Who Pays for AI Data Centers? Nvidia's $105 Billion OpenAI Guarantee, Anthropic's $5 Billion Quarter of Short-Term Compute, and Who Carries the Debt

Who pays for AI data centers is now written down in the filings. Nvidia has guaranteed up to $105 billion of an OpenAI affiliate's data center leases in Ohio and holds $279 billion of supply commitments; Anthropic paid for $5.0 billion of short-term compute in a single quarter, which took its adjusted operating margin from 19% to 15%; Microsoft, Alphabet, Amazon and Meta plan roughly $720 billion to $740 billion of capital spending for 2026 and are funding the shortfall with new debt and, this season, new equity; and CoreWeave, the listed neocloud, carries $35 billion of debt at 7% to 15% on its term loans.

Put together, the documents say this: the chip seller and the four largest clouds are carrying the construction on their own balance sheets, the AI labs pay a premium when they need compute faster than they contracted it, and the lenders, from the public bond market to the private term-loan funds, hold the debt that pays for the rest. This article reads each party's own filing, dated August 11 to November 2, 2026, sets the bear case and the supply case beside them, and lists what settles the argument and when. It is the second half of a pair: Why is RAM so expensive in 2026? followed the memory price from the makers' margins into Apple's guide. This one follows the compute price one tier up, from the data center into the lab's margin and onto the guarantee table. It makes no forecast of its own. The permitting side of the same question, which states have actually paused data centers and which have only proposed to, is in our data center moratorium tracker.

Who pays for AI data centers? The table

Each row is one party's own filing. The figures use each company's own definitions, so the column is read down, not summed.

Who	What they are carrying, in the filing's terms	Filing and date
Nvidia, the chip seller	Guarantees with a maximum gross exposure of $108.5 billion: up to $105 billion on an OpenAI affiliate's Ohio leases, signed August 2026, plus $3.5 billion of land, power and shell guarantees for AI cloud partners. $279 billion of supply and capacity commitments, up from $119 billion a quarter earlier. $36 billion of AI cloud agreements under which Nvidia buys capacity its partners cannot sell. $20 billion of third-party data center leases it intends to reassign. $99 billion of equity investments, with $25 billion more committed.	Form 10-Q for the quarter ended July 26, 2026, filed August 26
Microsoft	$45.2 billion of capital spending in the September quarter including finance leases ($39.4 billion in cash), against $58.0 billion of operating cash flow. Calendar 2026 "approximately $170 billion", with "over $50 billion" in the December quarter. $25 billion of notes priced October 29, ten tranches maturing 2028 to 2066, rated Aaa and AAA.	8-K and call, October 28; pricing term sheet, October 29
Alphabet	$51.4 billion of capital spending in the quarter against $42.3 billion of operating cash flow; free cash flow negative $9.1 billion. 2026 range "$195-$205 billion", unchanged; 2027 "to increase significantly". $85.2 billion of net debt proceeds in the first nine months; long-term debt $125.1 billion. Off the balance sheet: $75.7 billion of backstop obligations relating to data centers, $130.5 billion of payments for leases not yet commenced, $924.4 billion of material purchase commitments.	8-K and call, October 28; Form 10-Q for the quarter ended September 30
Amazon	$77.3 billion of purchases of property and equipment in the quarter; trailing-twelve-month free cash flow an outflow of $42.4 billion. 2026 cash capital expenditures "approximately $220 billion". $115.1 billion of proceeds from long-term debt in the first nine months, $48.1 billion of it in the quarter; long-term debt $173.1 billion.	8-K, October 29
Meta	$34.9 billion of capital spending including finance lease principal, against $32.8 billion of operating cash flow; free cash flow negative $2.1 billion. 2026 range "$135-145 billion". On October 29, 250 million depositary shares of 6.50% mandatory convertible preferred stock ($12.5 billion at liquidation preference, our arithmetic) and a forward sale of 40 million Class A shares. Two data center ventures, in Louisiana and Texas, with about $27 billion and $14 billion of estimated development cost, in which Meta holds 20% and to which it has given residual value guarantees.	8-K and Form 10-Q, October 21; 8-K, October 29
Oracle	$28.5 billion of capital expenditures in the quarter ended August 31, against $8.5 billion a year earlier. $125.0 billion of senior notes and other long-term borrowings, after $43.0 billion of notes issued in fiscal 2026. $19.9 billion of common stock sold in the quarter through its at-the-market program. Remaining performance obligations $664 billion, from $455 billion a year earlier.	Form 10-Q for the quarter ended August 31, filed September 11
Anthropic, the AI lab	Approximately $895.4 billion of committed spend under purchase commitments and leases, about $447.5 billion of it before the end of 2030 and about $69.8 billion due in the next twelve months. $5.0 billion of compute under short-term commitments in the September quarter, $6.5 billion in the nine months. $6.2 billion of debt, all convertible notes held by Amazon; up to $77.0 billion of financing capacity under note purchase agreements with partners and up to $5.0 billion of equity commitments. $141.1 billion of capital raised to date.	Amended Form S-1, filed November 2
CoreWeave, the neocloud	$31.4 billion of recourse debt and $3.7 billion of non-recourse debt at June 30, 2026, from $21.4 billion six months earlier. Effective interest rates of 7% to 15% on the delayed-draw term loans and about 10% on the senior notes. $103.7 billion of remaining performance obligations. Contractual interest expense $592 million in the quarter, from $250 million a year earlier.	Form 10-Q for the quarter ended June 30, filed August 11
The capital providers	The filings describe the structures and mostly do not name the lenders. Nvidia signed memorandums with "several large capital providers" in August 2026 for platforms intended to mobilize more than $500 billion of third-party capital, and holds $3.3 billion of equity-method investments in infrastructure financiers. Anthropic says a portion of its buildout "has been funded or supported by capital or credit support provided or facilitated by certain of our CSPs and chip partners", and that part of its Broadcom equipment lease was assigned to "a designated financial partner". Partners hold the other 80% of Meta's two ventures. CoreWeave's term loans are from "various lenders" with a bank as agent.	The same filings

