The AI Build-Out Ledger — a running record of how the AI boom is financed, reported and disclosed

01The $3 trillion in the footnotes

Every quarter the large technology companies publish a capital-expenditure figure, and that figure is what most coverage of the AI build-out is built on. The Wall Street Journal reported on August 17, 2026 that it does not describe the scale of what the companies have actually committed to, because a large share of their future obligations never reaches the balance sheet at all.

Reading the footnotes of the most recent securities filings at nine companies — Alphabet, Amazon, Meta, Microsoft, Oracle, Nvidia, Broadcom, SpaceX and Advanced Micro Devices — the Journal put their combined off-balance-sheet commitments, mostly AI-related, at about $3 trillion.most recent filings Trailing capital expenditure across the same set was about $600 billion, and the off-balance-sheet total was, per the Journal, roughly triple what those companies owe under their outstanding leases and long-term borrowings put together.

All three figures are nine-company totals — Alphabet, Amazon, Meta, Microsoft, Oracle, Nvidia, Broadcom, SpaceX and AMD — not the four companies broken out in the graphic in Entry 02. Source: WSJ analysis of company filings, Aug 17, 2026. Data as of June, except Nvidia (April) and Meta (July).

The Journal also notes what the disclosures do not resolve. The commitments are described in general terms, they span many years — Alphabet's energy agreements run as far out as 2054 — and in some cases they cover things other than compute: Nvidia committed to make $27 billion in equity investments between April 26 and the end of its fiscal year in January 2027, and some companies' off-balance-sheet exposures include agreements to buy other companies' stock in future or to backstop leases for other tenants.

The Journal returned to the point on September 24, in a piece on the size of the whole build-out. Stijn Van Nieuwerburgh, the economist whose estimate of that total is set down in Entry 11 and who is quoted in Entry 12, told the paper that technology companies often borrow from banks and private-credit firms through off-balance-sheet entities, in deals that usually come with little public reporting — which, he said, makes it hard to work out how large the financial risks are. The same article reports that a growing share of the hyperscalers' capital spending is financed with debt.

Two reports at the turn of October carry the point past leases and purchase commitments. On September 30 Bloomberg described the guarantees that large companies now give for smaller tenants' leases, which as contingent liabilities are disclosed in footnotes and not recorded on the balance sheet (Entry 13). On October 2, relaying the Financial Times, it reported that Amazon is exploring a vehicle to hold about $8 billion of chips outside its own balance sheet (Entry 04). Neither is part of the $3 trillion above, which the Journal counted from filings as of June.

Two more counts, on two more rosters

In October two outlets put their own totals on the same idea, each for a different group of companies. Neither says how its figure was assembled.

Bloomberg's “nearly $3 trillion” comes from a feature on who holds the risk of the build-out, and includes guarantees, which Entry 13 treats as a separate kind of obligation. The Economist's $2.4 trillion comes from an article about the reliability of the numbers in the AI trade, and it adds a comparison of its own: the commitments, it says, exceed what the four companies have borrowed from bond markets. The three totals cover different companies and, in Bloomberg's case, a different set of instruments. They are printed side by side, as the two counts of local restrictions are in Entry 11, and not reconciled or added together. The Economist's comparison with bond borrowing is its own; it is not set against the bond-issuance figures in Entry 12, which cover yet another roster.

02On the sheet, and off it

The Journal's central graphic sets the two on-balance-sheet obligations most readers already track — lease liabilities and long-term debt — against the two that sit outside it, for four companies: Alphabet, Amazon, Meta and Microsoft.

Select commitments, on-balance sheet vs. off — Alphabet, Amazon, Meta and Microsoft

Graphic: The Wall Street Journal. Reproduced from "Why Big Tech's AI Spending Is $3 Trillion Higher Than It Seems," The Wall Street Journal, August 17, 2026. WSJ's note: "Data are as of each company's most recent quarterly filing and in billions, except where noted." Source line as published: WSJ analysis of company filings. The interactive original — the "tap or hover" prompt visible in the image belongs to WSJ's live version — is on wsj.com.

03The running ledger

The per-company figures behind the graphic, in a form that can be extended each quarter. Blanks are meaningful: they mark companies the WSJ included in its $3 trillion total but did not break out in the published graphic, and categories the sources on file do not disclose.

04Anatomy of an off-book data center

The Journal uses one project to show how an obligation of this size ends up outside a balance sheet: Meta's "Hyperion" campus in Louisiana, a site it describes as about the size of 1,700 football fields.

Meta is building Hyperion. Neither the campus nor the $27 billion of debt financing its construction appears on Meta's balance sheet. Funds managed by Blue Owl Capital own the majority of a joint venture that owns the campus; a holding company, Beignet Investor, holds the Blue Owl stake and raised the construction financing in a bond sale. Meta is the minority partner — and the tenant. Its lease payments are the cash flow that helps pay the bondholders.

Two weeks after this entry was written, the same campus appeared in a different kind of story. Reporting on September 5 on the local resistance now facing data-center construction, the Journal's Christopher Mims noted that at Hyperion Meta is committing to lower the energy bills of nearby residents and to return more water than it takes out. Those are pledges rather than disclosures — the column gives no baseline and verifies no delivery — and they attach to the same project whose $27 billion of construction debt sits outside Meta's balance sheet. What the site owes bondholders is in a filing; what it owes the county is not. The wider count of local restrictions is in Entry 11.

Figures and terms in this entry: WSJ, Aug 17, 2026, except where marked — the Sep 5 WSJ column on local pledges, and MIT Technology Review, Sep 15, 2026, in the first panel above; and Bloomberg, Oct 2, 2026, relaying the Financial Times, in the panel on chips.

05Alphabet's $811 billion

One line in the graphic is larger than any other and moved further than any other. Alphabet's purchase commitments and contractual obligations stood at $811 billion as of June 30, 2026 — against $332 billion reported three months earlier.

The Journal's account of what is and is not known here is brief and worth keeping intact. As with the other companies, it is hard to tell from the disclosures what precisely Alphabet intends to buy: the company said the commitments primarily relate to "technical infrastructure and inventory" and to "agreements to secure energy for data center usage." Alphabet did not detail why the obligations increased so much from the figure three months earlier. The commitments span several years, with the energy agreements running as far out as 2054.

Figures per WSJ, Aug 17, 2026, from Alphabet's filings. The $708.0B and $479B comparisons are arithmetic on the figures in the WSJ graphic and article, not WSJ statements.

06Where the profits came from

Three days before the WSJ piece, The New York Times published a separate finding drawn from the same quarter's filings — this one about earnings rather than obligations. A large share of the Magnificent 7's second-quarter profit did not come from selling anything.

The seven companies — Alphabet, Amazon, Microsoft, Meta, Apple, Tesla and Nvidia — reported $315.6 billion in net profit for the second quarter (the three months through June, except Nvidia, whose financial year differs). Of that, $134.6 billion, or about 42%, came from investment gains, according to calculations by Matt King of Satori Insights, who adjusted the reported numbers for taxes. The comparable share in the previous quarter was 5%. Without those gains, King told the Times, the group's profits would have been roughly equal to the previous quarter — not growing.

07The circle

The Times frames the composition of those gains as evidence of something it calls circularity: the companies pushing the market higher are increasingly dependent on one another's success, and money invested in AI companies is frequently spent buying products and cloud services back from the investors.

