An analysis has found that if AI-driven financial instability becomes reality as global artificial intelligence companies ramp up borrowing, South Korea’s economy could take the biggest hit. The argument is that South Korea — the biggest beneficiary of the AI boom and tightly connected to global supply chains — could become a transmission channel for real-economy and financial shocks.

The Bank of Korea’s Research Department reached this conclusion in a report titled “Key Features and Risk Assessment of AI Industry Fundraising” released on the 27th. According to the report, combined corporate bond issuance by the five major cloud providers — Alphabet, Microsoft, Amazon, Meta and Oracle — came to about $230 billion (~305.1 trillion won) in January-September of this year, already more than double last year’s full-year issuance. That is equivalent to 33% of U.S. Treasury issuance over the same period (two-year and longer maturities, excluding inflation-linked bonds). The five companies’ share of outstanding U.S. investment-grade corporate bonds also more than doubled in a year, from 2.0% to 4.8%.

Kim Yun-kyung, a department head, and Wang Jae-gon, a research officer, who wrote the report, flagged two warning signs in the structure of AI investment. First, the “vendor financing” structure — in which suppliers lend to customers so they can buy the suppliers’ own products — resembles the practices of telecom equipment makers during the dot-com bubble. Second, off-balance-sheet transactions through special purpose vehicles (SPVs) and joint ventures (JVs) are increasing, making it harder to identify where risk ultimately lands — a pattern similar to the precursors of the global financial crisis.

The most notable point is circular trading among companies. When a large cloud company or Nvidia invests in an AI developer, the developer uses that money to buy the investor’s cloud services and semiconductors. According to estimates citing Bank for International Settlements (BIS) data, circular financing accounts for more than 50% of fund flows among AI developers, cloud providers and chipmakers. Contracts with OpenAI and Anthropic are estimated to make up an average of 38% of the order backlogs at the four cloud companies — Alphabet, Microsoft, Amazon and Oracle.

Cloud ProviderShare of AI Company Circular Transactions (Based on Order Backlog)Oracle47%Microsoft41%Alphabet39%Amazon20%

Note: Estimates based on BIS data cited by the Bank of Korea and the bank’s own analysis.

This structure raises concerns that it may inflate apparent real AI demand. The International Monetary Fund (IMF) also noted that revenue at companies participating in circular financing can diverge from fundamentals, potentially overvaluing the companies. Credit rating agency Standard & Poor’s (S&P) downgraded Oracle’s credit rating by one notch from BBB to BBB- in July, citing its high revenue dependence on OpenAI.

The risk is spreading to the financial sector as well. In August, Nvidia signed a memorandum of understanding (MOU) with six financial institutions — including Apollo, Goldman Sachs and KKR — to bring in more than $500 billion in external funding. It also announced a plan to guarantee up to 25% of the residual value of graphics processing units (GPUs) deployed in individual projects. The report said such arrangements ease the burden on individual companies but can broaden the scope for losses to spread to banks, insurers and private equity funds.

The weak link is young AI infrastructure companies. Total debt at CoreWeave, a so-called “neocloud” provider of GPU rental services, is 10.6 times its EBITDA, an overwhelmingly higher repayment burden than Alphabet (0.6 times) and Microsoft (0.7 times). CoreWeave’s bonds maturing in 2032 fell below 80% of face value on the 14th of this month, and its credit default swap (CDS) premium has risen 270 basis points (1bp = 0.01 percentage point) from the end of September.

Global investment banks are sounding warnings in the same direction. Justin Biemann, global investment strategist at JPMorgan Private Bank, said in a report on the 23rd that “the AI label alone no longer reassures lenders” and that “companies must demonstrate that they can raise investment funds, repay debt and ultimately convert spending into cash flow.” According to Morgan Stanley, weekly trading volume in CDS linked to Oracle, Alphabet, Amazon, Meta, Broadcom and Nvidia exceeded a September average of $9 billion (~11.9 trillion won), far above the $2 billion (~2.7 trillion won) per week in July. Biemann said five-year CDS for a particular AI infrastructure company had soared to 1,100bp, implying a roughly 50% probability of default within five years — up sharply from below 30% a year earlier.

Still, the Bank of Korea assessed that an AI-driven financial crisis is unlikely to arrive immediately. The reasoning is that major Big Tech companies are in sound financial shape and that AI investment is based on real demand rather than phantom demand. Indeed, only 2-5% of the fiber-optic cable laid during the dot-com bubble was actually used, whereas the data center vacancy rate in major North American markets stood at just 1.4% in the first half of this year. More than 80% of data center capacity currently under construction is estimated to already have tenants lined up. The debt maturity structure is also different. Of the five cloud companies’ outstanding bonds, 61% are long-term debt with 10 or more years remaining, so even if a shock hits, it is more likely to unfold gradually than explosively. In the first half of this year, U.S. corporate savings as a whole exceeded investment, with the gap at 0.4% of gross domestic product (GDP). During the dot-com bubble (1997-2001), investment exceeded savings by 1.9% of GDP.

Nevertheless, the report focuses on South Korea’s exposure. Since SK Hynix and Samsung Electronics are key suppliers of high-bandwidth memory (HBM), the linchpin of the AI semiconductor supply chain, a pullback in AI investment would reduce data center construction and GPU purchases, directly cutting HBM demand and weighing on exports. On top of that, if financial institutions that invested in corporate bonds and loans take losses, credit supply to other companies could shrink, amplifying the shock through both the real economy and the financial system. “South Korea, the biggest beneficiary of the AI boom, could see the intensity of real-economy and financial shock transmission amplified compared with the past due to its high connectivity with global supply chains,” the researchers said, adding that “vulnerabilities are gradually accumulating amid the phase of rising long-term interest rates.” They added that potential risks at neoclouds and AI developers — which lack sufficient financial buffers and disclose information opaquely — need close attention.

The market expects the five cloud companies’ operating profit to fall below capital expenditure for the first time next year. If Big Tech’s cash generation cannot keep pace with investment, its reliance on corporate bond markets for funding can only increase. The report’s core message is that when the AI investment cycle turns, it is the funding pipeline that wobbles first — and South Korea’s memory chip industry, standing closest to that pipeline, is on the front line of the transmission path.