How much debt is funding the AI boom?

How much debt is really funding the AI boom right now?

We estimate that roughly $600 billion to $800 billion of funded debt currently supports the AI boom. That is an editorial range, not an official count. No regulator or bank publishes one de-duplicated total.

The firmest public number comes from the IMF. In its October 2026 Global Financial Stability Report, it traced about $650 billion of debt tied to 14 disclosed AI transactions since 2023, of which about $320 billion sits with the companies receiving the investment. Those 14 deals are only a slice of the market, but in our reading part of that debt sits with the investors, whose borrowing also funds other activities. That is why our range brackets the IMF figure rather than starting above it.

The flow of new borrowing is a separate measure. FTSE Russell counted $218.1 billion of index-eligible investment-grade bonds from the five largest US hyperscalers in 2026 through early October. Goldman Sachs counted nearly $500 billion of AI-related issuance by early August on a broader definition. Part of that refinances older debt or funds ordinary corporate needs, so neither figure can simply be added to the IMF total.

The biggest deals of late October 2026 do not move the range yet, because most are not funded debt. Microsoft’s $25 billion of notes was priced on 29 October and is scheduled to settle on 4 November. Amazon’s facility of up to about $80 billion is undrawn capacity. The $63 billion Anthropic chip vehicle is an arranged financing for equipment delivered from 2027, and SpaceX’s $40 billion lease is a commitment that still faces closing conditions.

The wider exposure is larger again. Moody’s says the five largest US hyperscalers have about $969 billion of future lease commitments, including roughly $662 billion for leases that have yet to begin. Those contracts create real pressure, but they stretch across many years and are not cash borrowed upfront.

Why did AI suddenly become a debt story?

AI became a debt story when infrastructure spending grew faster than even the richest technology companies wanted to fund from normal cash flow.

The spending numbers explain the shift. In May 2026, Goldman Sachs modeled about $765 billion of AI infrastructure spending this year across compute, data centers and power, rising to roughly $1.6 trillion by 2031, or $7.6 trillion between 2026 and 2031. In October, Reuters reported that Moody’s expects the five largest US hyperscalers and CoreWeave to spend nearly $1 trillion on capital expenditure in 2027. Both are forecasts, not debt totals. Spending on that scale needs financing; we look at where the biggest rounds are going in our report on AI infrastructure.

The assets also have to be bought before customers produce the revenue that pays for them. A company must secure land, power and accelerators years before demand is fully visible. Lending now reaches that early layer too: in October 2026, Crusoe announced a $700 million facility from Neuberger Specialty Finance secured by power-generation equipment it had just ordered.

Debt lets companies build several waves of capacity at once. Goldman Sachs estimates that debt financed about 27% of hyperscaler capital spending in 2025 and roughly one third in 2026. The marginal AI campus is now financed through bonds, equipment loans, project vehicles and leases even when the company behind it is profitable.

What should we actually count as AI debt?

We should count borrowed cash tied to chips, data centers, cloud capacity and supporting infrastructure, while keeping leases, undrawn facilities, equity, guarantees and customer commitments in separate categories.

Corporate bonds count when their proceeds help finance AI expansion, even if the legal documents say “general corporate purposes,” as Microsoft’s October notes do. Bank loans and private-credit facilities count when they fund servers, GPUs, buildings or power connections, but only once drawn. Amazon’s new facility of up to $80 billion cannot be drawn before the first quarter of 2027, so it is capacity, not debt.

Leases need different treatment. A company may promise $20 billion of rent over 15 years, but the landlord does not hand it $20 billion on day one. The total includes financing costs, the owner’s return and payments for operating the site. We track that obligation because it can become painful, yet we keep it outside the funded-debt estimate.