The Times notes that this concern is not new — investors have watched tech giants, chipmakers and AI labs invest in or lend to one another for some time — and that the practice has defenders. Sam Altman, chief executive of OpenAI, "has described the deals as a creative way to free up the capital needed to supercharge innovation at a time of rapid change."

One return flow the database above left unquantified now has a figure, from Anthropic's draft prospectus: about $351 million paid to Amazon's and Google's cloud platforms in distribution fees on $2.16 billion of 2025 sales made through themdraft, by a Reuters analysis (Entry 11). It is a fee for distribution and is not the revenue share the database describes; the two are not treated here as one thing.

That is the circularity described by a participant rather than by a critic, and described as a financing gap: the labs had extraordinary demand for compute and could not secure the infrastructure independently. The two outside voices CNBC quotes read the equity positions the same way. Forrester's Naveen Chhabra says the optics stakes keep partners' tooling optimised for Nvidia's architecture and protect "the CUDA software moat" against AMD and cloud providers' in-house silicon; CCS Insight's Ian Fogg says equity gives Nvidia "a degree of control to encourage companies to take a Nvidia-related innovation path." Both describe control instruments. Neither describes returns, and no skeptic appears in the piece.

08Cash, and paper

The Journal's August 17 analysis and the Times's August 14 report end near the same question, from different documents: how durable are the numbers just reported? The Journal looks at cash generation; the Times looks at whether the gains hold.

A third instance of the same accounting mechanism, reported on August 31, runs in the opposite direction and at a counterparty rather than a customer. SB Energy's warrant liability — the warrants it issued to OpenAI — was remeasured from $3.6 billion in January to $5.5 billion at June 30, and the Journal attributes most of the widening of the company's first-half net loss, to $3.2 billion, to changes in that estimated value. Marks move reported profit in both directions, and at both ends of these deals. See Entry 10.

A fourth instance sits at the supplier. Nvidia's $5 billion investment in Intel was carried at $30 billion when CNBC reported its book on September 4 — a $25 billion increase that is appreciation, not cash. The $25 billion is arithmetic on CNBC's two figures, not a CNBC statement. CNBC describes the $99 billion total in Entry 07 as buoyed by rising technology stocks; the separable cash figure inside it is the more than $40 billion committed during 2026. So the same mechanism the Times found in the Magnificent 7's earnings — unrealised marks flattering reported profit — is running at the company selling the hardware, at a counterparty in Entry 10, and at the customers, simultaneously.

The cash side has since widened. Bloomberg reported on October 1 and again on October 4 that Alphabet, Amazon and Meta have all seen free cash flow turn negative on an annual basis, and that Microsoft is the only one of the four that has not. Those are the four companies of Entry 03. Neither article gives the figures or the period, so this is a pointer to the third-quarter filings due in late October, where the numbers can be read, and not yet a figure for the ledger. In the second article Robert Schiffman, an analyst at Bloomberg Intelligence, says that Meta, Amazon, Alphabet, Microsoft and Oracle have cash needs well beyond what they generate internally, which will raise their leverage over the next two years, and that their credit ratings have held because expected growth in EBITDA offsets the added debt. That is an analyst's view in a markets article, and is recorded as one.

Neither observation is a forecast in its source. The Journal's framing is conditional: if the assumptions about technology and demand prove wrong, "these deals to clinch future capacity could become a monstrous burden," companies would be "paying an expensive tab for infrastructure that they can't profitably use," and the obligations "could also lead increasingly indebted companies to have to borrow even more." The Times's is descriptive: investment gains "could continue to fluctuate as some of the private companies they are invested in go public."

09The range of views

Neither of the two August reports this page began with — the Journal's August 17 analysis and the Times's August 14 report — reaches a verdict, and neither does this page. What both do, along with the Journal's August 31 report on SB Energy, is record a spread of positions among the people whose job it is to price this. Here is that spread, in the sources' own terms.

10The landlord's book

Every entry above reads a filing by a company that is buying AI infrastructure. On August 31, 2026 the Journal reported from the other side of the contract: the draft IPO documents of SB Energy, a data-center developer majority-owned by SoftBank, which show what one of these commitments looks like on the books of the company receiving it — and what it cost to win the tenant.draft · terms not final

SB Energy is expected to make its IPO filing public within days of this edition, and to go public as soon as September, aiming to raise $5 billion to $7 billion. The Journal reviewed the draft documents and reports that the terms are not yet final and could change. What the drafts describe is a company whose data-center customers are also its investors: SoftBank and OpenAI are the planned tenants in three SB Energy data centers, while a fourth site — a 900-megawatt campus in Scurry County, Texas — has no customer yet. The Journal describes the documents as the first detailed look at the financial ties between SB Energy and the companies supporting its expansion into AI infrastructure.

Since then. SB Energy filed its registration statement publicly, on Form S-1, on September 1, for a listing on Nasdaq under the ticker SBE (SoftBank's announcement of September 2). Reuters reported on October 15 that it decided in September to delay the listing. The public filing has not been read for this page. Every figure in this entry is still the draft's, as the Journal reported it, and keeps its pill.

All figures in this entry are one private company's, from SB Energy's draft IPO documents as reported by the WSJ, Aug 31, 2026. They are not part of, and are not added to, the nine-company or four-company totals elsewhere on this page — see scope boundary 4 in the note near the top of this page.

What the warrants did to the income statement

The warrants are not only an inducement; they are a liability that gets remeasured every quarter, and the remeasurement runs through profit. SB Energy's net loss for the first half of 2026 widened to $3.2 billion, from around $250 million in the first half of 2025 — a move the Journal attributes largely to changes in the estimated value of its warrant liabilities. Over the same period the company's revenue, which comes primarily from its renewable-power unit rather than from data centers, was around $140 million, up 66% from the same period a year earlier.

Warrant values, net-loss figures and revenue: WSJ, Aug 31, 2026, from the draft IPO documents. The $1.9B six-month increase in the warrant value and the roughly $2.95B widening of the net loss are arithmetic on those figures, not WSJ statements. The June 30 measurement date is the same one behind the disclosures in Entries 02, 03, 05 and 06 — but these are one private company's figures, and are not combined with them.

A second ledger: the developers

Entry 03 tracks these commitments as they appear on the tenants' balance sheets. The same contracts appear on the developers' books as backlog. This table starts that second count, and is meant to be extended as reporting on other developers arrives.

Source for this entry in full: "The $5.5 Billion Perk SoftBank's Data-Center Venture Offered to Land OpenAI," The Wall Street Journal, August 31, 2026, reporting on draft IPO documents the Journal reviewed. SoftBank is also one of OpenAI's major investors. Because the documents are drafts and the terms are not final, every figure in this entry is provisional in a way the quarterly disclosures elsewhere on this page are not.

11What the private companies have signed

Entries 01 through 10 read documents that companies filed. The two largest compute books in the industry belong to companies that file nothing: OpenAI and Anthropic are private. What circulates in place of a disclosure is an outlet's estimate, a forecast relayed by someone who saw it, and a consultancy's projection to 2050 — three numbers of very different weight that are already being quoted in the same breath. This entry sets them down with their provenance attached. From Edition 7 it also carries the first figures from one of those companies’ own documents — Anthropic’s draft prospectus, not yet public, as Reuters reports it — kept beside the estimates rather than put in place of them.

