The AI money loop: who funds whom, and where it's cracking

Disclaimer: Not investment advice. This is an essay about how the AI infrastructure buildout is being financed, and about how to read the numbers around it. It assumes no background in finance; technical terms are defined where they first appear. Figures are from company filings, central banks, the IMF and the press, as of 19 October 2026. Where a number comes only from press reports and no company has confirmed it, I say so. Lab revenue figures are always run-rates, and I say how each lab counts them. Links to every source are in the text and collected at the end.

Two screens on 14 October

On Wednesday 14 October 2026, anyone watching AI stocks needed two screens to make sense of the day.

On the first, CoreWeave, the largest of the specialist AI cloud companies, fell 15.5% to $65.78. Its bonds due 2032, which pay a 9.625% coupon, were trading below 80 cents on the dollar, a price that implies a yield of roughly 14.5% to 15%, according to Yahoo Finance. That's the kind of yield the market charges companies it thinks might not pay it back. Nebius, Applied Digital and IREN, three smaller AI clouds, sold off with it, and Reuters reported Oracle down about 9%. The day before, the IMF had published its Global Financial Stability Report, which counts a sharp AI correction among the three likeliest triggers for a global market shock.

On the second screen, nothing had changed. TSMC, which manufactures nearly every advanced AI chip, was reporting record results the same week. Seven weeks earlier Nvidia had reported $96.2 billion of revenue for a single quarter, guided the next one to $108 billion, and told investors to expect its revenue to grow by about 70% in the following fiscal year, against a Wall Street consensus near 44% (Nvidia; CNBC). Its chief financial officer described demand as "supply-constrained."

Here's the odd part. These are not two different industries. CoreWeave buys almost all of its chips from Nvidia. Nvidia owns roughly a tenth of CoreWeave, is one of its customers, and has promised to buy any computing capacity CoreWeave can't sell until April 2032. One company's best quarter and the other's worst day are, in large part, the same money moving through the same pipes.

The rest of the week made the point harder to miss. On 15 October the price of insuring CoreWeave's debt against default for five years reportedly hit a record 1,194.5 basis points, according to ICE data cited by TNW: an insurer wanted almost 12% of the sum covered, every year. The yield on its 2031 bonds briefly passed 16%. The stock touched $59.51 before Jensen Huang, Nvidia's chief executive, said in Stockholm that CoreWeave needed no backstop and that its funding costs were falling; it closed at $69.69. The next day a hot US inflation reading revived bets on a December rate rise, and CoreWeave fell 11.9% to $61.39 (Reuters via The Star). It lost close to a quarter of its value in a week, with no company news behind any of it. It listed at $40 in March 2025 and peaked at $187 that June; it is now 67% below that peak.

That's what people mean when they talk about the "circular financing" of AI: the companies that sell the equipment invest in, lend to, guarantee and rent from the companies that buy it. The phrase usually comes with an accusation, that the whole boom is a merry-go-round of the same dollars counted several times. I wanted to know whether that's true, so I went through the filings, the deal announcements and the credit data, and built a ledger of every arrangement I could verify.

The short answer is that the loop is real, it is very large, and it is not a Ponzi scheme. The longer answer is the rest of this essay.

Fig. 12026: from AI demand to AI finance

The short version

The finding that organises everything below is that the AI money loop doesn't have one risk profile. It has three layers, and they are behaving in opposite ways.

The top is accelerating. Nvidia, TSMC, Microsoft and Alphabet have audited revenue, enormous cash flows and, in Nvidia's case, demand it can't fully supply. When people describe the loop as a fraud, this is the layer that refutes them.

The debt-funded middle is repricing. Oracle, the specialist clouds (CoreWeave above all), the special-purpose vehicles that own data centres on behalf of tech giants, and SoftBank all borrow against someone else's contract. Every visible crack of 2026 opened here. The weakest link is the specialist clouds, or neoclouds: GPU landlords that own almost nothing scarce, depend on one to three customers, and keep the risk that the chips are worth little when the contracts end.

The demand base is under-covered. OpenAI has signed roughly $730 billion of computing contracts against a revenue run-rate of about $50 billion. Behind the labs sits the number that is supposed to pay for all of it eventually, what businesses will pay for AI, and it's the least well-measured number in the whole system. In 2026 it started to wobble: big employers began capping staff AI spending, and business spending on AI tokens went flat for two months while usage kept rising.

Two other findings run through the essay. The first is that the loop migrated from cash to contingent credit during 2026. The most famous circular deal, Nvidia's $100 billion for OpenAI, never happened. What replaced it is a $30 billion cheque and up to $105 billion of guarantees on 20-year leases, which is bigger, quieter and harder to see. The same pattern repeats across the system: equity became guarantees, ownership became leases, and balance-sheet spending became off-balance-sheet commitments.

The second is a measurement asymmetry. Every dollar flowing into AI vendors is counted, audited and reported quarterly. The value those dollars create for the businesses that use AI is self-reported, hard to isolate and, in the most rigorous studies, still small at the level of the whole economy. Markets price what they can count, so the loop looks healthiest exactly where it is easiest to measure.

The one number underneath

Every layer of the loop ultimately depends on one number: what end customers, businesses and consumers, will pay for AI. Nvidia's sales depend on the clouds' capex. The clouds' capex is justified by their backlogs. The backlogs are mostly the labs' contracts. And the labs can honour those only if their own customers pay more every year. CoreWeave's lenders, Oracle's bondholders, SoftBank's margin lenders and Nvidia's guarantees are all, at one or two removes, bets on that number.

For most of the boom it was a footnote. Investors rewarded capex for its own sake, on the view that demand would follow supply. In 2026 it became the headline, for three reasons. The spending outgrew the cash: JPMorgan Asset Management estimates AI capex will absorb about 93% of the big cloud companies' operating cash flow this year, against 33% in 2023 (JPMAM). More of it is borrowed, so lenders now hold fixed claims on revenue that hasn't arrived. And the biggest buyers started to behave like cost-conscious customers. From spring 2025, many large employers had practised what staff called tokenmaxxing: pushing employees to use as many tokens, the units AI models are billed in, as possible, with mandates, leaderboards and near-unlimited budgets. Between May and August 2026, Uber, Meta, Microsoft and Amazon, among others, reversed course and capped it.

That's why the question near the end of this essay, does any of this pay?, now carries the rest of it.

Key terms

A few words recur throughout.

• Circular financing. A seller provides money (equity, a loan, a guarantee or a purchase commitment) to a customer, which then uses it to buy the seller's products. It isn't automatically improper. It's sometimes called vendor financing, and it's how aircraft, telecom equipment and cars have long been sold.
• Run-rate. A company's latest month (or quarter) of revenue multiplied up to a year. It's not audited and labs calculate it differently. OpenAI counts sales made through Microsoft net of Microsoft's share; Anthropic counts sales through Amazon and Google gross. Both methods are permitted under accounting rules, but they produce different numbers for the same business.
• Backlog, or RPO. Remaining performance obligations: revenue a company has contracted but not yet delivered. A cloud provider's RPO is the sum of its customers' multi-year commitments.
• Capex and operating cash flow. Capital expenditure is money spent on long-lived assets such as data centres and chips. Operating cash flow is the cash the business generates from its normal operations. When capex exceeds operating cash flow, the gap has to be borrowed, raised from shareholders or taken from savings.
• Off-balance-sheet. An obligation that doesn't appear as debt in a company's accounts, such as a lease that hasn't started yet or the debt of a partly owned joint venture.
• SPV. A special-purpose vehicle: a separate company set up to own one asset (a data centre, a pile of GPUs) and borrow against it, so the debt sits outside the sponsor's accounts.
• Credit default swap (CDS). Insurance against a borrower defaulting. Its price, in basis points (hundredths of a percentage point) per year, is a running market estimate of default risk.
• Take-or-pay. A contract under which the customer pays whether or not it uses what it ordered.
• Warrant. The right to buy shares at a set price. A warrant at $0.01 a share is, in effect, free shares delivered when conditions are met.
• Neocloud. A specialist AI cloud, such as CoreWeave, Nebius, Lambda or Nscale, that rents out Nvidia GPUs rather than offering the full range of services Amazon or Microsoft do.
• EBITDA and depreciation. EBITDA is earnings before interest, taxes, depreciation and amortisation. Depreciation spreads the cost of equipment over its useful life as an annual expense. For a company whose main cost is equipment bought with debt, EBITDA leaves out the two biggest costs.

How a dollar goes round

The easiest way to see the loop is to follow one dollar.

In February 2026 Amazon agreed to invest $50 billion in OpenAI. On the same day, OpenAI expanded its contract with Amazon Web Services by $100 billion over eight years, taking its total AWS commitment to about $138 billion, including roughly 2 gigawatts of Amazon's own Trainium chips (Amazon 8-K). The investment was announced as $15 billion up front plus $35 billion conditional on an IPO or an "AGI milestone". According to Amazon's own quarterly filings, it had paid the whole $50 billion by the third quarter of 2026, without either trigger being publicly met: $15.0 billion in the first quarter, $13.7 billion in the second and the remaining $21.3 billion after 30 June (GeekWire). Amazon gave no reason for paying early.

So: Amazon sends OpenAI cash. OpenAI promises to spend it, and much more, at AWS. AWS records the contract in its backlog and, as the computing is delivered, as revenue. To deliver it, AWS buys chips, some designed by Amazon and some from Nvidia. Amazon owns a stake in OpenAI that is valued at whatever OpenAI's next funding round says it's worth.

Nvidia's version is similar. It put $30 billion into the same OpenAI round, which closed on 31 March at $122 billion and an $852 billion post-money valuation (CNBC). Amazon and Nvidia together supplied about two-thirds of it. OpenAI, in turn, has a contract worth about $300 billion over five years with Oracle, which is borrowing heavily to build data centres filled with Nvidia chips, and a $250 billion Azure commitment with Microsoft, which also buys Nvidia chips.

None of that is hidden, and none of it is fraudulent. It's simply concentrated. The Bank for International Settlements, the central bankers' bank, studied 1,246 AI firms and found that 55.2% of the money invested in AI firms between 2021 and 2025 came from other AI firms. Of the circular relationships it mapped, 64% were "invest in and supply" pairs, and 73% started with a compute or infrastructure supplier. The IMF reviewed 14 such transactions since 2023 and found about $300 billion of equity and $650 billion of debt, exchanged for $1.4 trillion of future revenue commitments (IMF).

Fig. 2The AI money loop, October 2026

EquityCompute contractBackstop or offtakeGuaranteeWarrantLoan or bonds

Paid or contractedAnnounced, in talks or reportedDead

Fig. 3Follow one dollar

The $100 billion that never was

The most quoted number in the whole debate is one that never became money.

On 22 September 2025 Nvidia and OpenAI announced a partnership to deploy at least 10 gigawatts of Nvidia systems. Nvidia said it intended to "invest up to $100 billion in OpenAI progressively as each gigawatt is deployed" (Nvidia). It was a letter of intent, not a contract. The widely repeated "$10 billion per gigawatt" was an analyst's division, not a term of the deal.

For a few months it was the symbol of the loop: a chipmaker promising to fund its biggest customer's chip purchases. Then it quietly fell apart. At the end of January 2026 the WSJ reported the plan had stalled. Jensen Huang, Nvidia's chief executive, said it "was never a commitment" (Fortune), and on 4 March he ruled out a $100 billion investment altogether, citing OpenAI's plans to go public (Bloomberg).

What replaced it came in two parts.

The first was a straight $30 billion equity investment in OpenAI's March round, alongside binding OpenAI commitments for 3 gigawatts of inference and 2 gigawatts of training capacity on Nvidia's next-generation Vera Rubin systems (Investing.com). At the round's valuation, that's a stake of roughly 3.5%.

