For most of the AI boom, the constraints were chips, data centers and electricity. Capital was assumed to be available. This week’s news makes that harder to assume: Oracle sent a force-majeure notice over its giant New Mexico data center, Nvidia was reported to be talking to insurers about backstopping loans against its own chips, and Anthropic’s confidential IPO filing revealed a $518 billion infrastructure bill. Separately, a Columbia economist put a number on how large the whole buildout could get. And on September 30, Reuters and the Bank of England both pointed to the same pattern lower down the credit ladder: investors are charging more for AI debt.

None of this says the boom is ending. It says the price of financing it is now visible, and that price is starting to move.

The scale

Gartner forecasts worldwide AI spending of $2.67 trillion in 2026, up 49.5% from last year, with $1.48 trillion of it going to infrastructure. The IMF’s April estimate puts world GDP at about $126.3 trillion, so AI spending equals roughly 2.1% of global output and infrastructure alone about 1.2%.

Two caveats. The Gartner figure is a forecast, and it covers more than data centers, including software, services and devices. Spending is also not the same as investment. Still, this is no longer a niche technology boom financed from Silicon Valley cash flow. It is one of the largest destinations for capital in the world.

For the U.S. alone, Columbia’s Stijn Van Nieuwerburgh estimates in a new Brookings paper that data centers and related infrastructure will absorb $10.3 trillion between 2025 and 2032, an average of about 3.6% of GDP per year. The railroad boom, the previous record, absorbed about 2.2%.

Two projects, two warning lights

Project Jupiter. The Oracle-tenanted campus in New Mexico is financed by roughly $18 billion of bank loans from about 20 lenders. The arranging banks have struggled to sell the loans on, and they have been quoted at 89 to 91 cents on the dollar, according to the Financial Times.

The trigger was power. On September 24 Oracle sent a force-majeure notice to a Blue Owl unit over delays tied to the site’s electricity, according to Bloomberg. Bloomberg reported that if both sides agree such an event occurred, Oracle could defer rent for up to three years. That is a contract dispute, not a loan default, and Oracle says it is “fully committed to New Mexico.” But a modern AI campus draws power on the scale of a small city, and generation, transmission and permitting cannot expand as fast as servers can be ordered. Delay means interest keeps accruing while revenue is pushed out, and lenders want more compensation.

Hyperion. Meta’s Louisiana campus is financed partly by $27 billion of 6.581% senior secured notes due 2049, issued through a vehicle called Beignet Investor. S&P rated them A+, largely because Meta guarantees the project’s residual value. Meta owns only a minority stake, so the debt sits off its balance sheet. On Monday the bonds traded around 91 cents on the dollar, yielding as much as 7.55%, about 230 basis points over Treasuries. That is roughly a full percentage point above the coupon at issue. It is also below the record spread of about 255 basis points reached in July, so the pressure has come and gone before. No cause for panic, but not nothing.

Meta is in no difficulty. Oracle is not in default, but its credit-default-swap gauge hit a record high on September 24, according to Bloomberg, so it is under real market stress. The point is that debt on two enormous projects, backed by very strong corporate tenants, now trades at levels that ask investors to be paid more for the risk.

Those prices become the reference for the next deal. A sponsor that hoped to borrow at 6% may face 8% or 9%, or be asked for more equity, stronger guarantees or tighter covenants. On a $27 billion structure, each percentage point is about $270 million a year. The best projects will still be built. The more speculative ones get harder to justify.

The chips as collateral

The same pressure is showing up in hardware lending. The Financial Times reported that Nvidia has held early-stage talks with insurers, working with the broker Howden Re, on cover for lenders to smaller “neocloud” providers. If a borrower defaults and the pledged chips resell for less than the debt, the insurer would cover the shortfall. Nvidia has reportedly given at least one insurer data on chip depreciation and the expected future value of computing power. The talks may not produce any deal.