Where each number comes from

Anthropic — $517 billion. Since October, per The Information's analysis, the company has entered agreements for at least 14.8 gigawatts of compute capacity, on top of between 1 and 2 GW it held before, and could spend as much as $517 billion mostly over the next decade on cloud capacity, chips and data-center leases. The comparison the piece draws is with what Anthropic itself told investors in December 2025: $180 billion on rented servers through 2029. The two are not like for like — one is renting servers through 2029, the other is cloud, silicon and leases mostly over the next decade — so no multiple between them is computed here. The $517 billion is not a disclosure: Anthropic has never published a total compute figure, the estimate is assembled from public statements and the outlet's prior reporting, and several inputs rest on unnamed sources. It is a ceiling of announced commitments. The article says so itself, and lists the reasons announced capacity may never arrive: permitting, grid interconnection, component shortages and municipal opposition.

OpenAI — $750 billion and 30 GW. In April the company told investors it planned to have 30 gigawatts of compute by 2030. Last month it projected spending around $750 billion on compute through 2030, according to a person familiar with the forecast — up from a forecast of $665 billion earlier in the year. Neither figure is filed anywhere. The Wall Street Journal first reported the projections. On September 16 the Journal added a financing figure: OpenAI is in early talks with investors on a round that would value it at more than $1.2 trillion, ahead of a public offering now expected next year, against the $852 billion it was valued at in March after a $122 billion round from Amazon, Nvidia and SoftBank. The talks are attributed to one person familiar with them.est

A second baseline. On October 14 the WSJ, in its own voice and in a report on the departure of OpenAI's head of computing capacity, put the company's projected compute spending at around $750 billion through 2030, up from a projection of roughly $600 billion earlier in the yearest. The $750 billion agrees with the figure above; the earlier one does not. Which earlier forecast each outlet means is not stated, and both are printed. The same report describes OpenAI's August arrangement with SB Energy as a 10-gigawatt lease; Entry 10 separates the roughly 8 GW leased from the 10 GW of power behind it, and keeps that distinction.

PwC — $31.6 trillion. Global spending on AI infrastructure is projected at $31.6 trillion through 2050, with annual data-center capital expenditure rising from roughly $800 billion now to $1.8 trillion in 2050. The structural claim underneath it is the part that distinguishes this projection from a construction forecast: unlike earlier infrastructure booms, which taper after the initial build-out, this one is expected to accelerate, because servers and GPUs have to be refreshed every few years — and on PwC's reading recurring chip upgrades, rather than construction, will drive most of Nvidia's long-term capital investment. The figures reached this page through the WSJ's CIO Journal newsletter of September 8. The condition attached to them — that the later years are only fundable if adoption and pricing hold — is the newsletter's own reading of the scenario, not a PwC finding.

Ahead of the offerings

Both companies whose books this entry estimates are now preparing to go public, and the preparation has produced a new set of figures. None is filed. Each comes from unnamed people, or from a company document seen by one outlet and relayed by another.

Anthropic. According to The New York Times on September 18, the company could file publicly for an IPO within weeks and begin trading as soon as November; the people cited stress that the plans could change. Four people told the Times that annualised revenue is expected to pass $100 billion by the end of this yearest, and that investors are using that forecast to justify a potential valuation of $2 trillionest. Three put the company's compute at about 5 gigawatts by year-end, roughly doubling by the end of 2027est. Reuters, the same day and citing two people, reported that the listing could slip past the November midterms and that marketing to investors was not expected before mid-October. It also reported that the company is weighing what to spend on its next model against profitability.

The 5 GW is a fourth capacity figure for Anthropic on this page, and not the same kind as the other three: it is compute expected to be running at year-end, where the 14.8 GW above is capacity contracted since October, the 1–2 GW is what the company held before that, and the 11 GW from Amazon and Google sits inside the 14.8. No ratio between them is computed.

OpenAI. Bloomberg, relaying a Financial Times report on September 18, says a private company presentation in July projected negative free cash flow of $278 billion from 2026 through the end of 2030est, with revenue rising from $36 billion this year to $350 billion in 2030est while the company invests heavily in computing power and expects its expenses to far outpace that growth. The presentation was made in connection with a computing deal. The one person Bloomberg cites, who has knowledge of the deal, confirmed that the presentation took place but declined to discuss its figures, so the numbers rest on the FT's reporting and reach this page second-hand. Chief executive Sam Altman has said the company's IPO will not happen this year; Bloomberg adds that a new private round would give it room to push the listing back a further one or two quarters, or to fund acquisitions.

The $278 billion and the $750 billion above both run to 2030 and both are company projections, but one is cash flow after revenue and all spending and the other is compute spending alone, from forecasts made at different times. Neither is subtracted from the other.

The two figures describe the same talks. Neither outlet addresses the other's number, and they are printed side by side rather than reconciled. The same treatment applies to Anthropic's timing: as soon as November in the Times, possibly after the midterms in Reuters.

Since then. Bloomberg reported on October 1, from people familiar with the matter, that Anthropic could begin formal marketing as soon as the week of November 9 and begin trading before Thanksgiving, after pushing back earlier plans, with an investor day set for October 14; the timeline could still shift. Some of those people said prospective investors see a fair valuation of $1.8 trillion to $2 trillionest. That is a third account of the timing and is kept beside the other two.

A third figure for OpenAI's round came on September 29. Bloomberg, again from people familiar, reported that the company aims to raise at least $30 billion at a valuation of around $1.4 trillionest, as a bridge round in place of a public offering, which its chief executive has said will not come this year. Bloomberg presents the figure as a step up from the $1.2 trillion it had reported earlier, and says the talks are early and could change. One precision: the $1.4 trillion does not include the money to be raised, while the $852 billion from March includes the money raised then. The two are not on the same basis, and no multiple is computed. The last installments of that March round have since been paid (Entry 07).

The investor day. Anthropic presented to prospective IPO investors in San Francisco on October 14, and three outlets reported what was said, each from unnamed people briefed on it. All three put third-quarter revenue at $16.8 billionest, about 45% above the second quarter. The WSJ and Reuters give the second quarter as $11.6 billionest; the WSJ gives the first as just under $5 billion, having corrected an earlier $5.7 billion. The Information describes the third quarter as more than 3.6 times the company's revenue for all of 2025; that is its arithmetic, and the 2025 figure on this page is the draft prospectus's nearly $4.6 billion.

The two describe different periods, a year to date and a single quarter, and both can hold. Neither is a measure under accounting rules, and neither is set against the 2025 results in the draft.

The Information adds one thing the company volunteered and one it withheld. It estimated that 5% to 7% of its revenue comes from distillation — other companies using its models to gather data for training their own — and said that revenue will decline and weigh on the current quarter. And it gave no update to annualised revenue, last reported at $65 billion in July. The WSJ reports that investors expect annualised revenue above $110 billion by year-endest; the NYT's $100 billion of September 18, above, is kept beside it.

On valuation, three of the four people Reuters cites say some attendees discussed about $1.5 trillionest, against roughly $2 trillion previously expected. The Information says some investors thought the bankers should lower expectations to perhaps even less than $1.5 trillion, against the $2 trillion some bankers had floated, and gives the private valuation in May as $965 billion. The WSJ repeats its earlier reporting that the offering could raise as much as $100 billion at around $2 trillion. The figures are kept as each outlet gave them. On timing, The Information expects the listing before Thanksgiving, and the WSJ says the paperwork is expected to be made public in the coming weeks; as of its report it had not been. For context, Reuters notes that SB Energy decided in September to delay its listing (Entry 10), and The Information that Firmus Grid, an Nvidia-backed data-center company, pulled what would have been Australia's largest IPO in almost 30 years.

