How the AI buildout is financed
A map of the AI trade as it stands. Where the money comes from, where it goes, who is carrying the asset, who is carrying the liability, and which of those two lists the public is on. Every figure from a filing, an auction result or a regulator.
Where it stands
Microsoft, Alphabet, Amazon and Meta will spend $720 to $745 billion on capital in 2026. That is 2.24% of a $32.38 trillion economy, in one year, by four companies.
Nvidia is the largest company in the world at roughly $5.4 trillion. The seven largest US technology firms are worth more than $22 trillion together and make up 34% of the S&P 500, the highest single-cohort concentration in the index's modern history. Nvidia alone is 7.5% of it. Total US market value sits at 234.6% of GDP, against 150 to 172% at the dot-com peak.
Anthropic reached a $65 billion run rate in late July and posted a first operating profit of $559 million. OpenAI passed $40 billion in August while burning about $27 billion for the year. Google Cloud grew 82% year over year, AWS 37%, Azure 43%.
Where the money comes from
Operating cash flow. This remains the largest source. It no longer covers the spending. Alphabet spent $44.9 billion on capital in the June quarter against $39.1 billion generated by operations, producing its first negative free cash flow quarter since the 2004 IPO.
The bond market. Amazon, Alphabet, Meta and Oracle issued about $194 billion through 7 July. Goldman expects roughly $250 billion across the five hyperscalers this year and $400 billion next. The terms have moved. Orders covered the February deals nearly five times over. The first unscheduled deal of the second half was covered 1.4 times.
Equity. Alphabet raised $49.6 billion of stock in June. Oracle exhausted a $20 billion at-the-market programme inside one quarter, issuing 141 million shares for $19.9 billion net.
Other people's credit. Nvidia has guaranteed up to $105 billion of OpenAI's lease payments in Ohio. Google guarantees Anthropic's lease and power obligations. On 10 August, Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR announced platforms with Nvidia to "mobilize over $500 billion of third-party capital". Third-party capital means pension funds, endowments and insurers.
Where it goes
Most of it goes to Nvidia. The company booked $96.2 billion of revenue in the June quarter, $89.0 billion of it data centre, on $59.7 billion of net income.
From Nvidia the money goes to memory and fabrication. Nvidia has committed $279 billion to its own suppliers, up from $119 billion one quarter earlier. The filing describes the commitments as "primarily memory and manufacturing facilities" without naming counterparties. SK Hynix holds roughly 50 to 55% of the high-bandwidth memory market, Samsung 35 to 40% and Micron 5 to 10%. TSMC fabricates and packages the compute dies. At more than $4,800 of memory per Rubin-class accelerator, memory is the highest-cost line in Nvidia's bill of materials. That lands in South Korea, where semiconductors were a record $46.65 billion of exports in August, 47.5% of all goods exports, with memory roughly 90% of that.
The rest goes into land, power and construction. PJM's forward capacity cost went from $2.2 billion to $16.4 billion on the back of it. Texas froze new data-centre grid connections in September against 474 gigawatts of requests.
Nvidia, OpenAI and Oracle each appear on both sides
A London credit fund, Sona Asset Management, put $3.6 trillion of AI financing through 176 deals between 202 companies and flagged about 120 as ones where the same firms appear as investor, customer and supplier.
- Nvidia commits about $100 billion to OpenAI.
- OpenAI commits hundreds of billions to Oracle, Microsoft and Amazon for cloud capacity.
- Oracle, Microsoft and Amazon are among Nvidia's largest chip customers.
- Nvidia books the revenue.
Amazon's version is tighter still. It invested about $50 billion in OpenAI against OpenAI's commitment to buy about $100 billion of compute from Amazon.
The same structure appears within a single project. At the Ohio campus Nvidia guarantees the tenant's lease, holds $1.5 billion of equity in the landlord, supplies the only hardware the site is contractually allowed to host, then may earn a share of revenue from the capacity sold. Nvidia therefore occupies four positions against one asset and one stream of expected demand.