The second only became visible in August, in a note to Nvidia's quarterly filing (Nvidia 10-Q). Nvidia had agreed guarantees with affiliates of SB Energy, SoftBank's power company, "capped at a total of $105 billion". They back the land, power and building shell of a campus in Pike County, Ohio, with about 4.25 gigawatts of computing load, built in nine phases. The tenant is an affiliate of OpenAI. Each phase is leased for 20 years, the first is expected in Nvidia's fiscal 2029, and the site "will exclusively host NVIDIA AI infrastructure." Nvidia also holds an option to back about 3.8 gigawatts more. The guarantees fall away if OpenAI obtains a satisfactory credit rating.

There is a third number floating around that needs placing. On 26 July the WSJ reported that Nvidia was in talks to guarantee about $250 billion of lease and construction debt for a 10-gigawatt SB Energy campus in southern Ohio. It's the same site. The signed terms a month later, a $105 billion cap plus the 3.8-gigawatt option, look like the outcome of those talks. The $250 billion is a reported negotiating figure, not a separate exposure, and it shouldn't be added on top.

Fig. 4The $100 billion that became $30 billion plus $105 billion

The contrast with what Nvidia said nine months earlier is striking. In November 2025, answering short-sellers, Nvidia sent analysts a memo saying it had "a single disclosed guarantee with immaterial fair value" and did not rely on vendor financing to grow revenue (Investing.com). Michael Burry, the investor famous for betting against subprime mortgages, called the August guarantee a circular-financing red flag (Benzinga).

Both readings miss something. The guarantee isn't revenue, and no cash leaves Nvidia unless the OpenAI affiliate stops paying rent. But it's 3.5 times the size of the equity cheque, it runs for 20 years, and it is attached to a single customer that is losing money. The structural lesson of the whole episode is that the headline deal died and was replaced by something larger and less visible: capped cash became contingent credit.

My reading: whenever you see a dead headline in this story, look for what replaced it. It's usually still there, in a footnote.

Nvidia, the central bank of AI

Nvidia now sits on every side of the table. It's the supplier to almost every AI cloud, an investor in many of them, a customer of several, a guarantor of others' leases, and, through new financing platforms, the organiser of other people's capital to buy its products. The comparison people reach for is a central bank: the institution that supplies the money the whole system runs on.

Start with the portfolio. Nvidia's holdings of non-marketable equity (stakes in private companies) rose from $3.24 billion in April 2025 to $22.3 billion in January 2026 and $51.2 billion on the balance sheet by July 2026. Add $42.8 billion of listed shares and $5.0 billion of locked-up public stock and the equity book comes to about $99 billion. In the first quarter of its fiscal 2027 alone, Nvidia put $18.6 billion into private companies and infrastructure funds, about 23% of that quarter's revenue. Its own filing notes that some investees "may indirectly purchase or use our products in the cloud." The gains on those stakes are now a material line in its earnings: $23.7 billion in the first half of fiscal 2027, partly from funding rounds in companies that buy Nvidia hardware.

Then the rent-back. In its July filing Nvidia disclosed a new $36 billion line of "AI cloud agreements", made under partnerships in which "clouds procure our data center infrastructure products and we commit to cloud service agreements". In plain terms: clouds buy Nvidia chips, and Nvidia agrees to rent computing back from them. Add $29 billion of ordinary cloud-service commitments and Nvidia had promised to rent about $65 billion of capacity as of 26 July, up from $26 billion in October 2025 and $27 billion in January 2026 (Nvidia 10-Q; 10-K).

Fig. 5Nvidia's portfolio went vertical

Finally, the capital it mobilises. On 10 August Nvidia signed memoranda with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR to set up platforms that would raise "over $500 billion" of outside money to finance Nvidia-based computing for its customers (Nvidia). These are announcements, not commitments, but the direction is clear.

Adding it up, carefully

If you stack everything, you get a striking number: about $99 billion of equity, $25 billion of equity commitments not yet funded, $65 billion of rent-back and $108.5 billion of guarantees ($105 billion for Ohio and $3.5 billion of smaller lease guarantees for AI clouds). That's roughly $300 billion of balance-sheet or contingent exposure, against fiscal 2026 revenue of about $216 billion.

That figure is an upper bound on exposure, and it's not a measure of "self-funded revenue." A good share of the equity isn't customer financing at all. The Intel stake, bought for $5 billion and now worth about $30 billion, is roughly a third of the book on its own; Synopsys and Nokia are suppliers or partners, not chip buyers. The guarantees are contingent. I found no sell-side estimate of how much of Nvidia's revenue is funded by Nvidia, and I wouldn't trust one that claimed precision.

Fig. 6Nvidia's exposure stack

Two health indicators are worth watching more than the total. Nvidia's receivables, money customers owe it, rose from $38.5 billion to $63.1 billion in two quarters, which takes the average time to collect from about 51 to about 60 days. Five direct customers now account for 70% of receivables, each holding between 10% and 22%. One direct customer alone was 16% of second-quarter revenue. Rising receivables are what you'd expect if customers were stretching their payments, or simply if sales were jumping; the next two quarters will tell which.

Underneath, demand is real. Data-centre revenue went from $26.3 billion in the July 2024 quarter to $89.0 billion in the July 2026 quarter. The segment Nvidia now calls "AI clouds, industrial and enterprise", which includes the neoclouds it funds, grew 138% year on year and is now growing faster than sales to the big hyperscalers.

CoreWeave, the four-sided tie

CoreWeave is the cleanest example of how many sides one relationship can have. Nvidia is its sole GPU supplier. It is a shareholder: in January 2026 it bought $2 billion of stock at $87.20 a share, taking it to 11.5% of CoreWeave's Class A shares according to its January filing, probably closer to 10% today as the share count has grown (Nvidia 13G/A). It's a customer. And under a $6.3 billion order signed in September 2025 it is "obligated to purchase the residual unsold capacity through April 13, 2032," subject to conditions (CoreWeave 8-K).

At $61.39, CoreWeave's close on 16 October, Nvidia's January purchase is about 30% underwater.

Fig. 7CoreWeave's four-sided tie to Nvidia

The pattern repeats. Nvidia rents about 18,000 of its own GPUs back from Lambda for roughly $1.5 billion over four years, according to The Information, though none of the parties has confirmed it. It bought $2 billion of pre-funded warrants in Nebius in March. It is putting $1 billion into Nscale's $3.36 billion pre-IPO convertible, with funding expected in mid-November (Nscale).

Huang's defence is consistent and worth stating fairly. He has called circular-financing claims about CoreWeave "ridiculous," arguing that Nvidia's equity is a small part of what partners raise, and in August he and his finance chief argued that these arrangements secure supply of infrastructure rather than create demand. There's something to that: Nvidia's chips are sold out, so it doesn't need to fund customers to sell them today. What the arrangements do is shape who gets to buy them and on what terms, and spread Nvidia's exposure across the customers most likely to struggle if demand cools.

The demand base: two labs, two ratios

Every layer of the loop is ultimately underwriting a small number of AI labs. Their two balance sheets have diverged sharply.

OpenAI has signed about $730 billion of dollar-denominated computing contracts: about $300 billion with Oracle over roughly five years from 2027, $250 billion with Microsoft Azure from its October 2025 restructuring, about $138 billion with AWS, about $22.4 billion with CoreWeave and more than $20 billion with Cerebras. On top of those sit gigawatt agreements with no fixed price: 6 gigawatts of AMD chips and 10 gigawatts of custom accelerators from Broadcom.

Its revenue is growing fast, but the number depends on who's counting. OpenAI itself said it booked $13.1 billion of revenue in 2025. On 29 September Axios reported an annualised figure "nearing $70B". On 8 October the FT reported that OpenAI had told investors its run-rate was approaching $50 billion at the end of September, about $20 billion lower, because OpenAI counts revenue from sales through Microsoft net of Microsoft's share, while the higher figure had been built by investors using gross, Anthropic-style accounting (TechCrunch). The story knocked Nvidia down 2.9% and CoreWeave 7.8% that day. A day later Bloomberg reported that OpenAI expects its run-rate to reach or exceed $70 billion by year-end (Bloomberg). ChatGPT had about 900 million weekly users as of February 2026.

Anthropic's run-rate went from about $9 to $10 billion at the end of 2025 to $30 billion in April, $47 billion in May and more than $65 billion at the end of July, counted gross (CNBC). Revenue it actually recognised in 2025 was about $4.6 billion, according to Reuters' review of its draft prospectus. Its disclosed dollar commitments come to about $225 billion.

Put commitments over revenue and you get the clearest single picture of the demand base.

Lab	Commitments	Revenue base	Multiple
OpenAI	~$730B contracted	~$50B run-rate, end-Sep 2026, net	~14.6x
OpenAI	$856B compute plan, 2026–30	~$840B planned revenue, 2026–30	~1.0x
Anthropic	~$225B disclosed	more than $65B run-rate, end-Jul 2026, gross	~3.5x
xAI	$12.7B capex, 2025	$3.2B revenue, 2025	~4.0x

These are coverage ratios, not annual burdens: they set five-to-ten-year contracts against a single year's revenue. And the methods differ: Anthropic's gross counting flatters its ratio relative to OpenAI's net figure. In April, OpenAI's chief revenue officer told staff that Anthropic's then $30 billion run-rate was inflated by about $8 billion through gross booking; Anthropic says it books gross because it acts as the principal. Even allowing for that, the gap between the two labs is large.

Fig. 8The revenue race

Fig. 9Commitments versus revenue

OpenAI's own plan makes the point more starkly. An internal presentation from July, reported by the FT in September, projects about $856 billion of computing spend from 2026 to 2030 against about $840 billion of revenue, with revenue rising from $36 billion in 2026 (calendar revenue, not run-rate) to $350 billion in 2030. Computing alone absorbs roughly all planned revenue. Cumulative free cash flow over the period is minus $278 billion, more than twice the record March round, and the cash from that round runs out by 2028. The headline target has moved three times in a year, from $1.4 trillion over eight years (Altman, November 2025), to about $600 billion through 2030 (February), to $856 billion (July), so it's a plan, not a contract.

Funding is dominated by suppliers. Of the $122 billion round, Amazon and Nvidia supplied about two-thirds, and each recycles into its own products. SoftBank completed its $30 billion on 3 October, taking its cumulative investment to $64.6 billion, roughly 13% of the company (Manila Times). Another round of at least $30 billion at a valuation of $1.2 to $1.4 trillion is in talks.

Vendors paying the lab in equity

The most unusual feature of OpenAI's deals is that some of its suppliers pay it.

AMD granted OpenAI a warrant for up to 160 million shares, about 10% of the company, at one cent each (AMD 8-K). The first tranche vests when the first gigawatt of AMD's MI450 chips is deployed, the rest as deployments reach 6 gigawatts, and every tranche also needs share-price and technical conditions, with the final one requiring $600 a share. AMD first closed above $600 on 22 September. That alone vests nothing: as of 27 June, AMD's filing showed no tranche vested, and there's no public evidence the first gigawatt has been reached (AMD 10-Q).

Cerebras, a chipmaker with a $10 billion-plus capacity deal with OpenAI, issued it warrants over 33.4 million shares. In July OpenAI exercised the vested part: about 10 million shares, 4.22% of the company, for roughly $100. SB Energy, which is building a 1.2-gigawatt site in Texas for OpenAI under a 20-year lease, reportedly received $5.5 billion of OpenAI warrants, a flow in the opposite direction.

On a diagram of money, these are arrows pointing backwards: a supplier gives its customer a slice of itself to win the order. It's a discount paid in stock, and it makes the supplier's share price part of the lab's balance sheet.