Equity and guarantees are different again. Meta’s plan to raise up to $97.5 billion through common shares, mandatory convertible preferred shares and a future at-the-market program, as reported by Reuters, is equity, not borrowing. Broadcom’s residual-value support for the Anthropic chip vehicle is a contingent promise on the same loans, not new lending. Oracle’s $664 billion of remaining performance obligations, reported in September 2026, is a backlog, and the IMF’s $1.4 trillion of revenue commitments also stays out of any debt total.

Financing item	How we count it	Why
Corporate bonds and term loans	Funded debt	Cash is raised now and must be repaid
Equipment and GPU financing	Funded debt	The hardware directly supports AI capacity
Data-center project debt	Funded debt	A project vehicle borrows against the campus and tenant contract
Undrawn facilities and unclosed commitments	Financing capacity, not yet debt	Money becomes debt only when drawn or when the deal closes
Future lease payments	Debt-like exposure	Payments are fixed, but spread across many years
Guarantees and residual-value support	Contingent exposure	They pay only if the borrower or collateral falls short, on loans already counted
Customer prepayments and backlogs	Financing or orders, not debt	Customers provide cash or promise purchases without creating a repayment claim
Announced capital spending	Planned investment	A budget does not show how the spending will be financed

Does all this AI debt issuance mean as many new data centers?

No. The AI issuance headlines capture a huge borrowing wave, but they overstate how much brand-new infrastructure has been financed.

In July 2026, Goldman Sachs put AI-related debt issuance for the year at about $489 billion, already above the $322 billion it counted for 2025. By early August, Goldman described nearly $500 billion issued so far in 2026, with hyperscalers accounting for about $194 billion. The Financial Times reported on 30 September that Goldman expects more than $750 billion by year-end. A forecast is not money already raised, so we follow how much really went in, quarter by quarter, in our AI infrastructure funding report.

Definitions change the answer. FTSE Russell, counting only fixed-rate investment-grade bonds from five hyperscalers, reached $218.1 billion by early October. BNY cited an S&P forecast of as much as $400 billion of AI-related credit issuance in 2026. That sits well below Goldman’s projection because the two count different issuers and instruments.

Part of that money replaces older debt or funds ordinary corporate spending. There is another trap: one project can appear several times as it moves through the financial system. A private lender may fund construction, a bank may lend to that private-credit vehicle, and the finished campus may later be refinanced with bonds.

Physical delivery also lags the money. Oracle reported 850 megawatts of new data-center capacity delivered in the quarter to August 2026, a real but small fraction of what its order book implies. Microsoft said in October that some third-party capacity deliveries had shifted into the second half of its fiscal year. We use the issuance figure as a measure of financing activity, not of servers and buildings actually in place.

What does the IMF's $650 billion AI debt figure actually cover?

It covers debt linked to 14 publicly disclosed AI transactions since 2023, not every AI loan in the economy. It is still the most systematic public tracing of AI debt available today.

The IMF found about $300 billion of equity, $650 billion of associated debt and $1.4 trillion of future revenue commitments across those deals. About $320 billion of the debt sits with the investee firms themselves, partly supported by investor guarantees or held off balance sheet. The pattern the IMF describes is circular: a large company invests in an AI firm, which then commits to buy that company’s chips or cloud services.

Other layers reach beyond those 14 deals. The BIS counts more than $200 billion of outstanding private-credit loans to companies it classifies as AI, big data or cloud technology. Oracle raised $43 billion of debt in the fiscal year to May 2026. CoreWeave’s balance sheet showed about $35 billion of debt at the end of June 2026, while Reuters, citing LSEG data, put the debt behind its buildout at about $50 billion in mid-October; the two figures cover different dates and definitions. Alphabet reported $125.1 billion of long-term debt at the end of September 2026, a corporate balance rather than AI-only borrowing.

We cannot simply add those numbers. CoreWeave and Oracle may appear in the IMF deals, in lending datasets and in bond indexes at once. After allowing for overlap, a total below $600 billion no longer fits what is visible. Going above $800 billion would require counting a broad share of general corporate borrowing, future leases or facilities that have not been drawn.