Where the lending money originates
The sequence above moves revenue between firms. The money that funds the buildout arrives by five routes. Each begins with households.
Deposits. Households bank. Banks lend against commercial property and data-centre construction. When those loans stop being sellable, the bank pledges them to the Federal Reserve's discount window and takes cash without a sale, then sells the credit risk on the same loan through significant risk transfer. The synthetic securitisation pool has passed $670 billion and now references commercial real estate directly.
Pension and endowment money. Limited partners commit to Apollo, Blackstone, Brookfield and KKR. Those houses buy data centres. On 10 August they agreed platforms with Nvidia to raise $500 billion more of it.
Insurance premiums. The same houses own life insurers. Apollo holds Athene and KKR holds Global Atlantic. Private equity controls at least $1.5 trillion of life-insurer assets. In 2024, 49.5% of new investment by private-equity-owned insurers went into privately placed instruments, against 14% at independent insurers. Those insurers are the natural buyers of the credit risk the banks are selling.
Index funds. Retirement money tracks the S&P, where seven companies are 34% of it. It also tracks investment-grade bond indices. Those now carry the hyperscaler paper.
Utility bills. Capacity charges fund the grid the buildout needs. PJM's forward capacity cost is $16.4 billion for 2027 to 2028, built on interconnection forecasts that Wood Mackenzie puts at 72% phantom.
Each route begins with household money and ends at the same set of buildings. No party outside that group supplies material funding. Households therefore provide the lending and hold the contingent liability.
Who carries the asset and who carries the liability
| The asset | Who carries it | The matching liability | Who carries that |
|---|---|---|---|
| Ohio campus, 4.25 GW | SB Energy, as owner | 20-year lease | OpenAI, guaranteed up to $105bn by Nvidia |
| Data-centre buildings | SB Energy and other landlords | $288bn of leases not yet commenced | Oracle, as tenant, off its balance sheet |
| Louisiana data-centre venture | an unconsolidated vehicle, Meta holding 20% | ~$28bn residual value guarantee, no liability booked | Meta, disclosed in a 10-Q footnote |
| Compute contracts | OpenAI, as buyer | ~$1.4tn of commitments | OpenAI, against ~$40bn of revenue |
| GPUs and memory | Nvidia, as buyer | $279bn of supply commitments | Nvidia |
| Money owed by customers | Nvidia, $63.1bn, up from $38.5bn | longer payment terms granted to those customers | Nvidia |
| Office and multifamily loans | Banks, at carrying value | credit risk sold via SRT | Private credit funds |
| The first-loss slice of that credit risk | Private credit funds | policyholder claims | Life insurers owned by the same houses |
| Mag 7 equity, 34% of the index | Index funds | none, by construction | The household holding the fund |
| Grid capacity reserved | Nobody yet, 72% is phantom | capacity charges | Ratepayers, already paying |
On almost every row the asset and the liability sit with different parties. The party carrying the liability is further from the upside than the party carrying the asset.
Where the loss sits today
Very little has been realised. The realised losses fall in two places.
Equity holders who bought near the peak. South Korea's KOSPI went from about 5,000 in March to above 9,000 in June, then fell more than 16% across two sessions, removing roughly $2 trillion. Samsung was down 45% and SK Hynix 55% from their June peaks at the trough. SpaceX priced at $135 in June, peaked near $211 within three sessions, traded below its IPO price by mid-July, then closed at $152.71 on 20 September. Index trackers were obliged to buy it on 7 July at around $160.
Ratepayers, taxpayers and communities. Those transfers are already on the books. US residential electricity is 18.34 cents per kilowatt-hour, up 5% year over year. Georgia's projected annual data-centre sales-tax loss reached nearly $1.9 billion. Five firms reported $422 billion of profit and paid an effective federal rate of 4.5%.