Moving risk onto the landlords

The second shift of 2026 is that OpenAI stopped trying to own its infrastructure. It now calls Stargate an "umbrella term" and prefers to lease rather than own (Tom's Hardware). In April Microsoft took over the 230-megawatt Narvik capacity in Norway that OpenAI had planned to lease, and OpenAI will now reach it through its Azure contract. A planned 600-megawatt expansion at Abilene was dropped in March, according to Bloomberg, although Oracle called the reports false. Stargate UK is paused. And the WSJ reported in late April that OpenAI's finance chief, Sarah Friar, had warned she might not be able to pay future computing contracts if growth lagged; Altman and Friar called the report "ridiculous."

Each of these moves pushes risk from the lab to whoever signed the lease or borrowed to build the building. That's rational for OpenAI. It also means the people holding the risk are, increasingly, the debt-funded middle layer.

Anthropic and xAI

Anthropic's loop is tighter. Its compute contracts are paired almost one for one with equity from the same counterparty. Amazon put in $5 billion with up to $20 billion more on milestones, against Anthropic's commitment of more than $100 billion over ten years for up to 5 gigawatts of AWS capacity (CNBC). Microsoft and Nvidia committed up to $15 billion against a $30 billion Azure commitment. Google and Broadcom will supply about 3.5 gigawatts of TPUs from 2027, which Broadcom's filing makes conditional on Anthropic's commercial performance. The WSJ reported a $35 billion, six-year deal with Lambda at a Texas site where Nvidia holds the lease; no company has confirmed it.

The oddest deal has Anthropic renting xAI's Colossus 1 data centre from SpaceX for about $1.25 billion a month through May 2029, per SpaceX's prospectus. Either side can terminate on 90 days' notice, so the committed exposure is far smaller than the roughly $45 billion headline. One lab is monetising another's spare capacity.

xAI, now part of SpaceX, is the debt-heavy outlier. Per SpaceX's prospectus, it lost $6.36 billion on $3.20 billion of revenue in 2025 and spent $7.7 billion on capex in the first quarter of 2026 alone (Morningstar). Its Colossus 2 cluster was reportedly financed through a $20 billion SPV, $7.5 billion of equity (up to $2 billion from Nvidia) and $12.5 billion of debt, which buys Nvidia GPUs and leases them to xAI for five years.

The hyperscalers' hidden balance sheet

The hyperscalers, the giant cloud providers, turn the labs' commitments into concrete and silicon. In 2026 they began funding it from outside.

After the July earnings calls, 2026 capex guidance stood at about $220 billion for Amazon, $195 to $205 billion for Alphabet and $130 to $145 billion for Meta. Microsoft's calendar-2026 figure was reported at about $175 billion including finance leases, below the $190 billion cited in April, but that is a lease-accounting change, not a cut: more of its leases are now classified as operating leases, which don't count as capex (Benzinga). The big four together come to roughly $720 to $745 billion; with Oracle, $775 to $800 billion. Every company raised its guidance at least once during the year.

To put that in context: Microsoft's cash capex was $23.9 billion in its fiscal 2022 and $115.9 billion in fiscal 2026. Alphabet's went from $31.5 billion in 2022 to $91.4 billion in 2025, Meta's from $32.0 billion to $72.2 billion, Amazon's from $63.6 billion to $131.8 billion. Oracle's went from $6.9 billion in fiscal 2024 to $55.7 billion in fiscal 2026.

The latest quarter shows the cash running out. Amazon spent 119% of its operating cash flow on capex, Alphabet 115% and Oracle 123%. Meta roughly broke even at 98%. Only Microsoft, at 65%, still covered its spending comfortably. Epoch AI estimates capex has been growing about 70% a year against about 23% for operating cash flow.

Fig. 10Capex outruns cash

$2.35 trillion of promises

The spending is justified by backlog. Microsoft's commercial RPO was $678 billion at the end of June, Oracle's $664 billion at the end of August, Google Cloud's $514 billion and AWS's $496 billion: about $2.35 trillion in total (Yahoo Finance).

Much of it rests on one customer. In January Microsoft said about 45% of its RPO at December 2025, then about $625 billion, was tied to OpenAI; it has not updated the figure, and management attributed the latest quarter's growth to other clients. S&P says OpenAI is about half of Oracle's backlog. Put together, using those different dates, roughly $600 billion of backlog at two companies depends on a single lab with a net run-rate of about $50 billion.

Fig. 11$2.35 trillion of promises

The relationship cuts both ways. Microsoft earned $24.1 billion of revenue from OpenAI in its fiscal 2026, and its OpenAI stake contributed a $4.96 billion net gain to income that year, against a $3.62 billion loss the year before (Microsoft 8-K). OpenAI is now both one of Microsoft's largest customers and a source of paper profit.

Borrowing at record pace

The five hyperscalers sold $121 billion of US-dollar bonds in 2025, by Bank of America's count, which includes Meta's $27 billion Hyperion SPV bond, against an average of about $28 billion a year in 2020 to 2024. By 10 August 2026 they had sold about $220 billion, by JPMorgan's count, including about $62 billion in other currencies. Amazon's roughly $54 billion sale in March was the largest corporate bond deal on record. Alphabet sold the first 100-year bond by a tech company since 1997.

They are now about 5% of the Bloomberg US investment-grade index and, per the IMF, about 10% of global investment-grade issuance, up from under 1% in 2024. Demand is softening: between February and July, order books fell from about five times the bonds on offer to about two times, spreads widened from about 50 to about 80 basis points, and 78 of 91 hyperscaler bonds issued in 2026 were trading at higher yields in late July than when they were sold (FactSet).

The most telling move came in June. Alphabet, the company with the strongest balance sheet in the group, raised up to $84.75 billion of equity: $18 billion of common stock, $16.75 billion of mandatory convertible preferred, a $40 billion at-the-market programme to sell shares gradually from the third quarter, and a $10 billion private placement with Berkshire Hathaway (Bloomberg). When the best credit in the room chooses to issue shares rather than more debt, it's telling you something about the price of debt.

Fig. 12Bonds at record pace

What isn't on the balance sheet

The bigger obligations don't appear as debt at all. Microsoft has $329.1 billion of leases, mostly data centres, that it has signed but that haven't started yet, running from fiscal 2027 to 2033. Oracle has $288 billion, with 15-to-19-year terms. Those two alone come to $617 billion; FactSet puts the total across the group at about $820 billion.

Meta pioneered a different structure. Its Hyperion data centre in Louisiana sits in a joint venture, Beignet Investor LLC, 80% owned by an affiliate of Blue Owl, which contributed $7.0 billion in cash; Meta owns the other 20% (Meta 10-Q). The venture borrowed $27.3 billion through bonds paying 6.581% and due in 2049. Meta signed leases starting in 2029, initially for four years and renewable to 20, and gave a residual value guarantee: if it walks away and the property is worth less than a threshold that "effectively begins at $28 billion," Meta pays part of the difference. S&P excludes that guarantee from Meta's leverage. Meta repeated the structure with BlackRock at El Paso for about $14 billion, and that bond priced about 40 basis points wider.

Fig. 13Off the balance sheet

Accounting choices flatter the picture further. Microsoft, Alphabet and Oracle depreciate servers over six years and Meta over five and a half, while Amazon shortened some to five. Burry estimates that longer useful lives will understate depreciation by about $176 billion across the hyperscalers from 2026 to 2028, on the view that a GPU's economic life is two to three years (StreetInsider). That's his estimate, not an audited one, and Nvidia argues that older chips still earn rent. The GPU price data below suggests Nvidia has the better of that argument today.

My reading: the hyperscalers can afford all of this, but they are no longer paying for it out of pocket. The shift from cash flow to bonds, equity and off-balance-sheet leases is the single biggest change in the loop this year.

The cracked middle

Every visible crack of 2026 opened in the layer that borrows against someone else's contract.

Oracle: the investment-grade version

Oracle's backlog went from $137.8 billion in May 2025 to $455.3 billion three months later, when the OpenAI contract landed, and to $664 billion by August 2026. Only about 13% of it will be recognised as revenue within twelve months (Oracle 10-Q). To deliver the rest, Oracle has to build first and get paid later.

The cost is showing. Quarterly free cash flow was minus $5.4 billion. Debt stands at $125.3 billion. On 9 July S&P cut Oracle to BBB-, one notch above junk, projecting a free-cash-flow deficit of about $42 billion for fiscal 2027 (Computing). Oracle has turned to equity: $19.9 billion of at-the-market share sales and a mandatory convertible, within a $45 to $50 billion financing plan for 2026, about half of it equity. Its five-year CDS rose from about 144 basis points at the start of the year to a record of about 203 in late July and about 288 in mid-October, according to Reuters citing S&P Global Market Intelligence.

The project level shows it too. Around 24 September Oracle sent a force-majeure notice, a claim that events beyond its control excuse a delay, to a Blue Owl unit over Project Jupiter, a 2.45-gigawatt site in New Mexico, citing a gas-pipeline delay now expected to last until 1 February 2027 (TNW). Oracle calls it routine. The project's $18 billion loan was reportedly quoted at 89 to 91 cents, and a bank sold part of it below 90.

Fig. 14Oracle under strain

How a neocloud works

Strip away the jargon and a neocloud is a GPU landlord running on borrowed money. The business has five steps.

1. Sign a tenant first. A big customer such as Microsoft, Meta or OpenAI agrees to rent a set amount of computing for several years, usually take-or-pay.
2. Borrow against the lease. With the contract in hand, the neocloud borrows most of the cost of the chips. The lenders' security is the contract's cash flows plus the GPUs themselves.
3. Buy the chips. The money goes to Nvidia, usually through a server maker such as Dell, for GPUs, networking and storage.
4. Rent the building and the power. Most neoclouds don't own their data centres. They lease powered shells from developers and former bitcoin miners, often for 12 to 15 years, and buy the electricity.
5. Collect the rent and race the clock. The tenant's payments cover power, staff and interest, and must repay the loan before the chips go stale, because Nvidia ships a faster generation every year.

When the tenant is strong and the loan is repaid inside the contract, this is a good business. The trouble is everything outside that box.

Fig. 15How a neocloud works: the Meta contract

One contract, step by step

CoreWeave's contract with Meta is the cleanest example, because its financing is public.

The tenant. On 30 September 2025 CoreWeave announced a $14.2 billion capacity deal with Meta, reportedly running to December 2031 and since reportedly enlarged. It's take-or-pay, with payment deductions if capacity isn't available and protections if it arrives late.

The loan. In March 2026 CoreWeave set up a separate company, CoreWeave Compute Acquisition Co. VIII, which borrowed $8.5 billion through a delayed-draw term loan, drawn in instalments as equipment arrives (CoreWeave). CoreWeave didn't name the customer; Bloomberg reported it was Meta. The borrower is ring-fenced from a CoreWeave bankruptcy, and its cash comes from Meta, so Moody's rated the loan A3, solid investment grade, a rating CoreWeave itself is nowhere near. About half pays a fixed rate of about 5.9%, the rest SOFR plus 2.25%. After 15 months of drawing, it pays down to zero over about five years, with "no refinancing risk," in the words of the rating agency DBRS (DBRS).

The chips. The money pays Dell and Nvidia for Blackwell-class systems. At SemiAnalysis's estimate of about $3.9 million for a 72-GPU rack, all-in, $8.5 billion buys something like 150,000 GPUs. That's my arithmetic; CoreWeave hasn't disclosed the mix.