Observable financing layer	Current evidence	Our treatment
IMF-traced circular deals	About $650B of debt across 14 deals since 2023	Core anchor, but a subset of the market
AI-related private credit	More than $200B outstanding, per the BIS	Partly overlaps with the IMF deals
Large-tech corporate issuance	$218.1B of hyperscaler investment-grade bonds by early October 2026 (FTSE Russell); nearly $500B of wider AI-related issuance by August (Goldman Sachs)	Adjusted for refinancing and non-AI uses
CoreWeave debt	About $35B at June 2026; about $50B per Reuters in October	Strong direct link, with overlap across instruments
Late-October 2026 pipeline	$25B Microsoft notes priced; up to $80B Amazon facility; $63B Anthropic chip vehicle; $40B SpaceX lease commitment	Outside the estimate until settled, drawn or closed
Our funded AI-debt estimate	$600B–$800B	Editorial range after overlap adjustments

Are Big Tech companies borrowing because AI now eats their cash flow?

Increasingly, yes. None of them is running out of cash, but third-quarter 2026 results show AI spending now exceeds the cash some of them generate, so outside funding has moved from optional to routine.

Alphabet spent $51.4 billion on capital expenditure in the third quarter and posted negative free cash flow of $9.1 billion, with its 2026 capex guidance kept at $195 billion to $205 billion. Amazon spent $75.9 billion of cash capex in the quarter, and its free cash flow over the trailing twelve months was negative $42.4 billion. Microsoft reported $45.2 billion of capex including finance leases, or $39.4 billion in cash, and priced $25 billion of notes on 29 October for general corporate purposes.

Meta shows both sides. In the second quarter of 2026, it generated $31.9 billion of operating cash flow and kept just $784 million of free cash flow. Its third-quarter capex, including finance-lease principal, reached $34.9 billion, and it narrowed 2026 guidance to $135 billion to $145 billion. In late October it went to equity: $12.5 billion of mandatory convertible preferred shares closed on 29 October, and about $25 billion of common shares were sold under forward agreements that pay Meta only when settled, by November 2027 at the latest. Reuters put the whole plan, including a proposed $60 billion at-the-market program, at $97.5 billion.

Oracle is in a tighter position. In its quarter to August 2026, it spent $28.5 billion on capital expenditure and posted negative free cash flow of $5.4 billion, even with $11.4 billion of customer prepayments counted in operating cash. It also sold $20 billion of new shares. At its October 2026 Investor Day, it capped further capital raising through fiscal 2030 at $20 billion, split equally between equity and gross debt and excluding refinancing, and targeted positive free cash flow in the second half of fiscal 2028.

Strong hyperscalers borrow for speed and to keep cash in reserve. Stretched providers borrow because they cannot execute the plan otherwise. Both sit in the same AI-debt totals, with very different credit risk.

Which AI companies actually depend on debt to keep expanding?

CoreWeave and Oracle rely much more heavily on outside financing than Microsoft, Alphabet, Amazon or Meta, and CoreWeave remains the clearest case.

In the second quarter of 2026, CoreWeave spent $6.4 billion on property and equipment while generating $679 million of operating cash flow. It raised $13.5 billion of new debt in that quarter alone and repaid $3.9 billion. Net interest expense reached $640 million, about a quarter of its $2.58 billion revenue, and it posted a $626 million net loss. Its third-quarter results are scheduled for 10 November 2026.

The borrowing continued in the third quarter. In August 2026, CoreWeave arranged a $2.6 billion delayed-draw loan led by JPMorgan and MUFG. In September, it priced $3.7 billion of convertible notes due 2033 at a 2.875% coupon. Its revenue backlog of about $104 billion at the end of June explains why lenders keep funding it: the contracts are real, but they mostly sit with a few very large customers. Contracts that mostly sit with a few very large customers are one of the open problems we dig into in our AI infrastructure funding report.