What is Nvidia guaranteeing for OpenAI?

Leases, not loans. In August 2026, after the close of its quarter, Nvidia entered into guarantees capped at $105 billion "to provide credit support on a land, power, and shell buildout" by affiliates of SB Energy on behalf of an affiliate of OpenAI, the 10-Q says, covering leases for approximately 4.25 gigawatts of IT load at SB Energy's PORTS Technology Campus in Pike County, Ohio. The mechanics matter as much as the cap. Each guarantee becomes effective when its lease commences, and the amounts step up as each of nine construction phases is completed, the first of which Nvidia expects in its fiscal 2029. The leases run 20 years. Nvidia pays only on certain tenant defaults, and only "defined portions of lease and power payments and not the full cost of the site". The guarantees end if OpenAI achieves a satisfactory credit rating, or when each lease term is complete. In exchange, the campus "will exclusively host NVIDIA AI infrastructure, subject to limited exceptions", and Nvidia holds an option, in its sole discretion, to support about 3.8 gigawatts more as the site scales. OpenAI has agreed to reimburse and indemnify Nvidia for certain losses.

The older guarantees are smaller and already live: $3.5 billion of maximum exposure on land, power and shell guarantees for "select AI cloud partners' data center lease obligations in the event of their default", on terms of five to seven years, with $712 million placed in escrow by the partners. The filing classifies them as credit derivatives.

Around the guarantees sits a larger set of commitments that are not guarantees but carry the same question, which is whether the customer at the end of the chain pays. Supply and capacity commitments, mostly for memory and manufacturing, rose from $119 billion to $279 billion in one quarter. Under a business model introduced in the quarter, Nvidia's AI cloud partners buy its systems and Nvidia commits to buy cloud services back from them, typically for six years, $36 billion in total; the risk factors state it plainly: "if AI clouds do not successfully sell committed capacity to third-party customers, we have agreed to purchase that capacity." Investment-grade customers also get payment terms "ranging from 90 days up to one year" on large data center builds.

Nvidia's own explanation is in the risk factors: "access to capital can be particularly constrained for less-capitalized companies," and AI clouds and AI model makers "currently lack the ability to secure long-term infrastructure contracts and investment-grade financing capacity to secure the AI infrastructure necessary to grow." It expects its large cloud customers and investment-grade enterprises "to continue to secure land, power, and shell commitments independently." The guarantees are for the buyers who cannot borrow on their own name. The next 10-Q, for the quarter ended October 25, shows whether the table grew again.

How much does Anthropic spend on compute, and why did its margin fall?