The buildings. The racks go into leased third-party data centres. DBRS singles out two North Dakota sites, at Applied Digital's Polaris Forge campus, where CoreWeave leases 400 megawatts for about $11 billion over about 15 years, as the only ones with fully hedged power.

The rent. CoreWeave doesn't disclose the economics of a single contract, so this is an illustration built from an analyst's model (Rittenhouse Research), not company figures. For every 100 megawatts, the customer pays about $100 million a month. Building rent takes about $14 million and power about $6.5 million, leaving about $80 million. Wearing out the GPUs over six years costs about $37.5 million a month and interest about $11 to $13 million, so the contract earns roughly $30 million a month before tax. But repaying the loan takes about $33 million a month, so nearly all the cash goes to the lender until the contract ends.

The end. Around 2031 or 2032 the loan is repaid, Meta is free to leave, and CoreWeave owns a fleet of five-year-old GPUs with years still to run on the building lease.

In this deal Meta is safe and the A3 lender is safe. The risk sits with CoreWeave's shareholders and unsecured bondholders, who fund the slice of every project that isn't borrowed against a contract, the construction delays, the long leases and the fleet's second life.

CoreWeave: the contract works, the company doesn't

Add up all the contracts and the picture changes. CoreWeave's revenue grew from $229 million in 2023 to $1.9 billion in 2024 and $5.1 billion in 2025 (CoreWeave 10-K), and it guides to $12.4 to $13.2 billion for 2026. In the second quarter of 2026 it earned $2.58 billion of revenue and $1.51 billion of adjusted EBITDA, a 59% margin that looks like a software company's (CoreWeave). Then come the two costs EBITDA leaves out.

The first is depreciation. A GPU fleet costs more than two years of the revenue it earns and wears out in four to six, so a large slice of its cost has to be charged as an expense every year. After depreciation, adjusted operating income was $128 million, a margin of about 5%. Under standard accounting rules (GAAP), it was a $49 million operating loss.

The second is interest. Net interest expense was $640 million, a quarter of revenue and five times adjusted operating income, and CoreWeave guides the third quarter to $860 to $940 million. The net loss was $626 million.

Why does the contract make money while the company loses it? Because the company grows faster than its contracts mature. It's always paying for buildings and chips that aren't billing yet, and much of its debt isn't the cheap, contract-backed kind but unsecured bonds with 9% to 10% coupons. An independent analysis of the first-quarter filing found that the equipment in service yielded about 12.8% a year in cash, against an average cost of debt of 9.1% (Cape Fear Advisors): a spread of under four points on an asset that loses value every month. Jim Chanos, the short-seller, made the same point on the record in December 2025, saying the cost of capital is "higher than the return on" these data-centre deals (Yahoo Finance).

Fig. 16One contract versus the whole company

Of its $35.1 billion of debt, about $31.4 billion is recourse, meaning lenders can claim against the whole company rather than a single project. Analysts' consensus has free cash flow at about minus $25 billion in 2026 and minus $28 billion in 2027. This is a depreciation-and-interest problem, not a pricing problem: CoreWeave charges enough to make a healthy margin on each GPU hour, then spends it on wearing out the GPUs and paying for the money it borrowed to buy them. D.A. Davidson's Gil Luria says the downside case is that "the equity value is going to be zero."

Fig. 17CoreWeave: growth versus interest

Four clocks that don't line up

The deeper problem is timing. A neocloud runs four clocks at once, and they don't match.

• The contracts. CoreWeave's $103.7 billion of contracted backlog at the end of June runs off fast: 41% within two years, 39% in years three and four, and only about 20% after that (CoreWeave 10-Q). The average committed contract lasts about five years.
• The chips. CoreWeave depreciates its GPUs over six years, longer than Nebius (four) and reportedly Lambda (five), while Nvidia ships a new generation every year.
• The buildings. Its leases run far longer: $16.3 billion of lease liabilities with a weighted average of 12 years left.
• The debt. The contract-backed loans pay down inside their contracts. The rest doesn't. Unsecured bonds mature from 2030 to 2033 and will have to be refinanced at whatever the market charges then, and $7.5 billion of debt was due within twelve months at the end of June.

So when a first contract ends, CoreWeave still owes rent on the building and still has bonds to roll over, and the asset it has to pay with is a fleet that's no longer the newest. CoreWeave says its cash and available facilities are sufficient for what falls due within a year. The danger isn't the next twelve months. It's the second contract.

Fig. 18Four clocks that don't line up

The counterparty sets the price

The rest of the neocloud tier teaches one lesson clearly: the cost of capital is set by who signed the contract, not by the GPU. CoreWeave's own borrowing shows it best, because the same kind of chips, inside the same company, borrow at very different rates depending on the customer.

Facility	Backed by	Price
DDTL 4.0, $8.5B (Mar 2026)	Meta contract	Rated A3; SOFR+2.25%, or ~5.9% fixed
DDTL 3.0, $2.6B (Jul 2025)	OpenAI contract	SOFR+4.00%
$3.1B facility	CoreWeave parent guarantee	Rated Ba2; SOFR+4.50%
DDTL 2.0	Earlier contracts	~11% effective rate
Unsecured bonds, 2030–32	CoreWeave alone	9.25–9.75% coupons; yields of 14.5–16% in mid-October

The lenders are pricing Meta, OpenAI and CoreWeave, not the chips (DDTL 3.0 release; Cape Fear Advisors). OpenAI, unrated and loss-making, costs about 1.75 percentage points more than Meta for the same hardware. Ring-fencing sends the cheap money to the projects and leaves the parent's shareholders and bondholders with everything else, which is why the parent's bonds trade at distressed prices while the Meta loan, as far as anyone can see, is fine.

Fig. 19Same chips, different price

IREN, a former bitcoin miner, borrowed $3.65 billion against a $9.7 billion Microsoft contract, ring-fenced in a subsidiary, at SOFR (the benchmark short-term dollar rate) plus 2.13% to 2.25% (Blockspace). Lambda's Microsoft- and Nvidia-backed loans were rated A (low) and Baa1, solid investment grade. Nebius limits borrowing by getting customers to prepay: about 70% of its second-quarter deals pre-fund half or more of the capex, and it sold about $4.3 billion of convertible bonds at 1.25% to 2.625% after a Meta deal worth up to about $27 billion. At the other end, CoreWeave's unsecured bonds yielded 14.5% to 16% in mid-October, and Applied Digital pays 12.75% in kind on preferred shares sold to Macquarie.

Google found a quieter variant. It guarantees the leases of Fluidstack, a small cloud provider, at data centres owned by TeraWulf and Cipher Mining, and in return receives warrants in those companies at one cent a share (TeraWulf 10-Q). It turns an unrated tenant into Google credit without any Google capex. It's the same move as Nvidia's Ohio guarantee, at smaller scale.

Fig. 20The cost-of-capital ladder

Investment-grade counterpartyLab creditNeocloud credit

What isn't broken yet: GPU prices

The bear's simplest story is that GPU rental prices will collapse, leaving the middle layer with expensive hardware nobody wants to rent. That isn't in the data yet. SemiAnalysis's index for one-year H100 rentals fell from about $8 an hour in 2023 to $1.70 in October 2025, then rebounded about 40% to about $2.35 by March 2026 as capacity sold out (Seeking Alpha). A different index, Ornn's, read $2.92 on 1 October. Silicon Data's B200 index is up about 28% this year.

Fig. 21GPU prices bounced

Rental prices are only half of it. What matters when a contract ends is what an old GPU will fetch, and the two most recent generations tell different stories. Used H100s, the workhorse of 2023 and 2024, sold for about $14,250 in August, up 16% in a year, according to realised sales tracked by American Compute. Used A100s, one generation older, sold for about $5,400, down a third (IntuitionLabs). American Compute's conservative case puts the H100 at 11% to 19% of its original list price in 2027 and 5% to 7% in 2028. So residual value isn't zero, but it is cyclical: it depends on whether new supply outruns demand in the year your contract ends. CoreWeave's lenders reportedly insisted on six-year depreciation with no salvage value. In other words, the lenders assume the chips are worth nothing at the end, and the shareholders are the ones betting they're worth something.

So the middle layer's problem is duration and leverage. It has borrowed for five to twenty years against assets that lose value in one to six and contracts concentrated on a handful of customers. A wave of Blackwell and Rubin supply in 2027 and 2028 could repeat the H100 price decline of 2024 and 2025, but that is a forecast, not an observation.

Why neoclouds have no moat

A moat is whatever stops competitors or customers from taking your profits. Look at what a neocloud owns and it's hard to find one.

The chips come from a near-monopolist. Every GPU CoreWeave runs is Nvidia's. Early access to scarce chips was the neoclouds' great advantage in 2023 and 2024, but that allocation is Nvidia's to give. Nvidia likes having a spread of Nvidia-only customers as a counterweight to the hyperscalers' own chips (Google's TPUs, Amazon's Trainium, Microsoft's Maia), and it backs several neoclouds at once. If its priorities change, or supply loosens, the advantage goes.

The buildings and power are mostly rented. CoreWeave carries $16.3 billion of lease liabilities. Power, which Satya Nadella calls the real bottleneck, is exactly what neocloud deals sell to the hyperscalers, and the neocloud usually rents it too.

The demand sits with one to three customers. In the second quarter of 2026, CoreWeave's largest customer provided 36% of revenue, the second 26% and the third 10% (10-Q). Microsoft alone was about 67% of 2025 revenue.

Those customers rent because they can't build fast enough. Microsoft has committed more than $60 billion to neoclouds, and it's open about wanting flexibility. Nadella said on the Dwarkesh Podcast in November 2025 that Microsoft "didn't want to just be a host for one company," and that he didn't want five years of depreciation on one chip generation. Amy Hood, its finance chief, says about two-thirds of its capex goes on short-lived GPUs and CPUs, and that if demand changes "we can simply slow down" that part (Microsoft). Meanwhile it is stretching the accounting life of the data-centre buildings it owns from 15 to 25 years. My inference, not Microsoft's words, is that renting from a neocloud is a bridge: it lets a hyperscaler keep the option to walk away and hands the generation risk to the landlord.

The software helps but doesn't lock anyone in. CoreWeave's orchestration tools and Weights & Biases, the developer platform it bought in 2025, make clusters easier to run. But the biggest customers bring their own software and treat the capacity as bare metal, so switching at the end of a contract costs only the move.

What a neocloud owns outright is a fleet that depreciates in four to six years, plus a real but shrinking skill at bringing capacity online fast.

History isn't kind to businesses with that shape. Exodus Communications, the web-hosting star of 2000, was worth about $32 billion at its peak, borrowed about $3.5 billion in bonds to build some 47 data centres ahead of demand, and filed for bankruptcy in September 2001 (Computerworld). Core Scientific, a bitcoin miner that hosted other miners' machines, filed in December 2022 after the bitcoin price fell and a big hosting customer, Celsius, went bust (CoinDesk). In both cases the buildings and power outlived the original shareholders. Core Scientific's sites now house CoreWeave.

The newest neocloud to seek public money shows the same shape at an earlier stage. Nscale is pitching a $35 billion IPO valuation on $103 billion of contracts, "much not yet firm" according to Fortune, with Nvidia putting $1 billion into its pre-IPO convertible. Its filing shows first-half 2026 revenue of $140.6 million against cost of revenue of $189.6 million, so it lost money before paying for any overheads, and a net loss of $1.02 billion, including $457 million of fair-value charges. Its largest customer provided 52% of revenue. Management had recorded substantial doubt about the company's ability to continue as a going concern, because it relied on financing nobody had committed, and resolved it on the basis that it could defer or cancel capex (Fortune).