The market charges for that risk. AIStockWire, working from bond-fund holdings, estimated the median yield on tracked CoreWeave bonds at 14.04% on 29 October, about 9.3 percentage points over Treasuries. Oracle has a mature software business, but its AI cloud plan is larger than that business can fund, and its new cap on capital raising shows management wants to protect its credit rating. Microsoft keeps the strongest internal funding position, while Alphabet, Amazon and Meta now combine huge cash generation with growing outside financing.

Company or model	Current debt dependence	What the evidence says
Microsoft and Alphabet	Low but rising	Strong balance sheets; Alphabet’s free cash flow was negative $9.1B in Q3 2026, and Microsoft priced $25B of notes in October
Amazon	Moderate and rising	Trailing free cash flow of negative $42.4B; secured supplier facility of up to $80B drawable from 2027
Meta	Moderate and rising	Free cash flow near zero in Q2 2026, $25B of bonds, project vehicles and a $97.5B equity plan
Oracle	High for incremental AI capacity	Debt, $20B of new equity and customer prepayments fill the gap; further raises capped at $20B through fiscal 2030
CoreWeave and similar neoclouds	Very high	Borrowing is built into the GPU acquisition model
Independent data-center developers	Very high	Projects depend on loans, tenant contracts and refinancing

Are giant AI data-center leases basically debt?

Giant AI data-center leases behave a lot like debt when the customer has little practical freedom to walk away.

Moody’s estimated in early 2026 that the five largest US hyperscalers had about $969 billion of future lease commitments. Roughly $662 billion related to facilities whose leases had yet to start. Alphabet’s own filing shows the scale: at the end of September 2026, it had $130.5 billion of future payments on leases not yet commenced, mainly data centers and energy infrastructure.

The accounting label hides much of the economic relationship. A project vehicle borrows to build a data center or buy chips, while the technology company signs a long contract to use them. Lenders care about the tenant’s payments far more than the thinly capitalized vehicle that formally owes the debt. The IMF makes the same point: part of the $650 billion it traced is held off balance sheet or backed by investor guarantees.

The newest deals extend that model from buildings to chips. The $63 billion vehicle announced in October 2026 buys Broadcom accelerators and networking gear and leases them to Anthropic, with deliveries expected from 2027. SpaceX’s $40 billion commitment is a finance lease from an Apollo-managed entity, with lease schedules starting when GPUs arrive in the first half of 2027. Counting both the lessor’s borrowing and the tenant’s payments as separate loans would double-count the same chain.

Meta’s Hyperion arrangement shows the building version. Blue Owl funds own 80% of the $27 billion venture, Meta owns 20%, and Meta will lease all the facilities, with a capped residual-value guarantee if the campus is worth less than expected when leases end. Meta even counts finance-lease principal payments inside its capex guidance, which shows how thin the line between leasing and spending has become.

We keep lease commitments outside our funded-debt estimate because they are an undiscounted stream of future rent. Still, ignoring them would understate how much of the AI buildout depends on fixed payments from a small number of technology companies.

Who is lending all this money besides bond investors?

Private-credit funds, insurers, asset managers and structured-finance investors now provide a major share of AI infrastructure financing, with banks behind many of them and chipmakers increasingly standing behind the loans.

The BIS calculated in January 2026 that outstanding private-credit loans to AI-related companies had climbed from almost nothing to more than $200 billion, close to 8% of all private-credit lending. In September 2026, it widened the lens: private-credit lending to software and technology firms passed $1 trillion in 2025, about 44% of a US market of almost $2.5 trillion.

The average AI-related private-credit loan is large at about $169 million, versus $90 million for other borrowers. Yet maturity and pricing look almost identical: 4.7 years and a 6.2 percentage-point spread, against 4.8 years and 6.1 points elsewhere.