The quarterly figures are not annualised here. The $16.8 billion is not set against the $65 billion run-rate or either year-end expectation, its 45% is not compounded, and no valuation is divided by any revenue figure.

The book, by counterparty

Entry 03 counts commitments as they sit on the tenants' balance sheets, and the second ledger in Entry 10 counts the same contracts as backlog on a developer's. This third table counts one buyer's book by who is on the other side of each contract, because the counterparty and the length of the contract are what the totals above conceal.

What the draft prospectus says

On September 28 and 29 Reuters reported from Anthropic's IPO prospectus, which the company submitted confidentially to the Securities and Exchange Commission in June and has not made public. Reuters says it has seen the document; Anthropic declined to comment on the first report and did not respond on the second. The figures below are the company's own, but they reach this page through one outlet's reading of a draft that may change before it is filed publicly — the position Entry 10's SB Energy figures were in. They carry a draft pill rather than an est one.

The income statement. Revenue grew twelvefold in 2025 to nearly $4.6 billion. Operating expenses were $12.65 billion, of which compute and infrastructure were the $7.33 billion above, and the operating loss widened to $8.06 billion from $2.98 billion in 2024. The net loss of nearly $42 billion is far larger because it includes an accounting charge of about $34 billion, reflecting a rise in the estimated value of financing that could eventually turn into Anthropic shares — not, in Reuters' account, money spent running the business. It is a charge of the same kind as the warrant remeasurement Entry 10 describes at SB Energy: a change in the estimated value of instruments that may become shares, run through the income statement. Cash, cash equivalents and short-term investments were $20.28 billion at December 31.draft · FY 2025

How the revenue is reported.draft On September 29 Reuters reported, from the same draft, that 47% of Anthropic's 2025 revenue, about $2.16 billion, was sold through the cloud marketplaces of Amazon and Google, against 11% in 2023 and 32% in 2024. Anthropic books the full value of those contracts as revenue, on the ground that it sets the prices and delivers the service, and records the platforms' cut as a cost within its sales, marketing and partnerships line. By a Reuters analysis that cut was roughly $351 million, or some 16 cents of each dollar of those sales. OpenAI has told investors and employees that the approach inflates Anthropic's reported revenue by billions of dollars, Reuters reported in June; Anthropic said then that it follows established accounting practice as the principal in the transaction. The two positions are printed as given. The $4.6 billion above is revenue on Anthropic's own method, and no net figure is computed here. The same two companies also collect much of its cash: they were responsible for 60% of the $909 million of customer bills outstanding at the end of 2025, up from 42% a year earlier, and the draft warns that disputes or delays there could hurt cash flow. Of the year's revenue, about $3.8 billion was billed by usage and $789 million by subscription.

The report carries two further commitment figures. At the end of 2025 Anthropic had $54.6 billion of non-cancellable hosting and computing commitmentsdraft, and by early 2026 its total long-term commitments exceeded $417 billiondraft, covering 3.5 GW of dedicated computing capacity. Reuters does not relate either figure to the $518 billion. They are measured at different dates and described in different terms, so none is subtracted from another, and the 3.5 GW is not set against the other capacity figures in this entry.

What an adjusted figure leaves out.est On October 4 The Information reported a further item, from investors who were shown Anthropic's figures ahead of the offering and not from the draft as Reuters read it. The company booked a non-cash expense of more than $660 million in the six months from October 2025 to March 2026 for stock it issues to match employees' pledges of their own stock to charity: three to one for employees who joined in 2024 or earlier, on up to half their stock, and one for one on up to a quarter for later hires. Charitable contributions were $540 million for 2025 and about $125 million in the first quarter of 2026, which the outlet puts at 10% of employee expenses and 2% of operating costs. Anthropic leaves the expense out of its adjusted operating profit, together with stock-based compensation and a large legal settlement. The data firm Calcbench told the outlet that such an adjustment for charitable contributions is uncommon, and that the next-largest comparable figure among Fortune 500 filers in 2025 was BlackRock's $109 million. Because the match is paid in newly issued stock, it also dilutes other shareholders. The three amounts are printed as the article gives them. It does not reconcile the six-month figure with the full-year and first-quarter ones, and nothing is computed from them here. The piece is one of the outlet's Sunday analyses, and its expectation that the expense will reach billions of dollars after a listing is its own projection.

The two totals are a billion dollars apart; one is a ceiling and the other a floor. They are printed side by side, not reconciled. Their closeness is not taken here as confirming the estimate's method, and the partner-level figures below differ from the estimate's rows in several places.

The same book, as the draft itemises it

What has to happen before any of it is built

Announced capacity becomes capital expenditure only where it can be sited, and the count of places refusing is itself contested. Two organisations have counted local restrictions on data-center construction, using different methods and publishing a month apart. Both are given here.

Behind the larger count: more than 150 towns and counties passed bans in July alone, many in emergency meetings, against more than 300 active bans when the same tracker was first published in late June. More than 400 of the active restrictions are in the Midwest and South. Two states moved to the same posture from opposite parties — New York's governor signed an executive order in July pausing approval of new hyperscale data centers drawing 50 MW or more, the first statewide moratorium; Texas ordered a pause on new construction pending an audit of the strain on utility systems, with the state evaluating grid-connection requests totalling more than 474 GW. Massachusetts and Nebraska walked back builder incentives. The WSJ's September column adds the industry's own account of why: nondisclosure agreements that keep a data center's arrival from the public, and, in two operators' words, "fly-by-night" entrants. The two counts above measure the same object and are not reconciled here.

A county's side of the ledger. Bloomberg's October 12 feature gives the fiscal exposure of the place with the heaviest concentration of data centers. In Loudoun County, Virginia, more than 40% of tax receipts — $1.2 billion in fiscal 2026 — come from data centers. The county set up a $120 million revenue-stability fund in 2024 against a collapse in that revenue, and is drafting a plan for data centers that shut down, in a crisis or through obsolescence, that could require operators to post a decommissioning bond to cover the costs. The figures are the county's, as Bloomberg reports them.

Two figures for money already spent

Against announced commitments and projections, two figures for outlays actually occurring. Goldman Sachs calculates that spending on the equipment AI runs on — chips, servers, data centers, cooling and power — is running roughly $500 billion a year above its 2022 level. Census Bureau data put data-center construction alone at an annual rate of more than $75 billion, nearly 60% higher than a year earlier. Both reach this page through The Economist of September 4, in an article whose subject was employment rather than capital; neither is a company disclosure, and neither is added to anything above.

The build-out against the size of the economy

A third kind of figure measures the build-out against the U.S. economy rather than any company's books. The three below reached this page through one WSJ article of September 24, and none is a disclosure.

The Journal hedges the largest of these itself: projecting investment is tricky, it says, and total spending might end up substantially lower. On its account no single industry's build-out has taken so large a share of the U.S. economy, and the railroad boom of the late nineteenth century was the last time one industry accounted for more than the 1.9% Goldman expects this year. It also reports that a growing share of the five companies' spending is financed with debt, and that the Federal Reserve's chair, Kevin Warsh, has named borrowing by those companies as one reason long-term interest rates are up.

The three are not the same measurement and are not combined. The $10.3 trillion is an eight-year average for the whole economy; the 1.9% is one year, from a different forecaster; the $4.2 trillion is five companies' total capital spending as analysts forecast it, not an AI-only figure. Those five are a subset of the WSJ's nine in Entry 01, and their forward four years are not set against the nine companies' trailing ~$600 billion. None of the three is set against Goldman's $500 billion a year or the Census rate above, which measure spending already occurring.