The bull case deserves a fair hearing, because at the level of the contract it's strong. The big deals are take-or-pay with investment-grade customers, and the best loans pay off inside the contract. Nebius gets customers to prepay half or more of the capex on most new deals. Nvidia has promised to buy CoreWeave's unsold capacity through 2032, and Google guarantees Fluidstack's leases. GPU rental prices have rebounded, and CoreWeave said in September that some new short contracts were priced higher. Michael Intrator, its chief executive, told staff the backlog "more than covers" the debt (Reuters via The Star). All true. But backlog is gross revenue that still has to pay for power, rent and the capex needed to deliver it, and none of these protections covers the second contract or the parent's shareholders.

Fig. 22CoreWeave since its IPO

IPO $40Nvidia's entry $87.20

My reading: the neocloud is the residual claimant on the fastest-depreciating asset in the AI stack. It earns a thin spread over its cost of money when things go well and absorbs the loss when they don't. That's why it cracked first, and it's the first place a disappointment in end-customer demand would land.

Who's lending

Someone has to fill the gap between what the hyperscalers earn and what everyone is spending. Morgan Stanley puts data-centre capex at about $2.9 trillion through 2028, of which about $1.5 trillion must come from outside capital and about $800 billion from private credit (Cleary Gottlieb). JPMorgan sees more than $5 trillion over five years and says about $650 billion of annual revenue would be needed for viable returns (DCD).

Adding up the observable 2026 sources (hyperscaler bonds at an annualised pace of about $300 billion, $30 to $40 billion of data-centre securitisations, $50 to $100 billion of project loans and bonds, tens of billions in GPU SPVs and more than $140 billion of lab equity) gives roughly $500 to $600 billion. That's my estimate from the pieces, not a sourced total, but it suggests the gap is being filled.

Fig. 23Who funds the gap

Stargate: headline versus money

Stargate shows the difference between an announcement and a financing. The January 2025 launch promised $500 billion, with initial equity reportedly set at $19 billion each from SoftBank and OpenAI and $7 billion each from Oracle and MGX, the Abu Dhabi AI investor. OpenAI says it has "secured" more than 10 gigawatts. Epoch AI's site tracker showed only Abilene operating, at about 0.3 gigawatts, in April 2026, against about 9.6 gigawatts planned across seven US sites (Epoch AI).

The money that has actually built anything is project debt sized against Oracle's leases, not Stargate equity: about $9.4 billion of JPMorgan loans at Abilene, a roughly $38 billion JPMorgan and MUFG package for two Vantage sites in Texas and Wisconsin (DCD), and $16 billion at Saline, Michigan. Saline is instructive. Blue Owl walked away in December 2025; Blackstone replaced it, and about $14 billion of project bonds paying 7.5% and due 2045 were anchored by PIMCO (Blackstone). Oracle's credit, not SoftBank's or OpenAI's, is the load-bearing beam.

Fig. 24Stargate: secured versus built

$500B headline (Jan 2025) · initial equity reported: SoftBank $19B, OpenAI $19B, Oracle $7B, MGX $7B

Abilene, TXC

Crusoe, Blue Owl and Primary Digital. JPMorgan loans of $2.3B + $7.1B ≈ $9.4B. A 600 MW expansion was dropped in Mar 2026 (Bloomberg; Oracle disputed).

Vantage: Shackelford, TX and Port Washington, WI

≈$38B of project loans led by JPMorgan and MUFG, on leases to Oracle.

Saline, MIBXP

Related and Blackstone; ≈$14B of 7.5% bonds due 2045, PIMCO as anchor. Blue Owl walked away in Dec 2025.

Milam, TXSBE

SB Energy, 1.2 GW. OpenAI $500M + SoftBank $500M + Ares $800M of preferred equity.

Project Jupiter, Doña Ana, NM

Oracle, 2.45 GW. Force majeure notice ≈24 Sep 2026; its $18B loan reportedly quoted at 89–91 cents.

SoftBank: a loan on a mark it set

SoftBank is the most leveraged equity holder in the cycle, and its chain is the cleanest example of circularity in the credit sense. In late 2025 it sold its entire Nvidia stake for $5.83 billion and part of its T-Mobile holding for $9.17 billion. In March 2026 it took a $40 billion bridge loan maturing in March 2027. In May it cut the target for a margin loan backed by its OpenAI shares to $6 billion amid creditor hesitation (Bloomberg), then closed $10 billion in August at SOFR plus 425 basis points, with margin-call provisions (Bloomberg). It is seeking a second $10 billion.

SoftBank reports a loan-to-value ratio of 13%. The catch is that the collateral, its OpenAI stake, is valued at a private round that SoftBank itself co-led. If OpenAI's next round came in lower, SoftBank's asset value and its loan covenants would be hit at the same moment.

Fig. 25SoftBank's chain

The substitutes, and the regulators

Gulf money enters at the top of capital structures where banks won't go. MGX closed a $49 billion first fund, co-led OpenAI's round, and sits inside the BlackRock-led consortium whose roughly $40 billion buyout of Aligned Data Centers closed in July. Securitisation, packaging data-centre leases into tradable bonds, went from $2.4 billion of issuance in 2020 to $15.5 billion in 2025, and an SEC staff letter in July eased the rules (InvestmentNews).

Regulators now name the loop explicitly. Besides the IMF and BIS work above, the IMF's October World Economic Outlook includes a "market correction" scenario in which a steep fall in AI investment and asset prices leaves global output 1.1% below baseline by 2028, and US output 1.8% below (IMF WEO, Box 1.2). The Financial Stability Board found AI accounted for more than a third of 2025 private-credit deals. In the Fed's spring survey, the share of respondents naming AI as a salient risk rose from 30% to 50%.

So far, though, every stress episode has been absorbed by a substitute provider: Blackstone for Blue Owl at Saline, Apollo among SoftBank's lenders, PIMCO for retreating banks. The system has shown repricing, not refusal. That distinction matters. Repricing raises costs and squeezes the weakest borrowers. Refusal, when no one will lend at any price, is what turns a squeeze into a bust.

Does any of this pay?

All of this eventually rests on one question: will businesses and consumers pay enough for AI to service the commitments? Here the measurement asymmetry becomes the story.

Why the question moved to the front

Until this year the market treated that question as a matter of time. It no longer does. JPMorgan Asset Management found that in the latest earnings season, higher capex plans lifted share prices only when revenue estimates rose with them (JPMAM). Meta showed what happens otherwise: its free cash flow fell to $1.2 billion in the first quarter of 2026 from $26 billion a year earlier, its shares fell about 8% when it raised capex guidance, and Jefferies said it would stay "in the penalty box pending clearer capex ROI" (CNBC).

Small changes in corporate behaviour matter because lab revenue is concentrated. Among Ramp's customers, the top 1% of spenders account for about 80% of OpenAI and Anthropic spending, and those are exactly the firms that started rationing this year. A modest cap at the top moves lab revenue, which sets the growth rate the clouds' capex and the neoclouds' debt were underwritten on.

The easy side to count

Vendor revenue is audited and enormous. Nvidia sold $96.2 billion in one quarter. Microsoft's Azure passed $100 billion in fiscal 2026, and it has more than 30 million paid seats for its Microsoft 365 Copilot. Palantir's US commercial revenue grew 149% year on year in the second quarter (Palantir).

Usage is growing even faster. Google processed 9.7 trillion tokens (the fragments of words AI models read and write) a month in 2024, about 480 trillion in May 2025, and more than 3.2 quadrillion by May 2026 (Google). But the price of a fixed level of AI capability falls about tenfold a year (a16z), so revenue grows much more slowly than tokens. More AI is being used every month; each unit earns less.

Even the vendor side has a soft spot. The private labs' run-rates are unaudited and calculated differently, and part of their revenue is funded, indirectly, by investment from their own suppliers.

Against $700 billion to $1 trillion of annual capex, the sourced AI revenue stack (Anthropic's run-rate of about $65 billion, OpenAI's about $50 billion, Microsoft's AI business at about $37 billion, with some overlap) comes to roughly $160 billion before double counting is removed. The ratio is improving, because lab revenue grows faster than capex. The absolute gap is not narrowing. Bain estimates $2 trillion of annual revenue will be needed by 2030 to justify the spending, against an $800 billion shortfall (Bain).

Fig. 26Tokens versus dollars

The hard side to count

On the enterprise side, the answer depends on the instrument.

The US Census Bureau's Business Trends and Outlook Survey, the best random sample of American firms, found 17% to 20% of firms using AI between December 2025 and May 2026, and about 32% weighted by employment (Census). The latest reading, 23.8% for late August to early September, has been reported but not yet published in a Census story. Large firms reached 37% by May. The question was reworded in November 2025, which nearly doubled the measured rate, so the series has a break.

Ramp, a corporate-card company, sees a different world in its transaction data: 55.9% of businesses paying for AI tools in September (Ramp). But the median adopter spends about $12 a month per employee, against about $7,200 for the top 1% in August (a figure Ramp later revised up). Its latest data are covered below. In September Anthropic led OpenAI in business adoption, 43.8% to 40.6%.

Fig. 27Two ways to count adoption

Executive surveys conflict, and none of them is strong evidence. MIT's NANDA initiative found that about 95% of organisations saw "no measurable P&L impact" from generative AI, from a preliminary, self-reported study covering six months of 2025. Wharton and GBK found about 74% of executives reporting positive returns (Wharton). McKinsey found only about 6% of firms attributing 5% or more of their operating profit to AI (McKinsey). MIT's 5% and McKinsey's 6% agree more than the headlines suggest: a small minority is booking enterprise-level gains.

Big at the task, small in the economy

The rigorous evidence draws a clear two-level picture.

At the level of individual tasks, randomised and field studies show large gains. Customer-support agents with an AI assistant resolved 14% more issues per hour, and novices 34% more (NBER). Professionals writing with ChatGPT took 40% less time and produced work rated 18% better (Science). Developers at Microsoft, Accenture and a Fortune 100 company completed 26% more tasks (SSRN). Consultants at BCG were about 25% faster on tasks inside AI's capabilities, but 19 percentage points less likely to get the right answer on tasks just outside them (HBS). There are real negatives: experienced open-source developers were 19% slower with AI in METR's 2025 trial while believing they were faster, and METR's 2026 follow-up points to a speed-up whose confidence interval includes zero (METR).

At the level of the economy, the effects are hard to find. Danish administrative data covering 25,000 workers rule out effects on earnings or hours larger than 2% two years after ChatGPT's launch (NBER). The clearest labour-market signal is at the hiring margin: Stanford's latest "Canaries" update finds that 22-to-25-year-olds in the most AI-exposed jobs are 19% below where they'd be had they kept pace with less-exposed peers (Stanford). That figure isn't directly comparable with the 13% and 16% in earlier versions, which measured something slightly different. Daron Acemoglu, a Nobel-winning economist, estimates AI will raise total factor productivity by no more than about 0.66% over ten years (NBER); Goldman Sachs projected a 7% lift to global GDP over the same span.

Fig. 28The evidence ladder

easier to countharder to count

1. 1. Vendor revenueaudited
   $96.2B Nvidia revenue, Q2 FY27

   >$100B Azure revenue, FY26

   30M+ paid Microsoft 365 Copilot seats

   +149% Palantir US commercial, year on year
2. 2. Usagecompany-reported counts
   >3.2 quadrillion tokens a month processed by Google (May 2026), ≈7× a year earlier
3. 3. Task-level trials and field studiesrandomised, narrow
   +14% support issues resolved per hour; +34% for novices

   −40% time, +18% quality professional writing tasks

   +26% tasks completed by developers

   ≈25% faster, −19 pts BCG consultants, inside vs outside the frontier

   −19% METR 2025: experienced developers slower with AI

   includes zero METR 2026: confidence interval
4. 4. Firm surveysself-reported
   ≈95% MIT NANDA: no measurable P&L impact

   ≈74% Wharton/GBK: report positive ROI

   ≈6% McKinsey: get 5% or more of EBIT from AI
5. 5. Economy-wideleast measurable
   >2% ruled out Denmark: effects on earnings and hours

   19% below young workers in exposed jobs vs counterfactual (Aug 2026)

   ≤0.66% Acemoglu: total factor productivity gain over 10 years

So, for now, AI shows up in GDP mainly as investment spending. ING attributes "around a third" of recent US growth to tech investment (ING), and Epoch puts AI data centres and hardware at about 0.8% of GDP. It doesn't yet show up as productivity.