Anthropic's amended prospectus, filed November 2, prints the buyer's side of the same market. Revenue for the September quarter was $17.0 billion, up 1,127% from a year earlier and 47% from the June quarter's $11.6 billion (our arithmetic on the two filed figures). The operating result went the other way: a loss of $567 million against operating income of $974 million in June, "primarily driven by $5.0 billion of compute under short-term commitments, which currently costs significantly more than comparable compute procured under our multi-year commitments, and $2.0 billion of non-cash charges related to our donation matching program." Adjusted operating income still rose, to $2.5 billion from $2.2 billion, but the adjusted operating margin went from 19% to 15%, "primarily driven by the $5.0 billion of short-term compute."

The filing explains the mechanism without a number for the price. The majority of Anthropic's compute is contracted under multi-year commitments procured "12 to 18 months in advance of availability"; when demand or research needs exceed that capacity, it buys short-term compute at the higher price, and it "expect[s] our use of short-term compute to continue to increase in the near term". For the nine months the short-term line was $6.5 billion. The spot price of compute is not published anywhere official, and Anthropic's line is the closest thing to a filed reading of it.

The commitments behind the multi-year compute are the largest numbers in either filing. As of and after September 30, Anthropic had entered into purchase commitments and leases "with an aggregate committed spend of approximately $895.4 billion", about $447.5 billion of it before December 31, 2030, substantially all of the rest between 2031 and 2045, and about $69.8 billion due in the next twelve months. A significant portion relates to Anthropic Direct, its name for leasing chips itself and hosting them in dedicated facilities rather than renting through a cloud. Against that stand $131.9 billion of cash and short-term investments at September 30, $6.2 billion of debt in convertible notes held by Amazon, and up to $77.0 billion of financing capacity under note purchase agreements with partners, including up to $42.0 billion of convertible notes issuable to Broadcom or a designated financial partner. The prospectus also says what the guarantees above imply from the other side: "a portion of our compute and infrastructure build-out has been funded or supported by capital or credit support provided or facilitated by certain of our CSPs and chip partners," and reducing that reliance "depends in part on our ability to obtain financing from independent sources on our own credit." The price range of the offering is still blank in this amendment. The loss, the votes and the original compute commitments are in Anthropic's IPO prospectus, explained.

How is AI data center debt financed? The clouds and the bond market

The four largest clouds build most of it and pay for most of it from operations. In the September quarter, operating cash flow covered capital spending at Microsoft, $58.0 billion against $39.4 billion in cash, and fell short at Alphabet ($42.3 billion against $51.4 billion), Meta ($32.8 billion against $34.9 billion) and Amazon, whose trailing-twelve-month free cash flow was an outflow of $42.4 billion. The gap is closed in the bond market and, this season, in the equity market: Alphabet's and Amazon's nine-month debt proceeds, Microsoft's ten-tranche note deal and Meta's preferred stock and forward sale are in the table above. Oracle, which is building capacity for customers under contract, spent $28.5 billion on capital expenditures in the quarter ended August 31, more than three times the year-earlier figure, carried $125.0 billion of senior notes and other long-term borrowings, and sold $19.9 billion of stock through its at-the-market program in the quarter.

The commitments beside the debt are larger than the debt. Alphabet's 10-Q lists $130.5 billion of future payments for leases not yet commenced, $75.7 billion of "financial support in the form of backstop obligations relating to data centers" and $924.4 billion of material purchase commitments. Meta's shows a different structure: two ventures, formed in October 2025 and July 2026, that are building its Louisiana and Texas campuses with about $27 billion and $14 billion of estimated development cost; Meta holds 20% of each, will lease the sites from them starting in 2029 and 2028 on initial four-year terms renewable to 20 years, and has given each a residual value guarantee that pays the shortfall between a property's fair value and a threshold if Meta walks away. Meta's maximum exposure table shows $18.5 billion of lease commitments and $3.7 billion of future funding; the partners who hold the other 80% are not named. The eight-quarter capex series, cloud growth and the full financing table for the four are in Panic of 1873: the railroad bubble and what big tech's filings say about AI capex.

Who lends to the neoclouds?

