AI companies expand large-scale borrowing amid competition to secure chips
Financial-sector investment also grows through loan trading and securitization
Business setbacks could hit companies and the financial sector alike

Competition among artificial intelligence (AI) companies to secure chips is driving massive borrowing, allowing credit risk to accumulate in financial markets. Anthropic is seeking funding by offering annual returns above 10%, while CoreWeave, xAI and others have raised huge sums against chips and future revenue. Loan trading and asset securitization are also broadening the pool of investors betting on the success of AI businesses. But if expected earnings fail to materialize and the value of chips pledged as collateral also falls, companies’ repayment difficulties could readily spill over into financial-sector losses. In other words, the vicious cycle seen during the global financial crisis could recur: bad loans cause financial-sector losses, triggering a credit crunch that further weakens companies’ ability to repay.

Anthropic’s Borrowing Rate Rises 1.5 Percentage Points in Four Months

According to the Financial Times (FT) on the 18th, local time, Anthropic plans to raise USD 9 billion (approximately USD 9.28 billion) from Blackstone to secure additional chips. A Blackstone-managed fund will invest in a subordinated loan backed by Anthropic’s chip lease payments over the next five years. The loan carries an annual coupon rate of 10%, with investors’ total returns exceeding that level after accounting for the issuance discount and fees. Anthropic previously raised USD 4.4 billion (approximately USD 4.54 billion) at an annual interest rate of 8.5% through Project Big Sky in June. Returning to borrow just four months later, it faces a coupon rate alone that is 1.5 percentage points higher.

The financing is part of Project North Star, a USD 60 billion (approximately USD 61.9 billion) initiative. The project aims to secure Tensor Processing Units (TPUs), custom chips developed by Google, to establish 1.5 gigawatts (GW) of computing capacity. The funds raised will finance chips scheduled for delivery in 2027. Anthropic has agreed to begin making lease payments after receiving the chips. It plans to raise the required funding through USD 42 billion (approximately USD 43.3 billion) in senior loans and USD 18 billion (approximately USD 18.6 billion) in subordinated loans. The senior loans are guaranteed by Broadcom, which co-develops TPUs with Google, and arranged by Bank of America, Citigroup and Morgan Stanley. Blackstone will take up half of the subordinated loans, or USD 9 billion, while plans are underway to sell the remaining USD 9 billion (approximately USD 9.28 billion) to other investors over the coming months.

Anthropic’s choice of Blackstone as a partner for large-scale financing appears to have been influenced by their business collaboration this year. Beginning in March, the two companies discussed establishing an enterprise AI services joint venture with private equity firm Hellman & Friedman and others. In May, they officially announced the establishment of a company with Goldman Sachs participating. The plan is to combine Anthropic’s technology and development personnel with the investment firms’ capital and customer networks to introduce its Claude AI model into companies’ core operations. The main targets are the investment firms’ portfolio companies and outside midsize businesses lacking in-house development capabilities. The venture will handle everything from building AI systems tailored to each customer’s workflows to operational support. It reportedly also pursued acquisitions of technology services and consulting firms to secure the specialists needed for this work.

Table 1. Borrowing and Equipment Financing at Major AI Companies

USD 500 Billion Investment Plan Increases Reliance on External Funding

Anthropic’s enormous investment funding needs underlie the expansion of its collaboration with Blackstone from customer acquisition to large-scale financial support. Companies typically finance investments with cash on hand and funds generated from operations, covering any shortfall through loans or bond and share issuance. Anthropic recorded third-quarter revenue of USD 16.8 billion (approximately USD 17.3 billion) this year and two consecutive quarters of adjusted profitability, but it continues to need outside funding to support massive chip investments. Its planned chip and infrastructure investments alone exceed USD 500 billion (approximately USD 516 billion) over the next 10 years. External funding is needed to proceed with its investment schedule before proceeds from a public offering arrive, and as borrowing grows, lease payments and financing costs will add to its future cash outflows.

However, financing the necessary investment through private loans increases the costs associated with a company’s credit risk. The private credit market supplies funds even to companies facing constraints on bank borrowing or public bond issuance, but prices uncertainty over repayment into interest rates. In June, credit rating agency KBRA assigned Anthropic a rating one notch below investment grade. Given that publicly issued bonds from similarly rated companies yield around 7%, the returns of more than 10% offered to Blackstone entail substantial additional costs. The risks for those supplying the funds are also considerable. The subordinated loans Blackstone will take up rank lower in repayment priority in the event of default and lack the Broadcom guarantee attached to the senior loans, leaving them more exposed to losses. If Anthropic’s lease payments are disrupted, the loans relying on those payments for repayment could become impaired, likely passing losses on to the funds holding those assets and their investors.

AI Companies Raise Huge Sums Against Chips

Such credit risks are accumulating across the AI industry as companies competing to invest turn to private loans even at high interest rates. AI cloud provider CoreWeave has continued borrowing heavily against chips and customer contracts. It raised USD 2.3 billion (approximately USD 2.37 billion) in 2023 against graphics processing units (GPUs) and other collateral, then entered into a USD 7.5 billion (approximately USD 7.73 billion) loan agreement in May the following year, with Blackstone, BlackRock, Carlyle and others participating. It subsequently secured an additional USD 2.6 billion (approximately USD 2.68 billion) in July last year to expand computing infrastructure serving OpenAI. This year, it signed loan agreements for USD 8.5 billion (approximately USD 8.77 billion) in March, followed by USD 3.1 billion (approximately USD 3.20 billion) in May and USD 2.6 billion (approximately USD 2.68 billion) in August. Separately from those loans, it finalized terms for convertible bond issuances of USD 3.5 billion (approximately USD 3.61 billion) in April and USD 3.7 billion (approximately USD 3.82 billion) in September. CoreWeave’s total debt stood at USD 50 billion (approximately USD 51.6 billion) as of the middle of this month.

