AI Supercycle
Anthropic’s Reality Check, The Capital Squeeze, & The New World Order
This week’s issue: for two years the AI build-out was priced as a growthstory. This week it started being priced as a credit story.
Anthropic’s amended IPO filing shows approximately $451.2 billion of signed, unstarted leases at 30 September against $443 million of third-quarter capex, about $1,018 per $1, by our arithmetic. The same ratio runs from about 2 to about 10 at Amazon, Alphabet, Microsoft and Meta.
A note on sourcing. Anthropic figures and quotes are from its SEC Form S-1 of 26 October 2026 and Form S-1/A of 2 November 2026; interim figures are unaudited. Hyperscaler lease figures are from their 10-Qs as compiled in our earlier coverage, and their capex from our Big Tech AI capex tracker. Global Data Center Hub’s figures and quotes are from its post of 3 November. Third-quarter capex and operating cash, all ratios and the reconciliation are our arithmetic.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