CoreWeave's 10-Q for the quarter ended June 30 is the clearest filed statement of what the marginal borrower pays. Its debt rose from $21.4 billion at the end of 2025 to $35.1 billion in six months: $31.4 billion with recourse to the parent and $3.7 billion without. The table of instruments carries its own price list: delayed-draw term loans at effective rates of 15% (the 2028 facility), 11%, 9% and 7% (the non-recourse 2032 facility), senior notes due 2030 to 2032 at about 10%, equipment financing at 9% to 11%, and convertible notes at 2%. The newest term loan is secured on the equity and substantially all the assets of the subsidiary that holds the equipment, requires interest-rate hedges on at least 95% of its floating borrowings, and carries restricted-cash requirements. Contractual interest expense was $592 million in the quarter, against $250 million a year earlier, and $4.4 billion of principal is due in the rest of 2026 and $6.2 billion in 2027. On the other side of the ledger sits $103.7 billion of remaining performance obligations, 41% of it expected to be recognized over the two years to June 2028.

The lenders are the part the filings describe without naming. CoreWeave's facilities are "with various lenders" and a bank as agent. Nvidia's filing holds the structural backstop for this class of borrower: the $36 billion of AI cloud agreements under which it buys unsold capacity, the $3.5 billion of lease guarantees, the $3.3 billion invested in "infrastructure financiers", and the August memorandums "with several large capital providers" to mobilize more than $500 billion of third-party capital, with Nvidia providing, at its option, "limited residual-value support for a portion of specific projects". A neocloud in 2026 borrows against three things: its contract with a hyperscaler or a lab, the resale value of the equipment, and whatever the chip seller has agreed to stand behind.

Is AI a bubble in 2026? The bear case and the supply case

Two readings of the same filings, from two people who agree on the mechanism and disagree on its sign. Ed Dowd of Phinance Technologies made the bear case on the Risk Takers podcast, recorded October 30 and published October 31; Dylan Patel of SemiAnalysis made the supply case in a talk at Modal's Runtime 2026 conference, published November 2. Each is reported as that speaker's view.

Question	Ed Dowd, October 30	Dylan Patel, published November 2
What the guarantees mean	Vendor financing in the shape of Lucent and Nortel in 2000; Nvidia "will probably miss" within six to twelve months as neoclouds are shut out of the debt market	The only way the open ecosystem gets compute; Nvidia is "loading up their balance sheet" because the buyers cannot borrow; 6.5 gigawatts backstopped so far, by his count
The price of compute	High memory and compute prices are already destroying demand, in PCs and phones first	Rental prices for an H100 have risen from a trough of about $1.70 an hour to $2.40 to $3.20, by his firm's data, because every rented GPU is profitable at the token level
Neocloud debt	Priced out of the junk market; the bankruptcies follow, as the competitive carriers' did	Hard to raise, and bridged by selling the capacity to a hyperscaler or a lab and by Nvidia's backstops
Hyperscaler capex	Amazon's unchanged guide implies a 30% to 40% sequential drop in the December quarter: "growth is slowing or pausing"	About 40 gigawatts added worldwide in 2027, the labs taking roughly a third; $11 trillion of AI capex by the end of 2029, $7 trillion of it debt
Who pays	Lenders (private credit, insurers, pensions), neocloud equity, and Nvidia through a miss; no public bailout	Buyers pay more now; the capital markets fund the rest; the loser is whoever has no balance sheet

The claims that can be checked against a document, checked. Dowd's Amazon arithmetic holds on Amazon's own figures: approximately $220 billion for the year, less the $175.7 billion of gross purchases through September, leaves roughly $44 billion for the December quarter against $77.3 billion in the September quarter; that is our arithmetic, the release says "approximately" and nothing about a slowdown, and Amazon said on the call, per a published transcript, that it expects to spend meaningfully more in 2027. His claim that hyperscaler free cash flow "is now negative" holds for three of four: Alphabet's and Meta's quarters and Amazon's trailing year were negative, Microsoft's was not. His figure for PC shipments, down 4%, understates IDC's count, which was a 20.1% decline in the third quarter, a number that strengthens his demand point while being wrong as stated. His reading that Anthropic left the September quarter out of its prospectus because the numbers were weak was overtaken three days later: the amendment prints $17.0 billion, up 47% on the June quarter, with the margin compression described above. Patel's filed claim, Nvidia's backstops, checks out in shape and roughly in size: the filing shows 4.25 gigawatts guaranteed with an option on 3.8 more, against his 6.5; his rental prices, gigawatt counts and the $11 trillion are his firm's data and his forecast, and are reported here as such.

