AI stocks and bonds took a hit this week as the cost of money jumped. Private credit funds are some of the biggest lenders behind the data-centre build-out. Here is how the money flows, where the risk sits, and why the tax code helped make these deals so attractive in the first place.

The quick version: Building AI data centres costs trillions, and a big slice of that money is borrowed, much of it from private credit funds rather than banks. This week, hotter US inflation pushed borrowing costs up and knocked AI stocks and bonds. Nothing has broken yet. But investors are now asking how much of the AI boom runs on debt, who is on the hook if the revenue disappoints, and how much of the deal-making was shaped by tax rules that reward heavy investment and borrowing.

What happened this week

The trigger was boring: an inflation number. US consumer prices rose more than expected in September, and markets started to worry that interest rates would stay high for longer. AI stocks, which depend on cheap money and big promises about the future, took the hit first.

According to Reuters, on October 15 South Korea's KOSPI index fell about 8 per cent and Japan's Nikkei fell 4.8 per cent. Samsung Electronics and SoftBank Group each lost more than 9 per cent, and SK Hynix and Kioxia each fell more than 11 per cent. The day before, the S&P 500 had fallen 1.1 per cent and the Nasdaq 1.9 per cent. CoreWeave, a cloud company that rents out AI chips, lost about 15 per cent in that session.

The more interesting moves were in credit. A credit default swap is basically insurance on a company's debt. When it gets more expensive, lenders are more nervous about getting paid back. On October 15 the five-year swap on Oracle widened to a record 288 basis points, according to S&P Global Market Intelligence data cited by Reuters. CoreWeave's hit a record 1,194.5 basis points during the day, according to ICE Data Services, a level Bloomberg said implies a more than 60 per cent chance of default within five years. Its bonds due 2032 traded at about 77 cents on the dollar. Swaps on Amazon, Alphabet, Meta and Nvidia also reached record wide levels, though they stayed far below Oracle's.

Nvidia's chief executive, Jensen Huang, pushed back, telling Bloomberg he saw no need for Nvidia to backstop CoreWeave. "Their cost of capital is actually decreasing," he said.

Two days earlier, Bloomberg reported that JPMorgan and Goldman Sachs had taken the rare step of limiting their exposure to a loan for an AI infrastructure start-up, and that a planned listing by a Nvidia-backed data-centre operator had collapsed. Broadcom's chief executive, Hock Tan, likened the strategies of some investors drawn to the AI boom to gambling. "The AI machine is still running at full speed, but its fuel — capital — is getting increasingly expensive," Hebe Chen of Vantage Global Prime told Bloomberg.

Not everyone sees a crisis. Charu Chanana of Saxo Bank said the selloff in Asia looked less like a breakdown in the technology and more like "a repricing of the cost of capital". JPMorgan analysts said "financing risks and IPO timelines" were still at the centre of the story.

That is the key point for anyone in private markets. This is a story about financing, and private credit is in the middle of it.

Why AI needs so much debt

Training and running AI models takes enormous data centres full of chips, plus the power and cooling to keep them running. For a long time, the biggest tech companies paid for this out of their own cash. That is changing.

In an August 2025 analysis, Morgan Stanley estimated that spending on data centres and the hardware inside them would reach about $2.9 trillion over four years. It reckoned about $1.5 trillion of that would need outside financing, and that roughly $800 billion could come from private credit, mostly through what is called asset-based finance: loans secured on the buildings, equipment and long-term leases. "That's larger than the size of the high yield market and the leveraged loan market," Morgan Stanley's Vishwas Patkar said.

Private credit fits this job well. Funds managed by firms like Blue Owl, Apollo and Blackstone can write very large cheques, tailor the terms, and hold the loans for years. Insurers and pension funds like the steady, long-dated income. Banks, which face stricter capital rules, often prefer to arrange the deal and sell the risk on.

Who could fund the $1.5tn AI data-centre gap

Morgan Stanley estimate of outside financing for data centres over four years, in billions of dollars. Hover or tap a bar.

Source of financingEstimate
Private credit$800bn
Investment-grade bonds$200bn
Data-centre ABS & CMBS$150bn
Other (sovereign, PE, VC)$350bn
Source: Morgan Stanley, August 2025.

The deal that set the template

The clearest example is Meta's Hyperion data-centre campus in Louisiana, agreed in October 2025 and described at the time as the biggest private credit deal ever.