Elon Musk’s xAI has also tapped bond and loan markets to fund data center expansion and AI model development. It raised a total of USD 5 billion (approximately USD 5.16 billion) in borrowing arranged by Morgan Stanley in June last year, consisting of secured bonds and term loans. Uses for the proceeds included data centers and development of the AI service Grok. In January this year, large-scale financing also backed an equipment-leasing business serving xAI. Apollo-managed funds and affiliates led the financing, providing funds to an infrastructure fund run by Valor Equity Partners that was responsible for purchasing the equipment. The fund agreed to purchase USD 5.4 billion (approximately USD 5.57 billion) in computing equipment, including Nvidia GB200 systems, and lease it to an xAI subsidiary.

IREN also financed purchases of the GPUs needed to supply computing services to Microsoft (MS) through external borrowing. In June, it signed a USD 3.65 billion (approximately USD 3.76 billion) financing agreement backed by GPUs and contractual cash flows. Arranged by Goldman Sachs and JPMorgan, the deal combined USD 2.1 billion (approximately USD 2.17 billion) in private debt with a USD 1.55 billion (approximately USD 1.60 billion) loan facility that could be drawn as needed, at a weighted-average interest rate of 6%. In late August, it secured an additional USD 2.4 billion (approximately USD 2.48 billion) to purchase Nvidia Blackwell Ultra and other equipment for its Mackenzie data center in British Columbia, Canada. The transaction, involving Blue Owl and PIMCO entities, consisted of USD 1.2 billion (approximately USD 1.24 billion) each in senior secured loans and senior secured bonds, with a fixed interest rate of 9%.

Long-Term Capital Bets on Chip Lease Income and Residual Values

As borrowing backed by chips grows, loans have also begun to be resold to other investors. In July, a portion of a USD 35 billion (approximately USD 36.1 billion) private loan arranged to secure chips for Broadcom and Anthropic was traded through Wall Street banks. Issuance of asset-backed securities (ABS) is also being pursued to raise more funding for chip purchases. This involves bundling chip-related loans or rights to receive lease payments into securities and selling them to investors. Investors earn returns from loan principal and interest payments and lease payments, while financial firms can recover existing loans using the proceeds from securities sales and provide additional funding. An infrastructure financing business being prepared by Nvidia and financial firms has also discussed using this approach to expand participation by insurers, pension funds and others. The idea is to attract long-term capital by designing investment products based on chip lease income and residual values.

However, the chip lease payments and residual values underpinning returns on these investment products remain subject to uncertainty from technological change. As new chips become more widely available, demand for existing equipment may decline, pushing down both lease income and secondhand prices. Credit rating agency S&P Global has also identified rapid GPU obsolescence as a risk factor, noting that the shorter the equipment’s economic life, the harder it is to estimate recoveries in the event of default. Transactions that require remaining debt to be repaid through re-leasing or equipment sales after a lease expires are particularly vulnerable to falling prices. If several companies facing funding difficulties dispose of equipment pledged as collateral at once, the increased supply could drive secondhand prices down further. In that case, loan impairments and falling collateral values could reinforce each other, increasing losses and putting additional pressure on principal and interest payments on ABS issued against those loans.

Money Chasing High Returns in AI Could Set the Stage for Another Financial Crisis

A chain reaction of falling collateral values and impaired securitized assets also amplified the 2007–2008 global financial crisis. At the time, financial firms pooled subprime mortgages—home loans to borrowers with poor credit—to issue mortgage-backed securities (MBS), then used those securities as underlying assets to create collateralized debt obligations (CDOs). While home prices kept rising, even borrowers with limited repayment capacity could repay their debts by selling their homes or taking out additional loans. But when home prices turned downward, both refinancing and collateral sales became difficult, while rising delinquencies drove down the prices and credit ratings of related securities. Funding through securities sales also dried up, reducing new lending and weakening housing demand, creating a vicious cycle that further depressed collateral values.

The expectations of high interest income that supported risky lending then can also be seen in today’s AI financing. According to research by the U.S. Federal Reserve (Fed), initial interest rates on subprime hybrid adjustable-rate mortgages were around 8%, with rates 2–3 percentage points higher without the introductory discount. For investors, this offered an opportunity to earn high interest income; for borrowers, it meant repayment burdens would increase as interest rates rose. The returns of more than 10% offered by Anthropic likewise reflect both the repayment risk borne by investors and the financing costs the company must shoulder. If concerns about repayment capacity grow after an investment is made, more bonds may be put up for sale, with falling prices driving yields higher. In CoreWeave’s case, investor selling pushed yields on some bonds to around 15% in the middle of this month. Investor anxiety surfaced as falling bond prices drove yields upward. If such concerns lead to a contraction in lending, companies that have relied on borrowing will inevitably face delays in capital investment. If facilities start operating late and revenue fails to materialize on schedule, even their ability to repay existing debt could weaken. This means the AI investment market is also exposed to a vicious cycle in which fears of bad debt trigger a funding squeeze that, in turn, worsens credit problems.