Who Pays for the AI Build-Out? Capex, Debt, Neoclouds and the Circular Money Machine
In 2026 four companies alone will spend roughly $700 billion on data centres, more than the annual GDP of most countries. Part of it comes from the most profitable businesses ever built. A growing share comes from bonds, private credit, GPU-backed loans, off-balance-sheet joint ventures and deals in which chipmakers invest in their own customers. This guide follows the money layer by layer, asks whether the revenue is there to pay it back, and shows who is left holding the risk if it is not.
๐ฐ About $700 billion in one year
The earlier parts of this series followed electricity, factories, chips, packaging, networks and memory. Every one of them is paid for by someone. In 2026 the four biggest spenders, Amazon, Alphabet, Microsoft and Meta, are on track for combined capital expenditure of roughly $700โ745 billion, up from about $410 billion in 2025, according to trackers compiled from company guidance. Add Oracle, the neoclouds, the AI labs' own projects, sovereign programmes and Chinese spending, and the total is far larger.
For most of the boom this was paid for out of profits. That is changing. One estimate has capital spending absorbing about 94% of the tech sector's operating cash flow in 2025โ2026, and the gap is increasingly filled with borrowed money, much of it in structures designed to keep it off the biggest companies' balance sheets.
๐บ๏ธ The money map
The money moves in three stages: sources of capital, the builders who spend it, and the suppliers who receive it. Some arrows loop back: a supplier invests in a builder, who spends the money with that supplier.
๐ Follow a dollar
Click any box for details. Toggle "Show circular loops" to highlight money that returns to where it started.
Simplified. Flows are illustrative of structure, not to scale.
๐ฆ Layer 1: paying from cash flow
The hyperscalers' advantage is that they own some of the most profitable businesses in history: search and YouTube ads, Office and Azure, Instagram and WhatsApp ads, AWS. Their operating cash flows run to well over $100 billion a year each, so they can fund most of their build-out themselves. But capex is now growing faster than cash flow, squeezing free cash flow and buybacks, and pushing even these companies to borrow.
๐ 2026 capex guidance by company
Midpoint of reported guidance before Q3 2026 results. Ranges in the labels; sources differ.
Compiled from capex trackers (Introl, CreditSights, others). Amazon may have raised to ~$220B; Alphabet reported as $180โ190B or $195โ205B; Microsoft ~$175โ190B depending on how finance leases are counted; Meta $125โ145B. Check Q3 earnings releases (late October) for current figures.
๐ Layer 2: the bond wave
Until 2024, the biggest tech companies barely borrowed. Bank of America counts US corporate bond issuance by the five AI hyperscalers (Amazon, Alphabet, Meta, Microsoft, Oracle) at about $28 billion a year on average in 2020โ2024, then $121 billion in 2025. Big deals came one after another: Oracle $18 billion (September 2025), Meta $30 billion (October 2025), Alphabet $17.5 billion and Amazon $15 billion (November 2025), and Alphabet and Oracle together nearly $60 billion in February 2026. By August 2026 Big Tech had issued roughly $200 billion of bonds, and Goldman Sachs counted $489 billion of AI-related debt worldwide.
๐ Hyperscaler bond issuance
US corporate bonds, $ billions. The 2026 bar is a partial year and uses a different, broader count.
BofA via Reuters (2020โ24 average, 2025); Business Today (2026 to August, "Big Tech"). Not strictly comparable.
๐งฉ Layer 3: off the balance sheet
The biggest private-credit deal ever done shows the new playbook. To build its Hyperion AI campus in Richland Parish, Louisiana, Meta formed a joint venture with Blue Owl Capital. The venture, a special-purpose vehicle (SPV), sold $27.3 billion of fully amortising senior secured notes due 2049, anchored by PIMCO and structured by Morgan Stanley. Blue Owl owns about 80% and Meta about 20%. S&P rated the notes A+, yet they paid about 6.58%, well above Meta's own bonds.
๐๏ธ How the Hyperion deal works
Click the numbered steps.