The October 2026 deals show who now writes the biggest checks. Bank of America, Citi and Morgan Stanley arrange the first-lien debt in the Anthropic chip vehicle, about $43.5 billion according to Reuters, while Blackstone provides committed second-lien debt. Apollo and Goldman Sachs lead the SpaceX lease, with BlackRock, Goldman Sachs Asset Management and KKR-advised accounts as anchors. Smaller operators borrow too: TeraWulf raised its revolving credit line from $250 million to about $400 million.

Suppliers have become a quieter source of support. Broadcom provides residual-value support to the Anthropic vehicle, meaning it absorbs losses if the equipment is worth less than expected. Reuters, citing The Information, reported that Nvidia was in talks to provide credit support for Nscale’s Anthropic project, though no deal had been signed. Banks remain involved even when they disappear from the headline, through funding lines, guarantees or a difficult refinancing.

Is AI debt starting to look like the telecom bubble?

AI financing now resembles the late-1990s telecom buildout more than a normal technology investment cycle, and the IMF’s circular-deal warning makes the comparison harder to dismiss.

The pattern is familiar: companies rush to build capacity before rivals, lenders finance long-lived assets, and technology keeps changing the best equipment. Goldman Sachs has estimated that hyperscaler spending would reach the telecom boom’s peak intensity at around $700 billion in one year. Goldman’s own estimate of about $765 billion of AI infrastructure spending in 2026 is already at that scale, though it covers more than the hyperscalers.

Telecom also had vendor financing, where equipment makers lent to the customers buying their gear. The IMF’s 14 deals echo that: investors put money into AI firms that then commit $1.4 trillion of purchases back to them. Broadcom’s residual-value support and Nvidia’s reported guarantee talks add a new version, where the chip seller carries part of the risk on the chips it sells. Investors funding customers who then buy from them make a circular model; we compare which business models actually win in our AI infrastructure report.

AI demand could keep rising even if the financing disappoints, just as internet traffic kept growing after telecom investors lost money. Google Cloud revenue rose 103% and Azure 48% in the third quarter of 2026, though both include more than AI. The technology can transform the economy while specific lenders, landlords and cloud providers earn poor returns.

Today’s boom has stronger borrowers than telecom had. The fragile edge sits with projects that depend on one customer, one generation of hardware or repeated refinancing.

What would make AI debt go bad?

AI debt turns dangerous when compute prices, utilization or power connections fall short before the loans, leases and guarantees have run off.

The hardware can age much faster than the debt. A data-center shell may last 20 years, while a GPU can lose commercial value within a few product cycles. CoreWeave’s delayed-draw facilities are secured partly by GPU servers and customer cash flows, and the new chip vehicles lean on residual values. If newer systems deliver much more output per dollar, older collateral can lose value faster than lenders expected, and supplier guarantees could be called.

Power is now a visible risk too. Axios reported on 20 October 2026 that Goldman Sachs counts 86 data-center projects canceled or significantly delayed since January, 76 of them linked to local opposition. A campus that is financed but cannot connect to the grid pays interest without earning rent.

A single customer can also carry most of a project’s economics. If it delays deployment or runs into its own financing problem, the shock travels quickly through the structure. The IMF’s point is that in circular deals, the lender’s revenue and the borrower’s health are tied together.

Refinancing arrives long before many campuses wear out. The BIS finds an average maturity of 4.7 years for private-credit loans to AI-related companies, and Moody’s warned in October 2026 of a wave of private-credit loans coming due between 2027 and 2029. Overbuilding could do the most damage: lower compute prices help AI users but hurt anyone who borrowed on the assumption that today’s rental rates would last.

Could AI keep booming while lenders lose money?

Yes. AI usage can keep growing rapidly while lenders and data-center investors suffer large losses.

We saw this with the internet. Traffic and online services kept expanding after the dot-com crash, while many telecom networks and their investors failed.