CoreWeave, the best-known neocloud, pioneered borrowing against Nvidia GPUs in 2023. The uncomfortable question for lenders is what those chips are worth if a borrower fails. That depends on how fast Nvidia ships faster chips, how deep the used market is, what it costs to move and reinstall the machines, and whether other distressed borrowers are selling the same hardware at once. Much of this debt is investment grade because of the customer contract behind it, not the neocloud’s own credit or the chips.

Lending does not have to stop for conditions to tighten. A bank can lend 60% of a purchase instead of 80%, charge more, or ask for prepayments, guarantees or insurance. The project still gets financed, but someone has to supply the extra money.

The repricing is spreading down the credit ladder

Jupiter can be blamed on power and permitting, and Hyperion on the sheer size of its financing vehicle. Evidence from below investment grade is harder to explain that way.

Reuters reported on September 30 that low-rated AI borrowers have raised about $88 billion in U.S. leveraged finance this year, according to Goldman Sachs, against roughly $20 billion in the first eleven months of 2025, per Neuberger Berman. The periods are not identical, but the direction is clear. The prices are the more telling part. SoftBank, rated BB+, paid 8.625% on 3.5-year notes, 9.25% on 5.5-year debt and up to 9.75% on 7.5-year bonds, and weaker credits face borrowing costs of up to 14% to 15%. Morningstar’s Elizabeth Templeton said managers of collateralized loan obligations, the main buyers of leveraged loans, are becoming more cautious on certain names.

A third project financing shows the same thing. Zenith Arc, a vehicle financing an Oklahoma data center leased to Jane Street, sold $2.25 billion of five-year senior secured green bonds in August at 99.50 with an 8.875% coupon. The bonds dropped almost three points to a 96.75 bid right after issue and, according to Pender Fund Management, had lost more than seven points from the issue price by late August. That last figure comes from a single fund manager, and these bonds trade thinly, so treat it as indicative. It still matters because Zenith Arc is unrelated to Oracle or Meta.

There is counter-evidence, and it should not be skipped. BNP Paribas data show $40 billion of high-yield AI infrastructure supply so far this year against $12 billion in all of 2025, so lenders are still buying. But BNP’s Erin Brown also described overall high-yield volumes as “basically flat” year on year, which means AI is taking a larger share of a market that is not growing, not that the market is wide open. Nscale’s $3.36 billion pre-IPO raise, led by Third Point, is often cited as proof that money is available. It is a convertible financing, so it shows that equity-linked capital is available, not that credit is cheap.

Regulators are now paying attention. The Bank of England said on September 30 that Morgan Stanley estimated global AI-related debt issuance at about $450 billion by early September, double the 2025 total. Its Financial Policy Committee said high valuations rest on strong earnings expectations, and that a significant shock to those expectations could trigger a sharper repricing. Quartz’s summary adds that the committee sees a possible selloff steeper than July’s, with spillovers to sovereign debt.

What follows is inference, not reported fact. Capital is still available, but its marginal price is rising. A weaker borrower paying 12% to 15% cannot build on the economics of a hyperscaler borrowing near investment-grade rates. More projects will need extra equity, stronger customer contracts, guarantees or lower leverage before lenders participate. That is how financing conditions can ration a buildout without anything resembling a credit freeze.

Does this crowd out everyone else?

Goldman Sachs estimates, as cited by the Financial Times, that investors have provided about $500 billion to AI-linked borrowers this year, roughly $200 billion of it to Amazon, Alphabet, Meta, Microsoft and Oracle. Those hyperscalers are expected to issue more than $1 trillion of new debt over the next few years. The FT reports the volume is large enough that other companies, and even governments, now time their bond sales around hyperscaler deals.

Governments are borrowing in the same market. The Institute of International Finance puts global debt at a record $365.5 trillion in the first half of 2026. That figure covers governments, households, financial firms and companies, not sovereign debt alone. The IIF also noted that advanced economies paid over $3.3 trillion in interest on internationally traded government bonds last year, more than the world spends on AI.

