Is AI Financing About to Hit a Minsky Moment?

(The author of this article, Li Wei, is Vice Dean and Professor of Economics at Cheung Kong Graduate School of Business)

Anthropic is going public. According to news reports, Anthropic may raise $100 billion through an Initial Public Offering (IPO), while its own market value will reach as high as $1.5 trillion to $2 trillion. Another giant enterprise is about to be born.

However, little known is that just recently, Anthropic had just reached a massive debt financing agreement. The agreement is called "Project North Star." Under this agreement, Anthropic separately finalized two financings. One of them is about $42 billion, underwritten by a banking syndicate and guaranteed and endorsed by Broadcom; the other is about $18 billion in higher-risk subordinated debt, which does not enjoy Broadcom's guarantee, with half of it funded by Blackstone.

"Project North Star" is the largest AI (Artificial Intelligence) infrastructure debt financing in the world to date. Even more dramatically, Broadcom's Chief Executive Officer (CEO) Hock Tan was deeply involved in this transaction.

Why was Broadcom busy running around for Anthropic's financing? This requires starting from the beginning. Anthropic is a leading AI company that needs massive computing power to support it, and computing power comes from chips. In "Project North Star," the parties first set up a shell company. This shell company will borrow this over $60 billion in funds to purchase chips designed by Broadcom and Google called Tensor Processing Units (TPUs).

Anthropic rents the computing power of these TPUs, and the rent it pays will be used to repay the loans owed by this shell company, while Broadcom, relying on its strong financial strength, provides a guarantee for lenders to recover their funds. Since the borrower is this shell company, Broadcom only bears risk when Anthropic is unable to pay.

At the final juncture of facilitating the transaction, Hock Tan did two things: first, he provided a guarantee for the $42 billion syndicated loan; second, he raised the interest rate. Foreign media commented: "In this transaction, Broadcom simultaneously plays multiple roles as supplier, guarantor and negotiating party, and its CEO also personally entered the fray to mediate. All of these reflect the closely intertwined nature of the massive AI gamble that is currently reshaping the macroeconomic landscape, as well as the extremely high risk exposure faced by all participants."

In fact, the biggest problem now is that investors such as banks are very worried about giving huge amounts of funds to the AI industry, so they either choose to avoid such financing activities or demand more guarantees and higher interest rates. For a transaction involving a leading company like Anthropic, financing is still so difficult, so one can imagine the situation of the small fry in the industry.

Now the relatively consistent view inside and outside the industry is that AI will bring earth-shaking changes to the entire economy. According to the IPO filing provided by Anthropic, the company hinted that if all demand can be met, it will face a market opportunity of up to $40 trillion per year. This figure will exceed the massive market size estimate of $28.5 trillion given by SpaceX earlier this year.

On one side is enormous hope for the future, and on the other side is excessive fear of the debt market. How will the future market evolve? This is what this short article by the author wants to explore.

Unprecedented massive investment

Four years have passed since OpenAI launched the milestone ChatGPT in November 2022. AI has gradually gone from a niche topic among insiders to a hot global topic. No matter the industry, if one does not say a few words about AI, it seems one is somewhat out of touch with the times.

Take the Federal Reserve as an example. Current Chair Kevin Warsh has repeatedly said when discussing inflation that AI will ultimately permeate every aspect of the economy, greatly increase productivity, and thereby reduce inflation.

However, before bringing about a substantial increase in productivity, AI first brings enormous investment demand. For AI to work, computing power is needed—not a little, but massive amounts of it. Computing power is supported by chips, and to truly build the computing power that supports large AI models, data centers need to be built. Data centers are places where large numbers of chips are connected to function. The cost of a data center mainly consists of two parts: first, chips, which are the main cost; second, auxiliary items, including but not limited to buildings, cooling equipment, connecting wires, and so on.

In recent years, due to the rapid development of AI and people's frenzied expectations for AI, market investment in data centers has risen sharply. AI developers are all aggressively stockpiling chips, which has made NVIDIA, the producer of Graphics Processing Units (GPUs, the most mainstream AI chips), become the company with the highest market value in the world. NVIDIA's GPUs have long been in a state of supply falling short of demand.

So how large will AI investment actually be? Has anyone made a forecast? Believe it or not, someone has. According to estimates by Stijn Van Nieuwerburgh (Van Nieuwerburgh), a professor of finance at Columbia Business School, under a set of assumptions, U.S. investment in AI construction from 2025 to 2032 will reach as high as $10.3 trillion.

Van Nieuwerburgh also made a historical comparison. He calculated the share of gross domestic product (GDP) accounted for by some of the largest infrastructure investments in U.S. history, and found that AI construction may account for the highest proportion of GDP (Chart 1).

Source: Van Nieuwerburgh (2026).