The same economist's paper yields a fourth figure, of a different kind again. In a column of October 2 the Journal's Greg Ip reports Van Nieuwerburgh's estimate of the revenue AI would have to earn to justify the data centers now planned: $3.5 trillion in 2032, or 8.8% of GDPest. He started from builders' actual plans and the average cost of a data center, and assumed that some of those centers are cancelled, that each dollar of revenue yields 50 cents of cash flow, that builders need an unleveraged return of 10% on the capital invested, and that GDP grows 4% a year in nominal terms. The assumption he singles out himself is price. The figure supposes that today's prices for computing power, which scarcity is holding up, still apply in 2032 after capacity has roughly quadrupled. Ip sets beside it the price of bandwidth on fiber between London and New York, which fell 96% between 1997 and 2001.

This is a requirement derived from stated assumptions, not a forecast, and it reaches this page through a column. It is revenue in a single year, where the $10.3 trillion is investment over eight. Neither is divided into the other, and it is not set against any revenue figure in the box below.

A second requirement came on October 14, from a different starting point. The Economist cites researchers at Goldman Sachs: for the cloud giants to earn a return of 10% to 20% on what they invest in data centers this year and next, and for AI developers to earn a 30% margin, some $1 trillion of annual revenue is requiredest. It rests on different assumptions and a different horizon from the $3.5 trillion, and the two are not compared.

Sources for this entry: "How Anthropic Clinched $517 Billion in Compute Deals in 11 Months," The Information, September 6, 2026; "Data Center Bans Top 500 as New York, Texas Join Pushback," The Information, August 9, 2026; "Data Centers Will Cost Trillions More," The Wall Street Journal CIO Journal, September 8, 2026, relaying PwC's Global Data Centre Outlook; "How to Build AI Data Centers People Won't Hate," The Wall Street Journal, September 5, 2026; "OpenAI Considers Pre-IPO Funding Round at More Than $1.2 Trillion Valuation," The Wall Street Journal, September 16, 2026; "Anthropic signs first Australia data centre agreement," Reuters, September 16, 2026; "Anthropic Pursues IPO Despite Its A.I. Safety Warnings," The New York Times, September 18, 2026; "Anthropic considers releasing new AI model ahead of IPO, sources say," Reuters, September 18, 2026; "OpenAI Projects Burning Through $278 Billion by 2030, FT Says," Bloomberg, September 18, 2026, relaying the Financial Times; "The jobs apocalypse is postponed. An AI jobs boom is here," The Economist, September 4, 2026; "The AI Build-Out Is Becoming the Biggest Economic Bet in U.S. History," The Wall Street Journal, September 24, 2026; "Cracks in Oracle's AI Data-Center Build-Out Appear in Massive New Mexico Project," The Wall Street Journal, September 26, 2026 (the Oracle row); "Anthropic's IPO prospectus shows sweeping AI vision, surging costs," Reuters, September 28, 2026; "Anthropic's $518 billion AI buildout hinges largely on deals that cannot be canceled, filing shows," Reuters, September 29, 2026; "Anthropic says rogue AI agents pose uncertain legal risk for the company," Reuters, September 29, 2026. "Will America Spend 9% of Its GDP on AI? The Industry Is Counting on It," Greg Ip, The Wall Street Journal, October 2, 2026 (a column); "Anthropic Targets November IPO After Delays, Eyes SpaceX-Sized Debut," Bloomberg, October 1, 2026; "OpenAI Targets $30 Billion in Funding at $1.4 Trillion Value," Bloomberg, September 29, 2026; "Anthropic's Big Charity Bill for Shareholders," The Information, October 4, 2026; "How SpaceX's AI Unit Turned Itself Into an AI Cloud Firm," The Information, October 1, 2026; "Anthropic Discloses up to $84.5 Billion in SpaceX Compute Agreements," The Information, September 29, 2026 (a briefing). "Anthropic IPO prospectus lays bare deep dependence on Big Tech partners," Echo Wang and Krystal Hu, Reuters, September 29, 2026; "OpenAI Revenue Run Rate Approaches $70 Billion, Axios Reports," Bloomberg, September 29, 2026. From Edition 10: "OpenAI Computing Chief to Depart," Anissa Gardizy, The Wall Street Journal, October 14, 2026; "Anthropic's Quarterly Revenue Jumped 45% to Nearly $17 Billion," Kate Clark, Angel Au-Yeung and Corrie Driebusch, The Wall Street Journal, October 14, 2026; "Anthropic's Rapid But Slowing Growth Could Weigh on IPO Valuation," Cory Weinberg, The Information, October 14, 2026; "Investors temper Anthropic IPO valuation expectations over revenue growth slowdown, sources say," Milana Vinn and Anirban Sen, Reuters, October 15, 2026; "The AI boom is threatened by an epidemic of unreliable numbers," The Economist, October 14, 2026; "How the AI Boom Is Fueling Global Financial Risks," Bloomberg, October 12, 2026 (the Loudoun figures). Every figure in this entry is an estimate, a forecast, a third-party count or, from Edition 7, a figure from a draft that has not been filed publicly. None of it comes from a public filing, and none of it is combined with the figures in Entries 01 through 10.

12What the borrowing rests on

Entries 01 through 11 count what has been committed. This one asks what the borrowing behind those commitments is secured against. Items published on September 15 and 16 answer in different registers: a private-equity strategist's view, on the record; an account of a government order that the government denies, set down with the denial beside it; and the reported terms of a loan secured by the chips themselves. A report of September 26 adds the other side of the question: what lenders hold when the project behind a loan stalls. Two reports of September 30 and October 4 add a price: what lenders and bond buyers are now charging for it. Who has promised to pay if a borrower or tenant cannot is a separate question, taken up in Entry 13, and what the market charged in mid-October is in Entry 14.

How much of it is borrowed

MIT Technology Review's September 15 feature relays Morgan Stanley's estimate that more than half of the $2.9 trillion expected to be spent on data centers from 2025 to 2028 will be financed externally rather than out of the builders' own cash. It sets that beside a cash-flow figure: Alphabet posted a free-cash deficit of about $5.9 billion on roughly $120 billion of quarterly revenue, which the magazine says is its first since 2004 (the negative free cash flow already noted in Entry 08). The $2.9 trillion is a bank's projection, not a total of filings, and is not added to the $3 trillion in Entry 01, which measures something else.

The companies' own bond sales have been counted twice since, on different rosters. The Information (September 30, below) has Amazon, Google and Microsoft at nearly $160 billion of investment-grade debt. CNBC, citing LSEG, puts bonds issued this year by Alphabet, Amazon, Meta and Oracle at more than $228 billion as of October 13, more than double their total for all of 2025; all but Oracle are rated investment grade by S&P. Microsoft is in the first roster and not the second; Meta and Oracle are in the second and not the first. The two are not compared, and neither is set against the off-balance-sheet totals in Entry 01.

The same article describes the vehicles that route the exposure there: special-purpose structures like the one at Hyperion in Entry 04, which the author, David Rotman, flags in passing: "So-called special purpose vehicles (SPVs) are back!"

What the lenders are looking at

Speaking on Bloomberg Television on September 15, Jason Thomas, Carlyle's head of global research and investment strategy, said that lenders to AI projects often rely on the investment-grade credit of the sponsor rather than forming a view of the project itself. He compared that to 2005–07, when lenders relied on a bank's balance sheet rather than the mortgage collateral behind it, until the liabilities outgrew the balance sheet. He qualified it in the same breath: the pre-crisis condition does not hold today for investment-grade sponsors, but could as their liabilities grow and the range of possible data-center revenues widens. He also said risk premiums on AI exposure in both credit and equity have been shrinking, and that he considers a slowdown more likely than not. The write-up does not say a slowdown of what.