The crowding-out argument is plausible but not proven. Morgan Stanley’s head of U.S. credit strategy told an IIF briefing that AI-linked issuance is not the big variable for interest rates, in part because hyperscaler bond buyers and Treasury buyers tend to be different investors. The IIF’s own data show debt growth slowing, to $10 trillion in the half-year against $21 trillion a year earlier, as financing costs bite. Capital markets can expand; the question is at what price, and who pays it. The cleanest evidence today is in AI’s own paper, not in Treasury yields.

The gap between spending and returns

If AI delivers the productivity gains now assumed, the buildout pays for itself. Companies become more productive, new industries emerge, tax revenue grows and cash flows service the debt. The danger is the gap between when money is spent and when returns arrive.

Anthropic’s confidential IPO prospectus, seen by Reuters, shows the mismatch clearly. It commits the company to at least $518 billion of infrastructure spending with six partners over the next decade, and roughly 80% is non-cancelable or owed regardless of usage, according to Reuters. Against that, the filing shows $20.3 billion of cash at the end of 2025, so the commitments are about 25 times the cash on hand. It also shows an operating loss of about $8.1 billion for 2025 and a GAAP net loss of about $42 billion, roughly $34 billion of which was a non-cash financing and fair-value charge.

Two pieces of context cut the other way. The commitments are spread over seven to ten years, which averages roughly $52 billion a year if evenly spread over ten, and Reuters reports an annualized revenue run rate above $65 billion by the end of July. But a run rate is not profit, and Anthropic’s case to investors, that compute will be the binding constraint on AI, is a forecast that demand will keep growing fast enough to cover costs it can’t cancel. Skeptics are loud: the investor Michael Burry wrote on X on September 29 that markets should “tank hard” and prevent the Anthropic and OpenAI IPOs.

Van Nieuwerburgh’s paper frames the same problem. To earn a 10% return, the U.S. buildout would need about $3.7 trillion of annual AI revenue by 2032. Starting from roughly $100 billion at OpenAI and Anthropic combined, that implies growth near 80% a year. He also cautions that it is premature to call the buildout a systemic threat: strong demand, high utilization and better models could support it.

Two ways this ends

The benign outcome is a slowdown. Returns disappoint, lenders demand more protection, weaker projects are cancelled, and capital drifts elsewhere. The rest of the economy gets some relief in credit, equipment and electricity.

A disorderly reversal would look different. Projects financed on high-utilization assumptions get marked down. GPU collateral loses value. Banks and private-credit funds find that seemingly separate loans share the same demand assumptions. Special-purpose vehicles ask sponsors for more capital, and guarantors have to honor their commitments. Orders for chips, electrical equipment and construction are cancelled before the data centers themselves default. Governments lose tax revenue just as markets absorb credit losses, and central banks may find easing constrained if the boom also lifted electricity prices, construction costs and inflation.

That second path is a scenario, not a forecast.

What to watch

The risk is not that AI disappears. It is that the financial system builds for a level of demand that arrives late, or not at all. Jupiter’s loan quotes, Hyperion’s spread, Zenith Arc’s slide below par, hyperscaler order books and Nvidia’s attempts to insure its collateral are early prices on risks that barely existed a few years ago.

So far this is credit repricing, not a broken buildout. There is no broad sign yet of cancelled orders for GPUs, servers, networking, cooling or electrical equipment, and no evidence of widespread financing failures. The signal that would change that reading is a higher cost of capital starting to change behavior: projects downsized or postponed for financing reasons, falling loan-to-value ratios on GPU loans, mandatory residual-value insurance, larger sponsor equity requirements, or credible equipment-order cancellations. If those appear while Jupiter, Hyperion and Zenith Arc remain under pressure, the story moves from financial markets to the real economy.

For years the question was whether the industry could get enough chips and electricity. The question now is whether the world can finance all of it at a price that still makes the economics work.

Sources