Because the investment in AI construction is enormous, even the tech giants in Silicon Valley cannot bear the costs alone. According to Van Nieuwerburgh's research, the five major Silicon Valley cloud computing companies—Oracle, Amazon, Microsoft, Alphabet, and Meta—had total hyperscaler capital expenditures of $96.8 billion in 2020, while their total operating cash flow at the time was $252.2 billion. By 2025, these two figures had become $415.8 billion and $603.2 billion. In 2026, these five companies' spending on data center construction continues to soar, and we estimate that we will see a reversal in the relationship between these two figures for the first time. It is projected that the two figures will be $800.5 billion and $707.1 billion, respectively. This is why, as you have seen, these five companies have been especially active in raising funds in the market this year.

The Hyperion data center is a large data center with power demand as high as 2.0 gigawatts and total investment as high as $30 billion. Meta initially held the project entirely, but then sold an 80% equity stake to Blue Owl for about $2.5 billion. The resulting joint venture, Beignet, raised $27 billion in external debt in October 2025. The debt was rated A+ by S&P, one notch below Meta's own corporate credit rating, and represented the largest single investment-grade corporate debt issuance in U.S. history. The data center built by Beignet is leased by Meta, and Beignet uses the rent paid by Meta to service the debt.

Beignet's $27 billion in debt implies a debt-to-assets ratio of about 90%, while Meta's own debt-to-assets ratio is only about 43% (as of the third quarter of 2026). This contrast illustrates how off-balance-sheet project financing can support underlying infrastructure carrying higher leverage than typically appears on the balance sheets of investment-grade technology companies.

For Meta, Beignet is a special purpose vehicle (SPV) with its own separate balance sheet, and Meta's assets in it are off-balance-sheet assets. This structure isolates the project from Meta's broader corporate balance sheet and enhances protection for Meta's own equity holders and creditors, but the project's financial risk is mainly borne by the project company's majority shareholder Blue Owl and its creditors, while Meta, as a 20% minority shareholder, easily shifts most of the ultimate risk exposure to other investors. This makes it easy to understand why the project company's debt rating is one notch lower than Meta's.

Meta arranged the specific financing through five four-year lease terms starting in 2029 and extending to the bond's maturity in 2049. On each renewal date, Meta may terminate the lease for some or all of the Hyperion computing capacity. If it does so, the affected assets will be sold, and Meta must cover any shortfall between the sale proceeds and the contractually specified minimum value. This residual value guarantee gives Meta operational flexibility while retaining most of the long-term credit support that lenders typically receive in traditional long-term leases.

The accounting treatment of these commitments creates a further distinction between reported leverage and economic leverage. Under current Generally Accepted Accounting Principles (GAAP), Meta does not recognize future lease obligations or residual value guarantees on its balance sheet until the relevant lease or guarantee takes effect, even though one of these two commitments will ultimately be honored. As a result, the leverage ratio disclosed in Meta's financial reports understates the long-term economic obligations embedded in the contracts.

This off-balance-sheet structure is not without cost. Beignet's debt was issued at a yield of 6.58%, at least 100 basis points higher than the rate Meta could have paid at the time on comparable unsecured corporate debt. Over the financing term, this spread means Beignet will incur more than $500 million in additional interest expense, a cost ultimately reflected in Meta's lease payments. The benefit is that Meta isolates its balance sheet and avoids reporting the project's full leverage directly at the corporate level.

For Meta, therefore, the economic rationale is not cheaper financing but maintaining the image of a capital-light company. By placing the project in a separately financed SPV, Meta avoids adding the full amount of the related assets and debt to its own balance sheet and limits the impact on the capital intensity, leverage ratio, and credit metrics of its own financial statements. For a company valued partly on high growth and relatively low capital intensity, these benefits may outweigh the higher borrowing costs embedded in the lease.

Beignet bondholders receive a higher interest rate because they have no direct claim on Meta's balance sheet. Instead, repayment depends on the cash flowing through the project structure and on Meta's lease and residual value guarantees, creating additional structural, contractual, liquidity, and execution risks. The bonds' longer tenor and specialized private placement structure also contribute to the interest premium.

Thus, the Hyperion transaction may provide a template for future data center financing. It combines investment-grade tenant credit, high project-level leverage, a bankruptcy-remote entity, lease options, and residual value guarantees to attract substantial external capital while preserving the tenant's balance sheet flexibility. At the same time, the structure illustrates how financial engineering can make leverage and contingent liabilities less visible—remember those SPVs we saw in the subprime crisis? Does it feel a bit familiar?

Although the Hyperion financing is contractually complex and contains numerous contingencies, investors at least benefit from public credit ratings and the accompanying analysis by rating agencies. However, publicly rated debt accounts for only a relatively small share of data center financing; the larger sums are provided through syndicated loans and private credit facilities, like the Anthropic and Broadcom projects mentioned earlier, and there is almost no public information about these facilities.

Is a Minsky Moment coming?