The clearest measure of how much sits unbooked is Meta's Louisiana venture. Its June 2026 10-Q reports maximum exposure to loss on that one venture of $46.03 billion, against an equity carrying value of $2.92 billion, with no liability recorded because the residual value guarantee is not judged probable. Nobody in the chain has yet booked a credit loss. Nvidia earned $59.7 billion in the quarter. The visible stress is in the leveraged neoclouds, where CoreWeave's credit default swaps have traded near 855 basis points and where Oracle, CoreWeave and Nebius carry about $173 billion of debt at roughly 7.8 times quarterly revenue against Microsoft's 0.4 times.
Where the loss is going
Oracle's leases commence between fiscal 2027 and fiscal 2029, on fifteen to nineteen year terms, at $288 billion. Oracle is rated BBB.
Nvidia's guarantee becomes effective in phases from fiscal 2029. If OpenAI defaults, the filing says Nvidia "may assume the applicable lease". It holds an option on 3.8 further gigawatts at the same site.
Commercial mortgages maturing. $875 billion matures in 2026 and $652 billion in 2027. Banks and thrifts hold $1.844 trillion of it. Of regional banks, 54.8% sit above the 300% concentration guidance, at a median of 312%.
Office leases expiring. More than 32% of US office leases expire between the second quarter of 2026 and the end of 2028. Until a lease expires the tenant continues paying. Any AI-driven reduction in space demand can therefore first appear only inside that window.
The depreciation question. Five firms extended server useful lives toward six years, each adding roughly 4 to 5% to net income. Amazon has already partially reversed, shortening a subset from six years to five and citing "the increased pace of technology development".
Does the math add up
It does not. The gap has not moved since June.
Saturation revenue at current per-user economics runs $340 to $500 billion a year. The infrastructure to serve it runs $6.5 to $26.5 trillion. Simple payback is 13 to 78 years against a chip life the modelling puts at three.
The demand side has not improved either. McKinsey has 78% of organisations using AI in at least one function, against 55% two years ago. Between 5 and 8% report a measurable at-scale financial return. PwC found 56% of chief executives reporting neither revenue growth nor cost reduction from AI in the previous twelve months.
The labs are compounding revenue faster than the clouds. The capital spending sits with the clouds.
Where the public sits
Already paying. Ratepayers through capacity charges, state treasuries through abatements, federal receipts through accelerated depreciation. None of this is contingent. It is happening monthly.
Holding the equity. A third of the S&P is seven companies. Passive holders cannot step aside. In 2000 the Nasdaq fell 78% while the S&P fell 50%, since the speculative names sat outside the benchmark. That separation no longer exists.
Standing behind the credit, without having agreed to. The FDIC's Deposit Insurance Fund sits at $157.4 billion against banks holding $1.844 trillion of commercial mortgage debt. Fannie Mae and Freddie Mac guarantee a combined $7.8 trillion of mortgages including multifamily. The federal government has run both since 2008 and stands behind them. State guaranty associations stand behind the life insurers that hold the private credit. And Section 3 of Executive Order 14318 already authorises federal loan guarantees. Section 5(c) keeps them outside environmental review as long as they stay under 50% of project cost.
None of these has been called on. The Nvidia guarantee has not been triggered. No insurer has failed. No bank has been resolved. Each instrument exists and none has been drawn.
The condition the structure depends on
The structure depends on one condition. AI revenue must keep compounding fast enough that no participant is forced to sell an asset.
A sale establishes a price. A price requires holders of comparable assets to revalue. A revaluation reduces capital. Each mechanism described above avoids a sale. Nvidia's exposure is disclosed in a note. Oracle's leases are disclosed in a note. Banks pledge commercial mortgages to the discount window and receive cash without selling. Significant risk transfer moves credit risk while the loan stays in place. A federal loan guarantee enters the budget at the government's estimate of what it will eventually cost. The amount guaranteed never appears.
No participant has yet been forced to sell. The assets stay carried at estimates until one is.