The two theses fail on different evidence. Dowd's fails if Nvidia reports above its guide and raises the next one while a neocloud prices a bond; Patel's fails if rental prices fall while the guarantee table keeps growing. Both can also be right in sequence, the shortage first and the bust after, which was the order of 1999 to 2001.

What would settle it, and when

• November 10, 2026: CoreWeave's September quarter. Not yet confirmed by the company as of November 2, per the SentiSense earnings calendar. The debt table, the effective rates on any new facility, and remaining performance obligations.
• November 17, 2026, 2 p.m. PT: Nvidia's fiscal third quarter, the quarter ended October 25, confirmed by Nvidia on October 28. Data center revenue and the January-quarter guide are the first public numbers that test either reading. The 10-Q that follows (filed in late November in past years; the date is not announced) carries the guarantee table, the supply commitments and the AI cloud agreements: whether $108.5 billion, $279 billion and $36 billion grew.
• Anthropic's next amendment and its pricing day, neither dated. The price range, then the first quarterly report as a public company with the short-term compute line, the one filed reading of the spot price.
• Mid-December 2026: Oracle's quarter ending November 30. The SentiSense earnings calendar lists December 14, not yet confirmed by the company as of November 2. Capital expenditures, borrowings and the at-the-market program against $664 billion of remaining performance obligations.
• Late January and early February 2027: the four clouds' December quarters and 2027 capex. Alphabet promised details on its fourth-quarter call, Microsoft said spending would accelerate in the first half of calendar 2027, Amazon said meaningfully more, Meta is still building its budget.
• 2028 and 2029: the guarantees go live. Meta's Texas lease commences in 2028 and its Louisiana lease in 2029; Nvidia expects the first Ohio phase in its fiscal 2029. Until then the $105 billion is a cap, not a payment.

No sign-up required. See Nvidia's earnings, filings and news on its SentiSense page, live →

Sources

All read November 2, 2026.

• NVIDIA Corporation, Form 10-Q for the quarter ended July 26, 2026, filed August 26, 2026: Note 6 (equity method investments), Note 8 (guarantees as derivatives), Note 10 (commitments and guarantees, including the SB Energy guarantees), the MD&A and the risk factors. Conference call notice, October 28, 2026.
• Anthropic, PBC, Amendment No. 1 to Form S-1, filed November 2, 2026: prospectus summary, risk factors (indebtedness), MD&A (compute efficiency, results for the three months ended September 30, 2026, commitments, liquidity and capital resources).
• Microsoft, Exhibit 99.1 to Form 8-K, October 28, 2026; company-hosted transcript of the October 28 call (capex outlook, finance leases); pricing term sheet, October 29, 2026 (the ten tranches sum to $25.0 billion, our arithmetic).
• Alphabet, Exhibit 99.1 to Form 8-K, October 28, 2026; company-hosted transcript of the October 28 call (capex outlook); Form 10-Q for the quarter ended September 30, 2026 (debt, backstop obligations, leases not yet commenced, purchase commitments).
• Amazon, Exhibit 99.1 to Form 8-K, October 29, 2026. The 2027 remark is from a published transcript of the call, not a company document.
• Meta Platforms, Exhibit 99.1 to Form 8-K, October 21, 2026; Form 10-Q for the quarter ended September 30, 2026 (data center ventures); Form 8-K, October 29, 2026 (preferred stock and forward sale).
• Oracle, Form 10-Q for the quarter ended August 31, 2026, filed September 11, 2026.
• CoreWeave, Form 10-Q for the quarter ended June 30, 2026, filed August 11, 2026: Note 2 (remaining performance obligations) and Note 10 (debt).
• Ed Dowd, Risk Takers podcast, recorded October 30 and published October 31, 2026. Dylan Patel, talk at Modal's Runtime 2026 conference, published November 2, 2026. IDC, PC shipments in the third quarter of 2026, October 8, 2026.
• Earnings dates: SentiSense earnings calendar, November 2, 2026; CoreWeave's and Oracle's dates were not yet company-confirmed.

Figures are as stated in the documents above on November 2, 2026, in each company's own definitions, and are rounded to one decimal in billions. Arithmetic across filed figures is labelled as ours where it appears. Every forecast quoted is the named person's or company's own and can change. This article is educational and is not investment advice.