The end of tokenmaxxing

The newest evidence comes from the companies that were supposed to be AI's best customers.

From spring 2025, the fashion in big companies was tokenmaxxing: using as much AI as possible, measured in tokens. Shopify's chief executive made "reflexive AI usage" a baseline expectation in performance reviews. Meta ran an internal leaderboard of its top 250 token users, which staff reportedly gamed by setting agents to run parallel tasks just to climb it. Coding agents got near-unlimited budgets.

In 2026 the bills arrived. Almost all of what follows comes from leaked memos and press reports, several from a single outlet, so treat individual details with care. The pattern is consistent.

• Uber reportedly used its whole 2026 AI budget in about four months, then capped every employee at $1,500 a month per coding tool. Its president, Andrew Macdonald, said the link between soaring token use and better products "is not there yet" (Letsdatascience).
• Amazon removed an internal AI-usage leaderboard after staff gamed it (TNW).
• Meta told about 6,000 staff that internal AI use was on track to cost billions in 2026, ended its leaderboard and introduced central token controls (single source: The Information, via AI Weekly).
• Microsoft put a token spending cap on every division and moved its default internal model to a cheaper one. An executive wrote that "tokenmaxxing is not what we are optimizing for" (single source: 404 Media, via TNW).
• Accenture told staff not to use AI for trivial tasks, to conserve tokens; its agentic AI lead said AI is becoming "material to the cost structure" (single source: 404 Media, via TechCrunch).

None of these companies stopped using AI. They capped the heaviest users and moved routine work to cheaper models. That's rationing, not abandonment. But it bites exactly the customers who carry most lab revenue.

More tokens, flat spend

The best systematic evidence comes from Ramp, whose AI Index tracks card and bill-payment data from about 70,000 US businesses. Its lead economist, Ara Kharazian, titled the last three releases "Cracks in the AI thesis".

The finding is a gap between usage and spending. In August and September, business spending on AI tokens showed virtually no growth while the volume of tokens kept rising (Ramp). The reason is price. The effective price per million tokens fell 41% between March and August, from $1.15 to $0.68, as companies routed routine work to cheaper models, and the frontier models' share of tokens fell from 53% at its peak to 45% (Ramp). Anthropic's most capable model, priced at about twice its OpenAI rival, took only about 6% of Anthropic's tokens in July, which Kharazian reads as a new ceiling on what businesses will pay for extra capability (Ramp). Spending per employee at the most AI-intensive 1% of firms fell 3.1% from its July high by September. New-customer growth has "slowed down significantly," and Ramp says customers, not labs, now hold the leverage in contract talks.

This is a plateau, not a collapse. Total AI spending reached a new peak in early October, though only 1.5% above August's, and a similar lull in late 2025 was followed by renewed growth at year-end. Ramp's November release will show which pattern this is.

Fig. 29More tokens, flat spend

What the surveys say

Surveys of executives point the same way: spending is running ahead of measured returns.

• Bain's 2026 survey of 951 companies with more than $100 million of revenue found that most had targeted savings of 11% to 20% from AI, and about 40% of those measuring achieved 10% or less. Yet about 90% are raising AI budgets again, now for agents (Bain).
• In Pigment's third-quarter CFO index, 83% said consumption-based AI costs came in above expectations (reported; I haven't seen the primary).
• Redpoint's CIO survey found 45% are funding AI straight out of existing software budgets (The SaaS CFO), and Morgan Stanley's surveys show total IT budgets growing only about 3.7% in 2026 (Yahoo Finance). Part of AI's growth is coming out of other vendors' revenue, not adding new money.
• Deloitte's UK CFO survey, published in October, found 67% had realised only about a quarter of the change they expected from AI (Bloomberg).

Bain's result is the paradox in one line: savings undershoot, and budgets rise anyway, on the expectation of future returns. That gap is what investors started probing this year.

Agents meet locked doors

The most aggressive demand forecasts rest on agents: AI that acts, booking, buying and moving money, rather than just answering. Jensen Huang has reportedly said an agent can use about a million times more tokens than a simple prompt. In 2026 those agents started meeting locked doors.

• Retail. Amazon sued Perplexity over its Comet shopping agent and won an injunction in March. On 4 August the Ninth Circuit vacated it, holding that the user, not Perplexity, is the one accessing Amazon's servers, and sent the case back to the district court (Engadget). Amazon blocks agents technically anyway: in September it began blocking Meta's new Muse agent (Bloomberg). eBay requires prior approval for AI agents.
• Finance. On 13 October Jamie Dimon said JPMorgan won't admit unvetted outside agents: "You don't want rogue agents." They get read-only balances (PYMNTS). Schwab's chief executive said it stops dozens of attempts by major agents every week.
• Travel and local. Delta says no third-party agent has a partnership to book its tickets. Yelp admits only agents from companies that pay for its data (The Verge via Slashdot).
• Web infrastructure. Since 15 September, Cloudflare has blocked by default the mixed-use crawlers that feed AI search and agents, on ad-carrying pages of new and free-plan sites, nudging AI companies to pay (TechCrunch).
• Business software. SAP's April API policy bars outside agents that plan and execute sequences of calls. ServiceNow will meter and charge outside agents per action, which a JPMorgan analyst called "effectively a tax", and Workday prices agent work in credits (PYMNTS). Gartner has told security chiefs to block AI browsers for now.

There's a counter-trend: Shopify, Walmart and Expedia welcome agents through partnerships, and payment protocols from Google, OpenAI, Visa and Mastercard are maturing. But OpenAI retired its in-chat Instant Checkout in March after weak take-up.

This is friction, not a closed door: the likely end state is a toll-road web of agents that identify themselves, partner and pay. It barely touches coding agents and internal company agents, which run inside a firm's own walls and drive most agentic tokens today. It falls on consumer agents and on agents working across other companies' software, which are what justify forecasts of agent demand at internet scale. Every toll lowers the value of each agent task, and slower, costlier agents mean slower token growth.

Fig. 30Who is gating AI agents

BlockingCharging or permissionedWelcoming

Consumerretail, finance, travel, web infrastructure

Blocking

Amazon blocked Meta's Muse from 21 Sep 2026; robots.txt bars ChatGPT-User, Claude-User, Perplexity-User

eBay prior approval for AI agents from 20 Feb 2026

JPMJPMorgan no outside-agent logins; read-only balances via connectors such as Plaid (Dimon, 13 Oct)

Schwab blocks dozens of major-agent attempts a week (CEO, 16 Oct)

ΔDelta no third-party booking partnerships

Charging or permissioned

Yelp only agents paying for data licensing

Cloudflare default block of mixed-use crawlers on ad pages from 15 Sep 2026; Pay Per Crawl / Pay Per Use

Welcoming

Shopify storefronts open to agents

Walmart Gemini hands off to Sparky; a Muse partner

EExpedia Muse partner (single source)

RHRobinhood an agentic card with limits

Best Buy, Gap, Sephora, Wayfair Muse retail partners

Enterprise softwareAPIs and pricing

Blocking

SAP API Policy v4, Apr 2026

Salesforce / Slack 2025 API limits on bulk export and LLM use

Charging or permissioned

nowServiceNow Action Fabric, metered per action

Workday Flex Credits: 6 per résumé screened, 500 per contract review

Welcomingnone of note

In court: Amazon v. Perplexity

1. injunction, 10 Mar 2026
2. →vacated by the Ninth Circuit, 4 Aug 2026
3. →back in district court

Gartner advised clients to block AI browsers (Dec 2025). OpenAI retired Instant Checkout around Mar 2026.

The other side of the ledger

None of this means demand is falling. The vendors closest to business customers are still growing fast. Anthropic's revenue was more than $11.5 billion in the second quarter, against $787 million a year earlier, with positive adjusted operating income (Bloomberg Government). ServiceNow's AI products passed $1 billion of annual contract value, though its gross margin fell from 81% to 78% as AI raised its own costs (ServiceNow). Gartner forecasts worldwide AI spending up 47% this year, to $2.59 trillion (CIO Dive). And coding agents, the epicentre of the cost blowouts, are also the clearest case of return: engineering leaders in a Gartner survey report an average net productivity gain of about 19%, self-reported.

The closest precedent is the cloud cost-cutting wave of 2022 and 2023, when customers trimmed their cloud bills and growth slowed sharply without reversing. If that's what this is, spending keeps growing, more slowly and at lower prices. That's not a bust. But the loop was underwritten on the faster growth. The risk for the chain isn't that demand vanishes. It's that falling prices eat the volume growth, and that the heavy spenders who carried growth through the first half of 2026 have now installed caps.

My reading: none of this says AI isn't valuable. It says the value is diffuse, slow to appear in accounts, and partly shows up as hiring that didn't happen, which never appears as a line on anyone's income statement. And in 2026 the biggest buyers started treating AI as a cost to manage rather than a race to win. The loop is financed against the easy-to-count side of the ledger and will be repaid, if it is, from the hard-to-count side.

1999, rhymed

The telecom bust supplies the vocabulary for this debate, and about half the mechanics.

In 1999 and 2000, equipment makers lent heavily to the new telephone and internet carriers, many of them cash-burning start-ups known as CLECs (competitive local exchange carriers) funded by junk bonds. At the end of its 2000 fiscal year, Lucent had $8.1 billion of customer-financing commitments, $6.7 billion of loans and $1.4 billion of guarantees, equal to about 24% of its $33.8 billion of revenue. Only $2.0 billion had actually been drawn (Lucent 10-K). The next year it took $2.2 billion of provisions for bad debts and customer financing. The SEC later found that Lucent had improperly recognised about $1.148 billion of revenue in fiscal 2000 (SEC).

The overbuild was spectacular. From the end of 1997 to the end of 2001, US long-distance fibre capacity grew about 500 times while demand grew about four times, per Andrew Odlyzko of the University of Minnesota (slides). The belief that internet traffic was doubling every three months was a myth; it was roughly doubling every year. The bust erased about $2 trillion of market value and left about $1 trillion of debt, "much of which will never be repaid," as the sociologist Paul Starr wrote in 2002 (Starr).

Bigger in scale, better in quality

Measured against that, 2026 looks worse on scale and better on quality.

On scale, Stijn Van Nieuwerburgh's paper for the Brookings Papers on Economic Activity projects AI investment of about $10.3 trillion from 2025 to 2032, an average of 3.63% of US GDP a year. He said "my central estimate is 3.6% of GDP per year" (Brookings). For comparison, the US railroads averaged about 2.2% at their 1870–1890 peak (with individual years above 4%), the interstate highways about 1.1%, the telecom and fibre boom of 1996–2003 about 1.1%, canals about 0.66% and electrification about 0.5%. In his sensitivity runs the AI figure falls to about 3%, which would still be the largest of the lot. Note that this is a projection averaged over eight years, not today's level.