No deal, lender or figure is named in the interview. It is a view, not a disclosure, and is recorded here as one.

On October 1 the same firm put a number on the lenders' side. A Carlyle white paper, as Bloomberg reports it, estimates that private credit may need to supply roughly $1 trillionest to finance AI computing infrastructure, which Bloomberg says is more than half of the assets the industry now manages, and warns that failing to set clear limits on concentration in AI compute could prove its largest mistake. Mark Jenkins, Carlyle's co-president and head of global credit and insurance, said on Bloomberg Television that some of the firm's own investors are asking for limits of roughly 8% to 10% on these allocations, on the syndicated-loan side, and that financings which look diversified depend in the end on seven or eight high-quality names. Carlyle is not staying out, he said; it wants the exposure in balance. Like Thomas's remarks this is a view, here from a lender describing its own market, and it is recorded as one.

A price curve, taken down — or not

The second item concerns a narrower kind of collateral: the chips themselves. Two accounts follow, and they cannot both be right.

Whichever account holds, the reporting identifies a link between the figures on this page: a published price for renting a chip is also, for lenders, a price for the collateral behind loans like CoreWeave's. CoreWeave appears elsewhere on this page as a company Nvidia holds a stake in (Entry 07) and as a developer whose figures are not broken out (Entry 10). No loan figures for it appear in either source, and none are supplied here.

A loan secured by the chips

A day later Bloomberg reported a loan of exactly that kind, with its collateral named. A group of 10 banks is providing a $22 billion loan to Crux AI, the cloud venture Blackstone and Alphabet launched the week before (known internally as Project Braid), to buy Google's tensor processing units, or TPUs. Per the people Bloomberg spoke to, the debt is backed by the value of those chips and by Crux AI's customer contracts.

Bloomberg places the deal beside two others structured around the same chips. A group of lenders including Apollo and Blackstone provided a $35 billion debt package so that Anthropic could lease TPUs, in a transaction backed by Broadcom, which helps Google build them; a similarly structured deal of more than $60 billion is in discussion, its parties unnamed. A $60 billion chip financing arranged by Broadcom’s banks has since been reported with its parties named, and is set down in Entry 13; that it is the deal left unnamed here is likely but is not stated in the reporting, so this sentence stands as written. The three amounts describe different structures — a borrower buying chips, a lab leasing them, a deal not yet done — and are not added together. The $35 billion package is not placed in the Google row of Entry 11's counterparty ledger, because the reporting does not say it is the same arrangement.

The same article mentions Nvidia's August partnership with six financial firms, Blackstone among them, to help customers pay for its chips, which it puts at $500 billion. That is the announcement Entry 07 records from CNBC as partnerships to mobilise "more than $500 billion"; it is counted once, there. Crux AI's equity is not set against its debt as a ratio here: the equity is described as initial, and the lending group is not final.

Two more financings for chip purchases were reported in October, neither yet done. Bloomberg's credit column of October 10 says SpaceX is in talks with banks and investors to raise $40 billion to buy chips from Nvidia, which it calls among the biggest debt financings yet for the build-out, expected to close next year; and that Broadcom is sketching another deal to help OpenAI buy the custom chips the two are developing. The column does not describe the collateral. On October 12 Bloomberg reported that Tencent told bond investors its planned offering of about $5 billion, in dollar and offshore-yuan notes, would mainly finance AI capital spending, most of it on chips, per people familiar; its capital spending on AI projects and computing rose 176% to 52.8 billion yuan in the June quarter. That is a corporate bond, not a loan secured by chips. Tencent is the first borrower on this page from outside the U.S. companies and the frontier labs, and it is in none of the page's rosters. A loan whose chip-secured part the banks would not underwrite is set down in Entry 14. None of these amounts is added to the three above or to each other.

A lease that cannot be cancelled

Around a year ago OpenAI agreed to buy $300 billion of cloud computing from Oracle over five years. Oracle did not have the data centers to deliver it, so it leased new campuses in Texas, Wisconsin, Michigan and New Mexico, about 4.5 GW in all, branded as part of OpenAI's Stargate initiative, and banks lent tens of billions of dollars to build them. On September 26 the Journal reported on the New Mexico site, Project Jupiter in Doña Ana County. Per people with knowledge of the deal, Oracle's lease there is on hell-or-high-water terms: it cannot be terminated, and Oracle owes rent whether or not it has secured the power to run the buildings. It is a specific case of the Journal's general description in Entry 01, that signed leases “can't be canceled, for the most part.”

What stalled. Oracle replaced its original plan to power the site with gas turbines and diesel generators, which met local resistance, with Bloom Energy fuel cells, but it does not yet have the state air permits it needs, and a short stretch of a proposed 17-mile gas pipeline that would cross state land has been rejected twice. In recent months, per people familiar, it removed BorderPlex Digital Assets, a land and power partner on the project since late 2024, which is to be paid to give up its rights. In the week of the report Oracle served a force majeure notice on the developer, Stack Infrastructure, owned by Blue Owl funds, pushing back the date it begins paying full rent. The first phase had been due in the third quarter of 2028.

What the notice changes, and what it does not. Deferring rent until the power comes on automatically extends the lease, so on the Journal's account the total Oracle pays Stack does not fall: the notice buys time without reducing Oracle's exposure. The developer is covered in the meantime. Stack is earning Blue Owl the equivalent of a 9% annualised return, paid out of the money raised to build the project, and the financing set aside funds for up to three years of delay, after which the developer's forecast return rises to 11%.

The two accounts are printed as given. On one point they agree: the financial commitment is unchanged.

The tenant's credit. S&P cut Oracle's rating in July to one notch above junk, saying its rapidly expanding AI infrastructure business is increasing its overall credit risk, and the Journal notes that Oracle's capital spending on AI equipment and property has exceeded its operating cash flow for several quarters. Oracle is one of the nine companies behind the $3 trillion in Entry 01, not broken out there; the contract behind this lease now fills its row in Entry 11's counterparty table; and the developer's owner, Blue Owl, is the firm whose vehicle owns Hyperion in Entry 04. Most banks hold construction loans to maturity without marking them to market, per the Journal, so the $1.8 billion is a signal from the loans that traded, not a figure from anyone's books.

What the debt now costs

Four days after the Journal's New Mexico report, The Information described what has happened to the price of this debt. On September 30 it reported that lower-rated borrowers are having to offer concessions to sell bonds for data centers. CleanSpark, a bitcoin miner turned developer that is building a data center for Meta, sold $2.3 billion of bonds earlier in the month at 98.5 cents on the dollar, one of the largest discounts of the past year, with a coupon of 7.875% and an agreement to repay principal over time. In mid-August Zenith Arc, a developer backed by a Coatue venture and Fluidstack, sold debt at 99.5 cents to finance a data center to be leased by the trading firm Jane Street. The outlet notes that the end user in the first case is an investment-grade company, which has usually been enough to reassure buyers.