Why do it? The debt does not appear as Meta's borrowing; rating agencies treat Meta's commitment as a lease. That protects Meta's credit metrics and spreads the funding across investors hungry for long, high-quality paper. The catch: through the lease and a residual value guarantee, Meta still carries most of the risk. If it walks away and the campus is worth less than promised, Meta pays the difference. Critics note that S&P's rating leaned heavily on Meta's own AA- credit. Other hyperscalers and labs have been reported to be pursuing similar joint ventures.
โ๏ธ Layer 4: neoclouds and GPU-backed debt
Neoclouds are companies that rent GPUs, not general cloud services: CoreWeave, Nebius, Lambda, Crusoe, Fluidstack and hundreds of smaller operators. Their model is simple and highly leveraged. Sign a multi-year contract with a big customer (Microsoft, OpenAI, Meta, Google), borrow against that contract and the GPUs themselves, buy the GPUs, and repay the loan from the rent.
- CoreWeave
- The largest. Its Q2 2026 filing showed $35.6 billion of debt and $16.3 billion of operating-lease liabilities, against $103.7 billion of contracted future revenue (41% due by 2028). Its GPU loans sit in separate borrowing entities secured by chips and customer contracts. The August 2026 "DDTL 5.5" facility cost SOFR + 5.5% (about 9.12%), up from about 8.05% in May, and was rated one notch below investment grade. Goldman Sachs estimated it still needs up to ~$23 billion of new debt to deliver its backlog. Its bond yields and CDS spreads jumped between 15 and 20 October 2026.
- Lambda
- Closed a $1.0 billion fixed-rate loan at 6.78% in October, rated Baa1 (investment grade) because it is backed by two investment-grade customers.
- Nebius
- Funds its Microsoft and Meta capacity largely with asset-backed debt tied to those contracts, including a $775 million GPU-backed term loan.
- Crusoe ยท Fluidstack
- Crusoe has leaned on equity (a reported $3.9 billion Series F). Fluidstack has arranged roughly $10 billion of GPU-collateralised financing.
๐งฎ Does a GPU pay for itself?
๐๏ธ GPU loan payback model
One GPU bought with borrowed money and rented out. All defaults are illustrative, not any company's figures.
Simple model: straight-line depreciation over the useful life, interest on the average outstanding balance, operating cost per rented hour. Ignores taxes, residual value and contract terms.
๐ Layer 5: circular deals
The most debated financing comes from the suppliers themselves. In several deals, a company that sells chips or cloud capacity invests in, or extends credit to, a customer who then spends the money on that same company's products.
๐ The deals
Click a deal.
Estimates of all the interlocking commitments range from about $800 billion to $1.5 trillion, and the counts overlap. Defenders call this ordinary vendor financing, which industries from aircraft to telecoms have long used, and note the chips are real and in use. Critics point out that it inflates the suppliers' revenue with money they supplied, ties many companies' fortunes to one customer (OpenAI), and resembles the telecom-equipment financing that unwound badly in 2001.
๐ Is there enough revenue to pay it back?
In June 2024, Sequoia partner David Cahn asked "AI's $600B question": how much revenue would AI need to justify the GPUs being bought? By his 2026 update, as reported, about $1.5 trillion of infrastructure spending implies roughly $3 trillion of revenue. Goldman Sachs uses a different method and puts breakeven at about $308 billion a year of AI revenue on 2026โ27 capex. The answer depends on assumptions, which you can set below.
Revenue is growing very fast. Bloomberg reported that OpenAI was at roughly a $50 billion annualised run rate in late September and expects $70 billion or more by year-end, though its numbers are disputed because it and Anthropic count cloud-partner sales differently. Anthropic was reported at a $65 billion run rate at the end of July, with investors reportedly projecting $100โ110 billion by year-end. Cloud AI revenue at the hyperscalers adds more. But these figures are still small next to the spending.
๐งฎ The revenue gap calculator
How much annual AI revenue does a year of capex need? Illustrative model in the spirit of Sequoia's calculation.
Required revenue = (capex รท years) ร (1 + operating share) รท (1 โ margin). Comparison bar: OpenAI's expected ~$70B plus Anthropic's projected ~$100B year-end run rates (press reports, not audited, partly overlapping definitions).