Fig. 31Bigger than the internet build-out

On quality, the comparison favours now. The Nasdaq-100 trades at about 25 times forward earnings, against about 57 times in 2000; Nvidia at about 23 times in July (Motley Fool). Nvidia's customers pay in weeks. The core buyers, Microsoft, Alphabet, Amazon and Meta, generated $451 billion of operating cash flow in 2024; the CLECs had IPO proceeds and junk bonds.

On vendor exposure, the picture is mixed. A widely shared comparison puts Nvidia's financing exposure at 67% of revenue against Lucent's 24%, but that figure includes the dead $100 billion letter of intent (Tunguz). Nvidia's actual equity book of about $99 billion against roughly $303 billion of trailing revenue is about 33%: higher than Lucent's 24%, but much of it is equity in companies that aren't customers. If you count Nvidia's whole stack of equity, rent-back and contingent guarantees, it's closer to one times trailing revenue. Lucent's 24% was also mostly undrawn commitments, so the like-for-like number for Nvidia is larger, not smaller. The difference is in what it's made of: Lucent made unsecured loans to start-ups; Nvidia mostly holds equity, rents capacity and guarantees leases on buildings, with one very large exception tied to OpenAI.

Fig. 32Lucent then, Nvidia now

LULucent, 2000Nvidia, 2026

Vendor financing$8.1B of commitments at 30 Sep 2000 ($6.7B loans + $1.4B guarantees; $2.0B drawn) = 24% of FY2000 revenue of $33.8B≈$99B of equity actually held ≈ 33% of ≈$303B trailing revenue; the full stack with rent-back and guarantees, ≈$300B, is ≈1× a year of revenue67% the ratio only if you count the dead $100B letter of intent

FormUnsecured loans to new phone carriers (CLECs)Equity (much of it in non-customers), rent-back, contingent lease guarantees

Losses$2.2B of provisions in FY2001; the SEC found $1.148B of improperly recognised revenueNone to date

Buyers' fundingIPO proceeds and junk bonds$451B of operating cash flow (Microsoft, Alphabet, Amazon, Meta, 2024), plus a debt-funded middle layer

ValuationNasdaq-100 ≈57× forward earningsNasdaq-100 ≈25×; Nvidia ≈23× (Jul 2026)

OverbuildFibre capacity up ≈500× against demand up ≈4× (1997–2001, Odlyzko)GPU rental prices rising in 2026

Asset lifeFibre: decadesGPUs: one to six years

The asset-life difference cuts against the most comforting part of the analogy. Fibre laid in 1999 was still carrying traffic twenty years later; the overbuild was eventually used. GPUs lose most of their value within a few years. What will outlast this cycle are the buildings, the power connections and the grid interconnects, which is precisely what Nvidia's Ohio guarantee and Meta's Hyperion guarantee cover. That may be the most rational thing about them.

So the analogy fits the middle layer far better than the top. Oracle, the neoclouds, the SPVs and SoftBank are the closest living equivalents of the CLECs: borrowing long against concentrated customers, with funding costs that rise when sentiment turns. Nvidia, Microsoft and Alphabet look much more like the vendors that survived. In 2001, carrier funding dried up by April and vendor write-offs followed over the next six to twelve months. Whether that lag repeats is an analogy, not a forecast.

Bull versus bear, scored by layer

The bull and bear cases are both strong, and they're mostly describing different layers.

The strongest bear arguments are three. The revenue gap: Bain's $800 billion shortfall, and OpenAI's own plan, in which computing absorbs all planned revenue through 2030. Depreciation: Burry's $176 billion and the short economic life of chips. And the shift to debt: funding has moved from cash flow to bonds, leases, SPVs and margin loans, which by October was visible in prices, above all in CoreWeave's record CDS. Since the summer a fourth has joined them: the biggest corporate buyers have started rationing AI, and Ramp's data show spending flat while usage rises.

The strongest bull arguments are also three. Demand is supply-constrained: Nvidia guided about 70% growth when analysts expected 44%. Lab revenue is compounding at rates no telecom customer ever showed: Anthropic's run-rate grew more than sixfold in seven months, and its second-quarter revenue was more than $11.5 billion, against $787 million a year earlier. And the GPU rental rebound shows older chips still earn rent. Executives from Sundar Pichai to Jeff Bezos accept that parts of this are a bubble, while arguing, after the economist Carlota Perez, that bubbles fund infrastructure that outlives them (Stratechery).

Here's how I score it.

Layer	Bull evidence	Bear evidence	Score (bull–bear, out of 5)	Verdict
Top: Nvidia, TSMC, Microsoft, Alphabet	Audited, supply-constrained demand; ~70% FY28 guide; 25x not 57x; payment in weeks	~$300B exposure stack; receivables up to ~60 days; concentration on five customers	4–1	Bull
Middle: Oracle, neoclouds, SPVs, SoftBank	Take-or-pay contracts with investment-grade names; Meta-backed loan rated A3 and self-amortising; GPU prices up; substitutes keep lending	Oracle BBB-, ~288bp CDS; CoreWeave interest 5x operating income, bonds 14.5–16%, CDS a reported 1,194.5bp; neoclouds own nothing scarce and depend on one to three customers; margin loans on self-marked collateral	1–4	Bear, weakest at the neoclouds
Demand base: OpenAI, Anthropic, xAI	Run-rates compounding; Anthropic ~3.5x covered; 900M ChatGPT users	OpenAI ~14.6x; −$278B FCF plan; revenue counted differently; supplier-funded rounds	2–3	Lean bear, OpenAI-specific
Enterprise value	Large task-level gains; adoption broadening; tokens up ~7x a year; ~90% of Bain's firms raising AI budgets; Gartner +47% spending forecast	Near-zero economy-wide effects so far; median spend ~$12/month; tokenmaxxing reversed; Ramp token spend flat as price per token fell 41%; agents gated at high-value sites	2–2	Open, and now the crux

Fig. 33Bull versus bear, by layer

LayerStrongest bull dataStrongest bear dataBalance

Top

chips, foundry, cloud giants

TSMC

• Nvidia guides ≈70% growth for FY28 vs ≈44% consensus
• Nasdaq-100 at ≈25× forward earnings vs ≈57× in 2000

• Nvidia exposure ≈$300B; receivables ≈60 days of sales

4–1Bull

Middle

Oracle, neoclouds, SPVs, SoftBank

• GPU rental prices rebounded ≈40%

• Oracle at BBB−, CDS ≈288bp
• CoreWeave: interest ≈5× operating income; bonds 14.5–16%; CDS a reported 1,194.5bp

1–4Bear, weakest at the neoclouds

Demand base

OpenAI, Anthropic, xAI

• Anthropic run-rate ≈$9–10B → >$65B in seven months

• OpenAI commitments ≈14.6× revenue; −$278B planned free cash flow to 2030

2–3Lean bear, OpenAI-specific

Enterprise value

does AI pay for its users?

• Task-level trials: +14% to +26% productivity

• Denmark: effects larger than 2% ruled out
• Ramp: median ≈$12 a month; token spend flat as the price per token fell 41%

2–2Open, and now the crux

The scores are judgement, not arithmetic. But the pattern is robust to any reasonable reweighting: the top is in good health, the middle is not, and everything depends on whether the base grows into its commitments.

What would change my mind

A verdict is only useful if you say what would overturn it. These are the things I'm watching, in date order where there is a date.

1. Hyperscaler third-quarter earnings, around 28 to 30 October. The first 2027 capex signals, and whether Microsoft updates the OpenAI share of its backlog. A cut to 2027 plans from any of the four would be the most important bearish signal on this list. Another round of increases would extend the cycle.
2. AMD's third-quarter filing, early November. Whether the first tranche of OpenAI's warrant has vested, which would mean the first gigawatt of AMD chips is actually running.
3. CoreWeave's third-quarter results, November, and its next refinancing. Interest guided at $860 to $940 million against operating income in the low hundreds of millions. Any new borrowing, and its price, will say more than the stock. Unsecured money raised well below the 14.5% to 16% its bonds yielded in mid-October would weaken my bearish score on the middle; new money above the roughly 12.8% cash yield on its equipment would mean it is borrowing at a loss.
4. Nvidia's $1 billion Nscale convertible, funding expected mid-November, and the Nscale IPO behind it. A neocloud IPO that prices well would show the funding window is open.
5. SoftBank's September-quarter results, November, and whether the second $10 billion margin loan closes.
6. Ramp's November AI Index, mid-November. Whether business token spending grows again, as it did at the end of 2025, or stays flat while volumes rise. Two more flat months would move me towards the bear case on the base.
7. Nvidia's next quarterly report, guided at $108 billion. I'll be reading the notes rather than the headline: AI cloud agreements, guarantees, receivables and days to collect. If rent-back and guarantees keep growing faster than revenue, the loop is doing more of the work.
8. Oracle's next quarterly results, December, and S&P's outlook. One more notch is junk.
9. 1 February 2027, the revised gas-pipeline date for Project Jupiter.
10. March 2027, when SoftBank's $40 billion bridge loan matures.
11. The first public S-1 from Anthropic or OpenAI. Anthropic confidentially submitted a draft on 1 June 2026. A public prospectus would give the first audited, comparable revenue figures. If audited revenue tracks the run-rates, the demand base looks much stronger.
12. Amazon v. Perplexity, back in the district court. Whether Amazon seeks a rehearing or Supreme Court review, and whether the fight ends in a paid access deal, as JPMorgan's dispute with Plaid did. A licensing settlement would confirm the toll-road web; a renewed injunction would close more doors.
13. Token-cap news. Whether more large employers cap AI spending per employee, and whether Meta's enforceable team quotas arrive in 2027 as reported. The labs' next revenue disclosures will show whether the caps are biting.

Beyond the dates, three signals would move me towards the bull case: OpenAI reaching its targeted $70 billion run-rate by year-end on its own net basis; GPU rental prices and used-GPU values holding through the Blackwell and Rubin supply wave; and the Census adoption rate rising past 30% of firms with median spend rising alongside it. Three would move me further towards the bear case: any missed payment or restructuring by a neocloud or SPV; a down round at OpenAI, which would hit SoftBank's collateral directly; or business token spending staying flat into 2027 while Nvidia's guarantees and rent-back keep growing.

Verdict

Is the AI boom a circular-financing bubble? No, not as a whole. It's a credit cycle with a strong top, a weak middle and an unproven base.

The loop is real and large. Nvidia, Amazon, Microsoft and Google all fund customers who buy their products, and the BIS finds more than half of the money going into AI firms comes from other AI firms. But the top of the loop is not inflating its own revenue in any way that matters for its solvency. Nvidia's demand is audited and supply-constrained, its buyers have the cash flows the CLECs never had, and its valuation is less than half of what 2000 looked like.

What changed in 2026 is where the risk sits. Each time scrutiny landed on an obvious instrument, it was replaced by a less visible one. The $100 billion letter of intent became $30 billion of cash and $105 billion of guarantees. OpenAI's data centres became other people's leases. Hyperscaler capex became bonds, equity raises and $617 billion of leases that haven't started. The risk didn't leave the system. It moved down the stack, into Oracle's lease book, CoreWeave's bonds, SoftBank's margin loans and long-dated project bonds held by a handful of anchor buyers. That's where the repricing started, just as it did with the CLECs in 2001.

My base case for the next twelve months is that the top keeps growing and the middle keeps repricing, with at least one forced refinancing, restructuring or rescue in the neocloud and SPV layer before Nvidia's revenue stops rising. That would hurt lenders and equity holders in the middle. It wouldn't, on its own, end the buildout, because the substitutes so far have always shown up.