Where the supply comes from. The largest data-center operators, of which the article names Amazon, Google and Microsoft, will together spend some $700 billion on capital expenditure this yearest and have issued nearly $160 billion of investment-grade debt. Several are also moving spending onto other ventures that borrow in the high-yield market, where developers building for Anthropic and OpenAI are borrowing too. The discount counts above are Morgan Stanley's, and the bank led 13 of the 20 deals it counts. The $700 billion is given for a group the article does not list in full, and is not set against any other capital-spending figure on this page.

The banks. Per people arranging the deals, some lenders are becoming more selective, which narrows the pool available to syndicate large loans: one person named Société Générale and Sumitomo Mitsui Banking Corp., and a second named Mitsubishi UFJ Financial Group. The article notes what each had done before. Société Générale led a $7.1 billion financing for the first Stargate site, MUFG led with JPMorgan Chase a $38 billion financing for two more Oracle projects, and SMBC helped lead the $18 billion New Mexico loan described above.

Both things are happening at once: wariness about AI projects in particular, and a general rise in yields. Bloomberg reported on October 4 that in the previous week the 30-year Treasury yield reached 5.69% and the 10-year passed 5.3%, levels neither had reached since 2002. The Fed chair's remark that heavy borrowing by technology companies is one reason long-term rates are up is already recorded in Entry 11. No yield, spread or cost of borrowing is computed here from the prices and coupons above.

Sources for this entry: "What's at stake in AI's trillion-dollar gamble," MIT Technology Review, September 15, 2026; "Carlyle's Thomas Warns AI Financing Resembles Pre-Crisis Mortgage Model," Bloomberg, September 15, 2026; "Trump administration orders Kalshi to scrap AI price tracker over national security concerns," Semafor, September 15, 2026; "Blackstone, Alphabet Secure $22 Billion AI Chip Loan Backed by Major Banks," Bloomberg, September 16, 2026; "Cracks in Oracle's AI Data-Center Build-Out Appear in Massive New Mexico Project," Peter Rudegeair, Anissa Gardizy and Matt Wirz, The Wall Street Journal, September 26, 2026. "AI Buildout May Spur $1 Trillion in Private Credit, Carlyle White Paper Finds," Bloomberg, October 1, 2026; "Cracks Emerge in AI's Debt-Fueled Data Center Boom," Dakin Campbell, The Information, September 30, 2026; "Wall Street's AI Party Is on Edge as Soaring Yields Raise Risks," Bloomberg, October 4, 2026. From Edition 10: "CoreWeave stock sinks as bond yields surge near 15%," CNBC, October 14, 2026 (the issuance figure); "AI Debt Issuance Defies Rising Yields as SpaceX, Broadcom Plan Major Deals," Bloomberg, October 10, 2026; "Tencent's $5 Billion Bond Offering Targets AI Spending, Chip Purchases," Bloomberg, October 12, 2026. Nothing in this entry comes from a filing, and nothing in it is added to any figure elsewhere on the page.

13Who stands behind it

Entry 12 asks what the borrowing is secured against. This one asks who has promised to pay if the borrower or the tenant cannot. Reports published between September 30 and October 2 describe two versions of that promise: a chip supplier lending to its own customer and supporting the value of the chips it sold, and a large company guaranteeing the lease of a smaller one so that a landlord will build. In each, the party making the promise also sells to, or invests in, the party it is made for. And in most cases the promise is recorded, if it is recorded at all, in a footnote.

The supplier as lender

On October 1 Reuters reported a further term from Anthropic's draft prospectus, the confidential document Entry 11 already reads. Broadcom has agreed to lend Anthropic up to $42 billiondraft to finance infrastructure spending. The debt would take the form of notes that can be converted into Anthropic shares; Broadcom may designate a financing partner; and Anthropic says in the filing that it does not expect any notes to be sold before it completes its IPO. On Reuters' account the facility could finance about a third of a $125.2 billion commitmentdraft Anthropic has made for a five-year lease of TPU computing capacity, the chips Google and Broadcom have developed together.

Reuters adds that Anthropic is expected to become Broadcom's largest compute customer in 2027, and that Broadcom projects AI semiconductor revenue of about $115 billion in fiscal 2027 and $230 billion in fiscal 2028. The lender's largest expected customer is the borrower. An analyst Reuters quotes, Jay Goldberg of Seaport Research, describes Broadcom as following Nvidia in putting its balance sheet behind its chip sales, the practice recorded for Nvidia in Entry 07. "About a third" is Reuters' characterisation; no ratio of the loan to the lease is computed here.

A $60 billion syndication, and the seller behind its senior notes

A day later Bloomberg reported the banks' side. Per people with knowledge of the matter, Broadcom's banks, among them Bank of America, Citigroup and Morgan Stanley, are starting to syndicate $60 billion of chip financing that would benefit Anthropic and other companies. The deal has not been announced. It has two parts: a $42 billion Class A senior-secured tranche, and an $18 billion Class B junior tranche led by Blackstone, which is committing $9 billion from its own funds, with the remainder to be syndicated. Broadcom will provide what the people call residual value support, which effectively backstops the senior tranche. The company that sells the chips is standing behind what they will be worth to the lenders.

As with an earlier Broadcom chip financing, which the article does not name, part of the senior notes is likely to be refinanced later in the bond market. Bloomberg says the financing is being watched on Wall Street as a test of whether investors will still fund the build-out during a public backlash against data-center construction. Broadcom and the banks declined to comment. How much of the $60 billion is for Anthropic is not stated, and no pricing, term or detail of the support is given.

Neither report mentions the other's figure, and neither says the two are the same money. They are printed side by side, not reconciled, and neither is added to the $18 billion.

Two figures already on this page bear on these, and are likewise left as they stand. Entry 12 records, from Bloomberg's September 16 report, a deal of more than $60 billion then in discussion with its parties unnamed, and a $35 billion package from lenders including Apollo and Blackstone so that Anthropic could lease TPUs, backed by Broadcom. The October 2 article links back to an August Bloomberg report that Broadcom was seeking more than $60 billion in a new AI debt deal. That the syndication is the deal Entry 12 left unnamed, and the $35 billion package the earlier financing, is an inference from those links and not a statement in either article, so the entries are not merged. Second, the draft's table in Entry 11 carries Broadcom-related equipment leases of about $161.2 billion, from Reuters' September 29 report, where the October 1 report gives $125.2 billion for a five-year TPU lease and does not mention the earlier figure. Both are Reuters' readings of one draft. Whether one sits inside the other is not stated, and both are printed.

A guarantor as the price of a lease

The second kind of promise was described by Bloomberg on September 30. Brokers at CBRE, the largest commercial real-estate services firm, are telling the newer AI cloud companies — neoclouds, which rent out access to AI chips — that a partner with a strong credit rating to guarantee their lease payments is now a precondition, and several said they will enter serious talks only when one is in place. Developers and banks doubt that entrants with speculative credit ratings will last long enough to pay for scarce power and land. Vacancy in core markets fell to a record low of 1.4% in the first half of 2026, per CBRE, which lets data-center operators choose their tenants. The firm says it has no blanket ban on neoclouds.

Bloomberg gives one such guarantee with a price on it. When Fluidstack, a neocloud that provides access to Google's chips, leased a campus in Western New York from TeraWulf last year, Google provided a $3.2 billion guarantee and took warrants in TeraWulf, a former crypto miner. Per people familiar with the matter, investors were not told at the time that Anthropic would be the end user of the chips. Per one of them, Google's commitment let TeraWulf issue notes to finance the building at 7.75%, where it would otherwise have paid 10% to 12%.