โณ The depreciation fight
How long does a GPU last? The answer decides how much profit the build-out shows. Hyperscalers generally depreciate servers over five to six years. Investor Michael Burry has argued the real economic life of AI chips is closer to two or three years, because Nvidia releases a much better generation every year, and that the gap overstates profits by about $176 billion in 2026โ2028 (a figure critics say is directionally plausible but unproven). Amazon has already cut the life of some servers from six years to five.
๐๏ธ What useful life does to the books
$100 billion of GPUs bought in one year. Drag the life assumption.
Straight-line depreciation, no residual value. Real hardware is often redeployed to cheaper work (inference, smaller models) late in life, which supports longer lives.
โ๏ธ Who holds the risk?
If AI revenue falls short, the losses do not land evenly. They flow down a ladder from the most protected investors to the least.
๐ช The risk ladder and three scenarios
Pick a scenario to see who gets hurt. Qualitative and illustrative.
๐ฐ๏ธ Timeline
๐ญ What to watch
- Q3 capex guidance (late October) and 2027 previews: still rising, or plateauing?
- Neocloud refinancing. Whether CoreWeave and peers can keep borrowing, and at what rates.
- Lab revenue and IPOs. Whether OpenAI and Anthropic keep compounding revenue, and how public markets price them.
- Credit spreads and CDS on Oracle, CoreWeave and the SPV bonds.
- Depreciation policies. Any further cuts to server lives would hit reported profits.
- Token prices. As The Token Price Collapse showed, cheaper tokens grow usage but squeeze revenue per token.
๐ข Reading the numbers
- Capex figures are pre-Q3 guidance from trackers that disagree, partly over whether finance leases count.
- Bond totals use different scopes (five hyperscalers in the US, "Big Tech", all AI-related debt worldwide) and are not directly comparable.
- CoreWeave numbers come from reporting on its Q2 2026 10-Q; "backlog" and "remaining performance obligations" are measured differently.
- Lab revenue is run rates (current month ร 12) from press reports of private companies, with disputed definitions.
- Circular-deal totals overlap and are estimates. The NvidiaโOpenAI amount was a letter of intent, not a binding commitment.
- Sequoia's $3T and Burry's $176B are reported second-hand.
- All calculators use illustrative inputs. Nothing here is investment advice.
๐ References
- Capex and bonds
- Introl: Hyperscaler capex hits $690B in 2026
- CreditSights: Raising hyperscaler capex 2026 estimates
- Reuters: AI hyperscalers will drive higher US corporate bond supply (BofA figures)
- The Global Treasurer: The $489B AI debt wave (Aug 2026)
- Business Today: Is Big Tech borrowing too much for the AI boom?
- Bloomberg (via EnergyNow): The $3 trillion build-out and debt markets
- Livewire: Tech companies are piling on debt. Should investors care?
- Off-balance-sheet and neoclouds
- Global Data Center Hub: Meta and Blue Owl's $27B bet
- Details of Wall Street's biggest private-credit deal (Meta, Blue Owl)
- CCIR: CoreWeave's debt, entity by entity
- Goldman Sachs on CoreWeave's funding gap (BigGo Finance)
- Yahoo Finance: CoreWeave's debt sparks AI contagion fears
- MarkTechPost: GPU neoclouds 2026
- Circular deals, revenue and depreciation
- Global Finance: AI's financial circle game
- Noah Smith: Should we worry about AI's circular deals?
- Yahoo Finance: Nvidia, OpenAI and Oracle's financing circle
- Sequoia (David Cahn): AI's $600B question (2024)
- AI Weekly: Cahn pegs the 2026 revenue gap at roughly $3T
- Bloomberg: OpenAI expects $70B annualised revenue by end of 2026
- Yahoo Finance: OpenAI's annualised revenue and the methodology dispute
- Simon Willison: Anthropic's run-rate revenue (May 2026)
- YieldTheory: Is Michael Burry right about AI depreciation?
- Related on this site
- The Token Price Collapse ยท Frontier lab compute economics ยท Custom Silicon vs Nvidia ยท The Power Bottleneck