The same split can happen with AI. Better chips and more efficient models may increase useful work while reducing the price customers pay for each unit of compute. Enterprises could adopt AI faster precisely because infrastructure owners are forced to cut prices. BNY noted in October 2026 that excess issuance alone could push credit spreads wider, even if the underlying businesses do not deteriorate.

Debt makes that gap harsher. An equity-funded company can wait for demand to catch up. A leveraged data center must meet interest and principal schedules. CoreWeave more than doubled its revenue in a year, to $2.58 billion in the second quarter of 2026, and still lost $626 million, with $640 million of net interest. Growth alone does not cover the financing bill.

We should judge the technology and the financing separately. Strong AI adoption would support the long-term need for compute, but it would not rescue every campus built in the wrong place, every old GPU fleet or every contract priced too aggressively.

Could AI debt threaten the wider financial system?

AI debt is creating real pockets of stress, and official warnings have grown louder, but the wider financial system remains short of crisis.

The risk is still concentrated. The BIS says around 20% of private-credit funds have made at least one AI-related loan, but the average fund has only about 5% of its lending in the category. J.P. Morgan said in October 2026 that most global banks have limited direct exposure to CoreWeave. One in five private-credit funds has lent to AI; we track which investors are making the biggest bets in our AI infrastructure report.

The links are getting harder to trace. In October 2026, the IMF warned that its 14 circular deals could amplify a market correction, because one company’s failure would hit its investors’ equity stakes and their future sales together. The IMF also notes that hyperscalers now issue about 10% of global investment-grade bonds, up from less than 1% before the boom. Moody’s added that AI is now a driver of private-credit growth as refinancing risks rise.

Credit markets are no longer completely relaxed. In August 2026, Goldman Sachs said spreads on its basket of AI-leader bonds had widened to nearly twice their tight of 74 basis points over the previous year, and that 17 of 23 data-center joint-venture high-yield deals traded wider than at issue. The Financial Times reported in October that trading in credit-default swaps on exposed companies reached about $9 billion a week in September, against less than $2 billion in July. That is hedging activity, not losses.

Our judgment: pricing is still too generous for the riskiest borrowers, especially neoclouds and single-tenant projects, while the strongest hyperscalers keep the wider system away from crisis territory for now.

So how much debt is funding the AI boom?

The best current answer is roughly $600 billion to $800 billion of funded AI debt, an editorial estimate, plus close to $1 trillion of future lease commitments and a fast-growing pipeline of deals not yet funded.

As seen above, the IMF traced about $650 billion of debt behind just 14 AI deals. That figure sits inside our range: it covers only a slice of the market, but in our reading part of it is investors’ general borrowing. Issuance keeps climbing, with $218.1 billion of hyperscaler investment-grade bonds by early October and Goldman Sachs expecting more than $750 billion of wider AI-related issuance in 2026. Refinancing, general corporate use and repeated financing of the same assets stop us from adding these numbers together.

Our range also rests on more than $200 billion of AI-related private credit, Oracle’s $43 billion debt raise, CoreWeave’s debt stack of $35 billion to $50 billion depending on the measure, and multibillion-dollar project vehicles such as Hyperion. The late-October wave, from Microsoft’s $25 billion of notes to the $63 billion Anthropic chip vehicle, Amazon’s $80 billion facility and SpaceX’s $40 billion lease, will raise the total only as it settles, closes and is drawn.

The five largest US hyperscalers carry about $969 billion of future lease commitments, including around $662 billion for leases not yet started. Those contracts sit on top of funded debt, even though they are not loans received today and are not measured on the same basis.

Debt remains a choice for Microsoft, a growing habit for Alphabet, Amazon and Meta, and a necessity for Oracle, CoreWeave and independent developers. Meta and Oracle now lean on equity too, and chipmakers stand behind some loans. The AI boom still runs on cash, prepayments and equity, but its fastest-growing edge is unmistakably leveraged, and lenders have started to price that.