Here is how it worked. Meta and funds managed by Blue Owl formed a joint venture to build the campus, with about $27 billion of development costs. Blue Owl's funds own 80 per cent and Meta owns 20 per cent. Blue Owl put in about $7 billion of cash, and part of its share was funded with bonds sold privately to Pimco and other investors. Meta received a one-time payout of about $3 billion from the venture. The bonds were rated A+ by S&P but yielded 6.58 per cent, a level more often seen on riskier debt.

Meta then agreed to lease the finished campus from the joint venture, with an initial term of four years and options to extend. It also gave the venture a "residual value guarantee" for the first 16 years: if Meta ends the lease and the campus is worth less than a set minimum, Meta makes up the shortfall.

Why build it this way? According to Bloomberg, the structure lets Meta fund the project off its own balance sheet. Meta gets a giant data centre without adding $27 billion of debt to its books. Investors get a high-yielding bond backed by one of the world's biggest companies. And the deal turns a tech project into a real estate and infrastructure investment, which is exactly the kind of asset private credit and infrastructure funds love to own.

The appetite for this kind of lending is now being tested. Bloomberg reported on October 15 that Stack Infrastructure, a data-centre developer owned by Blue Owl, is seeking an A$8.6 billion (about $6 billion) five-year loan for a data-centre project in Melbourne, in the middle of the selloff. Bloomberg also cited a Goldman Sachs estimate that AI-related debt issuance worldwide had reached about $578 billion this year by September.

Where tax fits into all this

The financing structures are not just about balance sheets. Tax rules shape who owns what, who gets the deductions, and why investors accept the yields they do. Here are the four pieces worth knowing.

1. The depreciation prize. The One Big Beautiful Bill Act (OBBBA), signed in July 2025, made 100 per cent bonus depreciation permanent for qualifying property acquired after January 19, 2025, according to a summary by the accounting firm RSM. In plain English, a business can write off the full cost of qualifying equipment in the year it buys it, instead of spreading the deduction over many years. Data centres are packed with expensive equipment, so these deductions can be enormous. In a joint venture structure, the entity that owns the assets for tax purposes is generally the one that claims them, which is part of why who owns the building matters so much. The tax-paying owner gets a big early write-off, while the tenant deducts its lease payments over time.

2. Interest deductions on the borrowing. Debt-funded projects depend on being able to deduct interest. The Section 163(j) limit caps that deduction at a share of a business's earnings. From 2025, OBBBA switched the measure back to a friendlier, EBITDA-based version, which lets capital-heavy businesses deduct more interest because depreciation is added back. We covered that change here. But there is a catch: when interest rates rise, as they did this week, interest costs grow faster than the room to deduct them. Higher rates squeeze highly leveraged borrowers twice, once in cash and once in tax.

3. The wrappers that pass the tax through. Many of the funds doing this lending are business development companies (BDCs). Many of the landlords are real estate investment trusts, including specialised data-centre REITs. Both structures can avoid paying corporate tax themselves if they meet certain rules and pay out most of their income to investors, so the tax is paid by the investors instead. That makes them efficient vehicles for owning loans and buildings. It also means they live and die by income: if borrowers stop paying, distributable income falls, and so do the payouts investors rely on.

4. Foreign money and the "is this a business?" question. A lot of the cash behind private credit comes from overseas pension funds and sovereign wealth funds. For them, US tax can turn on whether the fund is just investing in loans or is actually running a lending business in the US. Interest that qualifies for the portfolio interest exemption can be tax-free for foreign investors, while income treated as effectively connected with a US business can be taxed at full rates. That is why private credit funds spend so much effort on structure, from blockers to the "season-and-sell" approach of buying loans after a bank originates them. As AI lending gets bigger and more bespoke, staying on the right side of that line gets harder.

None of this means the deals are tax-driven gimmicks. The projects are real, and so is the demand for computing power. But generous write-offs and pass-through structures make the numbers work better for owners and lenders, and that helps explain why so much money piled in so quickly.

The worry: a loop of money

The sharpest warning came from the International Monetary Fund. Its October 2026 Global Financial Stability Report, released on October 13, looked at 14 large disclosed AI transactions, according to a summary by Tech Times. It found that the biggest tech companies had provided about $300 billion of equity and about $650 billion of debt, including off-balance-sheet vehicles. In turn, AI companies had committed about $1.4 trillion of future revenue back to those same tech companies through data-centre contracts, computing purchases and service agreements.