Fluidstack and TeraWulf each have a row in Entry 11's counterparty table, taken from The Information: Fluidstack as an initiative to build its own sites in Texas and New York, TeraWulf as a direct lease in Kentucky. Bloomberg describes a Fluidstack lease from TeraWulf in New York with Anthropic behind it. The accounts concern overlapping parties and are not merged into one row; whether the New York campus is inside either existing row is not stated.

What a guarantee covers, and where it is recorded

The form of the promise is changing. Eamon Nolan, a project-finance lawyer at Vinson & Elkins, told Bloomberg that the large guarantors are avoiding commitments to pay all of a tenant's future rent and moving to residual value guarantees: a promise that the data center will keep some value if the tenant leaves. Many are triggered only when a default is imminent. That is what can keep them off a balance sheet for years. A contingent liability does not have to be recorded there, David Trainer of the research firm New Constructs told Bloomberg, and can be disclosed in the footnotes of quarterly and annual reports instead; Bloomberg identifies him as a former member of an advisory committee to the accounting standard-setter. The footnotes are where the Journal found the $3 trillion in Entry 01. Guarantees are a different kind of obligation from the leases and purchase commitments counted there, and nothing in the reporting says they are inside that total.

Bloomberg gives the terms of the largest, citing a filing. Nvidia's commitment at OpenAI's Ohio campus, leased from SB Energy and put at up to $105 billion (Entry 07, Entry 10), does not guarantee the full cost of the site or all of the tenant's obligations. Nvidia will provide support if the value of a replacement lease, or of a sale of the data centers, falls below a set level. Its guarantees cover an initial 4.25 GW, with an option to support about 3.75 GW more, and are triggered only by a specified tenant insolvency or an uncured payment default under a lease.

A fourth ledger: the backstops named on this page

For scale, the same Bloomberg article cites a CreditSights report from earlier this year projecting that the top five hyperscalers will spend roughly $750 billion in 2026est, and observes that guarantees add obligations on top of that spending. The article does not name the five, so the figure is not set against the WSJ's nine in Entry 01 or the five in the analyst consensus in Entry 11. It also reports a worry it attributes to some in the industry without naming them: that backstops prop up companies that would not survive alone, at a time when investors are already uneasy that circular financing of the kind in Entry 07 could deepen losses in a downturn.

A backstop declined. On October 15, per Bloomberg, Nvidia's chief executive Jensen Huang said the chipmaker, one of CoreWeave's suppliers and, per Entry 07, one of its investors, did not need to backstop the company, noting that its cost of capital “is actually decreasing.” It is a statement that a backstop is not needed, not a disclosure about whether any exists, and CoreWeave is not added to the table above. The prices behind the question are in Entry 14.

Sources for this entry: "Broadcom to lend Anthropic up to $42 billion to lease its chips, filing says," Echo Wang, Milana Vinn and Max A. Cherney, Reuters, October 1, 2026; "Broadcom Amassing $60 Billion to Fund Chips for Anthropic," Bloomberg, October 2, 2026; "AI Startups Face Hurdles Securing Data Centers Without Big-Tech Lease Guarantees," Bloomberg, September 30, 2026; from Edition 10, "CoreWeave Credit Risk Hits Record as Default Odds Top 60% on Borrowing Plans," Bloomberg, October 15, 2026 (Huang's remark). The Reuters figures are from a confidential draft; the Bloomberg figures are from people familiar and, for Nvidia's guarantee, from a filing Bloomberg cites. Nothing in this entry is added to any figure elsewhere on the page.

14What the market is charging

Entry 12 asks what the borrowing is secured against, and Entry 13 who has promised to pay if a borrower cannot. This one records what lenders and bond buyers charged for it in one week of October, at the most indebted of the companies that rent out AI computing and at a new one borrowing for the first time. Every figure here is a market quotation from a particular day, and the next day's replaces it. None is a disclosure.

CoreWeave's debt, two days running

CoreWeave is the neocloud whose stake Entry 07 records in Nvidia's book and whose figures Entry 10's developer ledger leaves blank. On October 14 its shares fell 15.5% and, per CNBC, the yield to maturity on its 9.625% notes due 2032 reached 14.8%, and on its 9.75% notes due 2031 about 14.9%. S&P rates the company B+, in junk territory. An S&P Global Market Intelligence pricing executive told CNBC it is natural, with rates rising, that highly leveraged companies such as CoreWeave are affected.

The next day Bloomberg reported that the cost of protecting CoreWeave's debt against default for five years rose by as much as 74.5 basis points to 1,194.5 basis points, according to ICE Data Services, the highest intraday level on record, with the swaps “pricing in a more than 60% chance of default within the next five years.” Its 9% notes due 2031 fell about half a cent to 82.75 cents on the dollar, a yield of 14.5% (Trace), while its 9.75% notes due 2031 rose 2.25 cents. Bloomberg ties the move to an interview the company's co-founder, Brannin McBee, gave the Financial Times: the goal, he said, is investment-grade ratings on more of its debt and ultimately on the company, to gain access to “unlimited” pools of insurance money. Bloomberg's reading is that the remarks underscored how much more it plans to borrow. The shares bounced after Nvidia's chief executive said the chipmaker did not need to backstop CoreWeave (Entry 13).

CNBC and Bloomberg quote different notes on different days, so the yields above do not conflict; each is printed with its coupon and date. The 60% is Bloomberg's statement of what the swap prices imply. It is not a rating or a forecast, and it is not computed here. No yield, spread or probability is derived from any price in this entry.

Bonds for the buildings

The same Bloomberg report follows the pressure into bonds sold to build data centers tied to CoreWeave. A $3.5 billion note that Galaxy Digital sold in July to fund AI computing capacity in Texas fell for a third straight day to 90.5 cents on the dollar, down 8.5% this month. The $3.3 billion of bonds Core Scientific issued in April to fund CoreWeave-leased data centers in five states traded at 92.52 cents after six straight declines. These are trades in bonds already outstanding. The issue prices in Entry 12, such as CleanSpark's 98.5 cents, are a different kind of price, and the two are not compared.

A loan the banks would not fully hold

On October 12 Bloomberg reported a $5 billion loan offering for Volta Infrastructure Holdings, a neocloud founded this year, led by JPMorgan and Goldman Sachs. The banks underwrote a $1.3 billion term loan but not the $3.7 billion chip loan, which per people familiar with the deal was a move to protect themselves; Bloomberg calls it a rare one. The term loan pays for leasing a data center in Norway that Volta will run with Bitdeer, a bitcoin miner. The chip loan buys about 36,000 Nvidia GPUs. Anthropic has signed a six-year agreement for compute at the site, per earlier Bloomberg reporting.

It is the second loan on this page whose collateral is the chips it buys, after Crux AI's in Entry 12. In Crux's, the banks lent and then syndicated; in Volta's, they declined to underwrite the chip-secured part at all. Volta is not a row in Entry 11's counterparty ledgers, which come from other sources, and its six-year agreement with Anthropic is not added to them. The two loans are not added together.

Sources for this entry: "CoreWeave stock sinks as bond yields surge near 15%," CNBC, October 14, 2026; "CoreWeave Credit Risk Hits Record as Default Odds Top 60% on Borrowing Plans," Bloomberg, October 15, 2026, which relays an interview in the Financial Times that was not read here; "JPMorgan, Goldman Limit Risk on Volta's $5 Billion AI Loan Offering," Bloomberg, October 12, 2026. Every figure is a market price from the day stated, or loan terms from people familiar. Nothing in this entry comes from a filing, and nothing in it is added to any figure elsewhere on the page.