The Minsky Moment is a concept proposed by American economist Hyman Minsky, referring to the turning point at which prolonged economic stability leads to debt accumulation beyond a critical threshold and asset prices suddenly collapse. The theory divides financing behavior into three stages: hedge financing (cash flow covers principal and interest), speculative financing (can only pay interest), and Ponzi financing (relies on asset appreciation to repay debt). The fragility of the financial system intensifies as leverage rises, ultimately triggering a liquidity dry-up and a collapse of market confidence.

Specifically, during a long economic boom, rising asset prices lead to an increase in speculative lending, and the debt generated by excessive speculation causes a cash flow shortage for investors—that is, the cash flow generated by assets is insufficient to pay the interest due on debt. When creditors demand repayment, investors mired in debt problems must sell investments whose asset prices are relatively stable to maintain cash flow; however, at that point there are no counterparties left willing to pay such high asset prices. A series of sell-offs will drive asset prices down further and tighten liquidity, ultimately causing a collapse in asset prices.

First, tenant and end demand are highly concentrated. AI data centers are mostly single-tenant facilities, and their cash flow depends heavily on continued demand from a small number of model developers and cloud providers. Second, technology becomes obsolete rapidly. IT hardware and cooling systems iterate extremely quickly (typically with an economic life of only 3 to 6 years), creating a serious maturity mismatch risk with the long duration of infrastructure financing. Third, there is severe revenue and payback pressure. At a 10% unlevered return and a 50% operating cash flow margin, the new computing capacity added from 2025 to 2032 would need to generate $3.725 trillion in mature-stage annual revenue each year, equivalent to charging about $5.5 per hour for each installed GB300 GPU (at full utilization). If the combined annual revenue of leading large-model companies such as OpenAI and Anthropic is to support this target, it must grow rapidly at about 80% per year through 2032. Fourth, asset price correlation has risen. The market beta of publicly listed data center Real Estate Investment Trusts (REITs) has risen from about 0.5 in the past to around 1, indicating that their performance is now highly tied to the broad market and AI stocks and that they no longer have the traditional attributes of defensive assets.

AI is still advancing by leaps and bounds, but concerns are growing both inside and outside the industry about whether it will collapse. Public information shows that U.S. data center-related bond issuance reached $80 billion in 2026, of which about $37 billion was rated junk, compared with $8.6 billion for all of 2025, indicating that issuance of lower-rated AI data center-related bonds is rising sharply. As of mid-October 2026, the average yield on junk-rated data center bonds had climbed to nearly 10%, a change reportedly occurring within just one week. Corporate bonds related to emerging cloud service providers that rent GPUs to AI customers are also under pressure. At the worst of the sell-off in October 2026, CoreWeave's corporate bonds traded below 80 cents on the dollar. Heavily indebted developers Firmus and AirTrunk have postponed or canceled IPOs amid market turmoil.

A small story recently published in The Wall Street Journal vividly illustrates the mood in the industry. According to the report, a bond fund manager said Morgan Stanley has recently struggled to sell a bond issue of about $300 million. The bond is backed by the bank's loan to a Kansas City data center project called "Project Blitz," operated by Lambda. Lambda, like CoreWeave, is a "neocloud" startup that uses borrowed money to buy chips and then rents out the chips' computing power to AI developers. The bond fund manager said that although Morgan Stanley has roughly doubled the yield on the Project Blitz bonds, few buyers have shown interest. Even Morgan Stanley's daughter is having trouble finding a match—one can only imagine that other families' daughters are faring even worse.

It may still be too early to assert that a Minsky moment is imminent, but some signs cannot be ignored. Major players are finding it harder to raise financing, small fry are scrambling for money everywhere, and stocks and bonds are being hit together—the chain always breaks first at its most fragile link, then spreads to other parts. The subprime crisis once left us a rather profound lesson in this regard. On May 17, 2007, after the U.S. subprime market had already begun to collapse and New Century Financial had filed for bankruptcy protection, Federal Reserve Chairman Ben Bernanke publicly declared that the impact of the subprime market was limited and that the subprime troubles would not seriously harm the overall economy.

The subprime crisis then spread rapidly from the housing market to the credit market, dealt heavy blows to major investment banks in Europe and the United States, and ultimately set off an unprecedentedly severe global financial tsunami and Great Recession when Lehman Brothers collapsed on September 15, 2008. In his later memoirs and reflections, Bernanke admitted that at the time the vast majority of officials and economists, including those at the Federal Reserve, had severely underestimated the breadth and depth of the subprime crisis and had failed to see in advance the transmission mechanisms of the crisis through the broader financial system.

In the eyes of many, AI is an invention on par with the Industrial Revolution. A creation of that magnitude is bound to have a profound impact on human society, but not all of that impact will be positive. While we are infatuated with it, we also need a measure of reason and calm—otherwise we may not live to see tomorrow morning's sun, or may even experience a full-blown Minsky moment. That may be what history teaches us.

Exclusively first-published on the Yicai Account of Yicai Media Group. This article represents only the author's views and does not constitute investment advice.