The variable that decides the longer story is the one hardest to measure: whether the labs' revenue, counted on a comparable and eventually audited basis, keeps compounding fast enough to honour roughly $1 trillion of contracts. If it does, the middle gets paid and the 1999 analogy fails. If it stalls, the contingent credit Nvidia and Google have written becomes real exposure, and the cost of short-lived chips lands first on the lenders. Today's best evidence, big gains on individual tasks that don't yet show up in the economy, broad but shallow adoption, and big buyers now rationing rather than retreating, doesn't settle it either way. What it does show is that the end customer has become price-sensitive, and the loop was underwritten as if it weren't.

So the right question isn't whether AI is a bubble. It's which layer you're exposed to. At the top, I'd worry about valuation, not solvency. In the middle, and above all at the neoclouds, I'd worry about solvency first. At the base, I'd watch the S-1s and Ramp's monthly numbers.

Sources

Nvidia

• Nvidia, OpenAI and NVIDIA announce strategic partnership to deploy 10 gigawatts of NVIDIA systems, 22 September 2025
• Nvidia, Form 10-Q, quarter ended 26 July 2026, August 2026
• Nvidia, Form 10-Q, quarter ended 26 April 2026, May 2026
• Nvidia, Form 10-K, fiscal 2026, February 2026
• Nvidia, Financial results for the second quarter of fiscal 2027, 26 August 2026
• Nvidia, Partnerships with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR to mobilise over $500 billion, 10 August 2026
• Nvidia, Schedule 13G/A on CoreWeave, January 2026
• CNBC, Nvidia's AI investments reach $99 billion, 4 September 2026
• CNBC, Nvidia Q2 fiscal 2027 earnings live updates, 26 August 2026
• Fortune, Jensen Huang on the OpenAI investment: "never a commitment", 2 February 2026
• Bloomberg, Nvidia's Jensen Huang rules out $100 billion OpenAI investment, 4 March 2026
• Investing.com, Nvidia and OpenAI's $100 billion deal is dead; what replaced it matters more, 2026
• Investing.com, Nvidia rebuts Burry-linked allegations in analyst memo, November 2025
• Benzinga, Michael Burry points to Nvidia's $105B OpenAI guarantee, August 2026
• TNW, Nvidia in talks to guarantee $250bn of OpenAI data-centre financing, 27 July 2026
• DCD, Nvidia signs $1.5bn deal to lease its GPUs back from Lambda, September 2025
• Nscale, Pre-IPO convertible financing, 25 September 2026
• Fortune, Nscale wants a $35 billion valuation, 24 September 2026

The labs

• CNBC, OpenAI closes record-breaking $122 billion funding round, 31 March 2026
• Amazon, Form 8-K exhibit on the OpenAI agreement, February 2026
• GeekWire, Filings: how Amazon's $50B OpenAI deal actually works, August 2026
• TechCrunch, OpenAI's revenue is reportedly $20 billion less than previously projected, 8 October 2026
• Bloomberg, OpenAI expects $70 billion in annualized revenue by end of 2026, 9 October 2026
• The Star (Reuters), OpenAI forecasts cash burn near $280 billion by 2030, FT reports, 19 September 2026
• Manila Times, SoftBank completes $30B OpenAI investment, 3 October 2026
• AMD, Form 8-K on the OpenAI agreement, 6 October 2025
• AMD, Form 10-Q, quarter ended 27 June 2026, 2026
• Tom's Hardware, OpenAI now prefers to lease compute and says Stargate is an umbrella term, 2026
• CNBC, Anthropic says annualized revenue climbed to $65 billion in July, 17 August 2026
• CNBC, Amazon to invest up to $25 billion in Anthropic, 20 April 2026
• Morningstar, Lifting xAI's hood in the SpaceX IPO, 2026

Hyperscalers

• FactSet, Hyperscalers tap external financing as AI capex outruns cash flow, 2026
• Benzinga, Microsoft's $15 billion capex cut isn't a cut at all, July 2026
• Microsoft, Form 8-K, fiscal 2026 fourth-quarter results, 29 July 2026
• Yahoo Finance, Big Tech's cloud backlog just hit $2.3 trillion, 2026
• Bloomberg, Alphabet upsizes equity offering to $85 billion, 3 June 2026
• Meta, Form 10-Q, quarter ended 30 September 2025, October 2025
• StreetInsider, Michael Burry warns of $176 billion depreciation understatement, November 2025

Oracle, CoreWeave and the neoclouds

• Oracle, Form 10-Q, quarter ended 31 August 2026, September 2026
• Computing, Oracle under pressure after ratings downgrade, July 2026
• TNW, Oracle force majeure on Project Jupiter, September 2026
• CoreWeave, Form 10-K, fiscal 2025, 2026
• CoreWeave, Form 8-K on the Nvidia capacity order, September 2025
• CoreWeave, Second-quarter 2026 results, 11 August 2026
• Yahoo Finance, CoreWeave stock plunges as bond yields spike, 14 October 2026
• Reuters via Investing.com, Tech stocks tumble as data-centre debt worries hit AI trades, 14 October 2026
• Blockspace, IREN closes GPU financing on Microsoft contract, May 2026
• TeraWulf, Form 10-Q, quarter ended 30 June 2026, 2026
• Seeking Alpha, Nvidia's H100 rental prices surge nearly 40% in six months, 2026

How a neocloud works, and why it has no moat

• CoreWeave, Form 10-Q, quarter ended 30 June 2026, August 2026
• CoreWeave via Business Wire, CoreWeave closes $8.5 billion delayed-draw term loan, 30 March 2026
• Morningstar DBRS, Credit rating of A (low) for CoreWeave Compute Acquisition Co. VIII, 3 April 2026
• CoreWeave, DDTL 3.0 press release (SEC exhibit), 31 July 2025
• Rittenhouse Research, CoreWeave unit economics and margin profile, 20 February 2026
• Cape Fear Advisors, CoreWeave: twenty-seven years, 12 July 2026
• Yahoo Finance, Jim Chanos sees risks in growing debt market backed by Nvidia's AI chips, 1 December 2025
• IntuitionLabs, Used AI GPU prices and resale trends, 2026
• Microsoft, Fiscal 2026 fourth-quarter earnings call, July 2026
• Teahose, Transcript: Satya Nadella on the Dwarkesh Podcast, November 2025
• Fortune, Nscale's $35 billion IPO filing and going-concern history, 21 September 2026
• Computerworld, Exodus seeks Chapter 11 bankruptcy protection, September 2001
• CoinDesk, Core Scientific files for bankruptcy protection, 21 December 2022
• TNW, Huang says CoreWeave needs no Nvidia backstop, 15 October 2026
• The Star (Reuters), CoreWeave stock and bonds set for weekly plunge as rate-hike fears hit AI bet, 16 October 2026

Capital, Stargate and SoftBank

• Cleary Gottlieb, Financing the data center boom, 2026
• DCD, JPMorgan: global data centre and AI infrastructure spend to hit $5 trillion, 2025
• Epoch AI, OpenAI Stargate: where the US sites stand, April 2026
• DCD, JPMorgan and MUFG lead $38bn debt package for Oracle-linked data centres, November 2025
• Blackstone, Related Digital announces financing for $16 billion Oracle data centre in Saline Township, April 2026
• Bloomberg, SoftBank cuts target for OpenAI margin loan to $6 billion, 8 May 2026
• Bloomberg, SoftBank secures $10 billion margin loan backed by OpenAI stake, 6 August 2026
• InvestmentNews, SEC exemption clears path for more data-centre asset-backed bonds, July 2026

Regulators and macro

• IMF, Global financial markets face rising risks from AI, bond yields and cyberattacks, 12 October 2026
• IMF, World Economic Outlook, October 2026, Box 1.2, October 2026
• BIS, Bulletin No 137: Circular relationships among AI firms, 1 October 2026
• Brookings, BPEA podcast: financing the AI buildout (transcript), 8 October 2026
• ING Think, How much is AI contributing to US economic growth?, 2026
• Bain & Company, $2 trillion in new revenue needed to fund AI's scaling trend, 2025

Enterprise value and adoption

• US Census Bureau, AI use by businesses, May 2026
• Ramp, AI Index, October 2026, October 2026
• Google, I/O 2026: Welcome to the agentic Gemini era, May 2026
• a16z, LLMflation: LLM inference cost, 2024
• Palantir, Second-quarter 2026 results, 2026
• Wharton and GBK, 2025 AI adoption report, executive summary, October 2025
• McKinsey, The state of AI, 2025
• Brynjolfsson, Li and Raymond, Generative AI at work, NBER, 2023
• Noy and Zhang, Experimental evidence on the productivity effects of generative AI, Science, 2023
• Cui et al., The effects of generative AI on high-skilled work, SSRN, 2024
• Dell'Acqua et al., Navigating the jagged technological frontier, HBS, September 2023
• METR, Early-2025 AI and experienced open-source developer productivity, July 2025
• Humlum and Vestergaard, Large language models, small labor market effects, NBER, 2025
• Stanford Digital Economy Lab, Canaries in the coal mine, August 2026 update, August 2026
• Acemoglu, The simple macroeconomics of AI, NBER, 2024

End-customer ROI and token rationing

• JPMorgan Asset Management, Technology and AI investment outlook, 2026
• CNBC, Big Tech capital expenditures now seen topping $1 trillion in 2027, 30 April 2026
• Letsdatascience, Uber president says AI spending is getting harder to justify, May 2026
• TNW, Tokenminimizing: companies cap employee AI spending, 17 June 2026
• TNW, Microsoft sets AI spending limits, 4 August 2026
• TechCrunch, Companies are scrambling to stop employees from maxing out AI budgets, 24 June 2026
• AI Weekly, Meta caps employee AI token use as costs approach billions, June 2026
• Ramp, AI Index, August 2026, 12 August 2026
• Ramp, AI Index, September 2026, 9 September 2026
• Bain & Company, Your AI budget is growing; your returns aren't, 2026
• Open Future Forum, CFO AI leverage report (Pigment CFO Index), October 2026
• The SaaS CFO, The AI budget reallocation is coming, 2026
• Yahoo Finance, Morgan Stanley survey shows modest IT budget growth, 2026
• Bloomberg, AI to cut headcount at more than half of UK firms, CFOs say, 13 October 2026
• Bloomberg Government, Anthropic revenue surges to over $11.5 billion in second quarter, 2026
• ServiceNow, Second-quarter 2026 results (8-K exhibit), July 2026
• CIO Dive, Global AI spending forecast for 2026, May 2026
• Gartner, Enterprise AI coding agent market, 2026

Agents and the sites that block them

• CNBC, Amazon wins court order to block Perplexity's AI shopping agent, 10 March 2026
• Engadget, Perplexity overturns Amazon injunction on its AI shopping bot, 4 August 2026
• Bloomberg, Amazon blocks Meta's Muse AI agent from its retail site, 21 September 2026
• PYMNTS, Jamie Dimon slams the door on rogue AI agents, 14 October 2026
• Slashdot (The Verge), Websites are closing their doors to AI agents, 18 October 2026
• TechCrunch, Cloudflare's new policy pushes AI companies to pay for publishers' content, 1 July 2026
• PYMNTS, ServiceNow, SAP and Workday make AI agents pay to play, 7 May 2026
• Gartner, Cybersecurity must block AI browsers for now, December 2025

History

• Lucent Technologies, Form 10-K405, fiscal 2000, December 2000
• Lucent Technologies, Form 10-K405, fiscal 2001, December 2001
• SEC, Lucent settles SEC enforcement action, 17 May 2004
• Andrew Odlyzko, What caused the telecom crash?, 2002
• Paul Starr, The great telecom implosion, September 2002
• Tomasz Tunguz, Nvidia and Nortel: vendor financing compared, 2025
• Motley Fool, Nvidia's forward P/E has actually fallen, 11 July 2026
• Stratechery, The benefits of bubbles, 2025