The concern is that the money goes round in a circle. The revenue meant to repay the debt flows back to the same companies that provided the capital. If AI revenue disappoints, the IMF warned of a "triple whammy": equity losses, trouble servicing debt and broken revenue commitments, all at once.

The IMF also flagged private credit directly. The report warned that "rising leverage and contingent liquidity risks in private credit could amplify stress", and that liquidity mismatches at insurers and pension funds have grown. "Insurance companies and pension funds are increasing allocations to illiquid assets," the IMF's Caio Ferreira said, according to Tech Times.

The second worry: software loans

There is a quieter AI risk in private credit too, and it points the other way. AI might not just need financing. It might also disrupt the companies private credit has already lent to.

A March 2026 study by the Bank for International Settlements found that private credit loans to software-as-a-service companies grew from almost $8 billion in 2015 to over $500 billion by the end of 2025, about 19 per cent of private credit funds' direct loans. BDCs, which make up about a fifth of US direct lending, had lent more than 15 per cent of their loans to these software firms.

Investors noticed. Between October 2025 and February 2026, software stocks fell almost 30 per cent and BDC share prices fell about 10 per cent on average. BDCs with the most software exposure did about 5 percentage points worse than the rest. The BIS linked that to fears that AI tools could undercut traditional software businesses.

So private credit sits on both sides of the AI trade. It is lending to build the infrastructure, and it has lent heavily to software companies that AI could disrupt.

The case for calm

There are good reasons not to panic.

First, many of the biggest deals are backed by the strongest companies in the world. A lease from Meta, Microsoft or Google, plus a residual value guarantee, is very different from an unsecured loan to a start-up. That is why the Hyperion bonds were rated A+.

Second, demand is real. Julius Baer's Mark Matthews told Reuters he was not worried, pointing to reasonable valuations, no sign of recession and demand that "still massively exceeds supply in most parts of the AI story".

Third, private credit funds are built to hold loans for years. Unlike a bank, a closed-end fund does not face a run on deposits if markets wobble. Lenders who underwrote conservatively can sit tight.

Fourth, this week's moves were mostly about price, not defaults. As Saxo's Chanana put it, "Investors are not questioning whether AI demand exists." The question is what to pay for that growth when rates are high and volatile.

The case for worry

The worries are just as serious.

The structures are new and untested. Off-balance-sheet joint ventures, short initial leases and residual value guarantees spread risk in ways that are hard to see from the outside. If a tenant walks away after the first lease term, lenders depend on the building being worth what the models assumed.

The money is concentrated. A handful of tech companies and AI labs sit at the centre of both the borrowing and the revenue, which is exactly the circularity the IMF flagged.

Higher rates hurt twice. They raise the cost of new debt and, through the interest deduction limits, can shrink the tax benefit of borrowing.

And private credit is less transparent than public markets. Loans are not traded every day, so losses can show up slowly. The BIS noted that deepening discounts on BDC shares may signal doubts about how private loans are being valued.

What it means for private markets

For private credit managers, the next few months are a test of underwriting. Deals with strong tenants, long leases and solid guarantees should hold up. Deals that relied on cheap money, short leases or start-up borrowers will be tested if rates stay high.

For private equity, the AI boom has been a source of both deals and exits. If financing gets tighter, infrastructure deals get more expensive, and listings for AI-linked companies get harder to pull off, as the collapsed IPO shows.

For real estate investors, data centres have been the hottest property type. Their value depends on tenants renewing leases and on demand for computing power staying strong.

For limited partners such as pension funds and insurers, the question is concentration: how much of a portfolio is now tied, directly or indirectly, to the same handful of AI companies?

What to watch

The Fed on October 28. Futures markets priced roughly a three-in-four chance that the Fed holds rates, according to CME FedWatch data cited by Reuters. A hike would add pressure on leveraged borrowers.

New AI financings and credit swaps. Watch whether big deals, like Stack's $6 billion Melbourne loan, get done, and at what price, and whether swaps on Oracle and CoreWeave keep widening. Wider spreads or banks holding more of a deal than planned would be warning signs.

BDC results. Third-quarter numbers will show whether loans to software companies are starting to slip into non-payment.

Tax guidance. Treasury and the IRS are still issuing rules under OBBBA. Any change to how depreciation, interest limits or pass-through rules apply to large infrastructure ventures would change the economics of the next wave of deals.

The AI boom is not over. But the bill for building it is getting more expensive, and private credit is holding a big share of it.

This article is analysis for general information and is not tax, legal or investment advice.

Sources