Issue #300

When Chipmakers and Insurers Start Acting Like Banks

When a chip company guarantees its customers' debt and an insurer bankrolls data centers, whose balance sheet holds the risk?

Business

Oswarld (Kwangseob Ahn)· Writer & editor of INLEVEL9

Add as a Google preferred source
When Chipmakers and Insurers Start Acting Like Banks

A Capitalism Mixed From Heroin and Cocaine

This week, The Economist titled its editorial on American finance “Speedball Capitalism.” A speedball is a cocktail of heroin and cocaine — a sedative and a stimulant mixed together so that each one masks the other’s signal. The editorial’s point is that America’s capital markets, having bet everything on AI, are now in exactly that state.

The numbers back that up; this isn’t just hyperbole. On July 10, SK Hynix raised $26.5 billion through a Nasdaq ADR1 listing — the largest stock sale ever by a foreign company in the US, surpassing Alibaba’s $25 billion in 2014. The Economist lines this up alongside SpaceX’s largest-ever IPO, Amazon’s jumbo bond issuance, and OpenAI’s record private funding round. In the span of a single year, the US market absorbed all of it.

The stimulant and the sedative arrived in the same week. According to an FT report on September 29, Anthropic’s IPO prospectus reportedly showed 2025 revenue of roughly $4.6 billion — 12 times the year before — alongside a net loss of about $42 billion. Around $34 billion of that net loss is an accounting charge reflecting the valuation of financing instruments that may later convert into equity. The same document’s risk factors reportedly even included a warning that AI could pose an existential risk to humanity. The very next day, September 30, the US 10-year Treasury yield hit an intraday 5.304%, surpassing its 2007 peak to reach its highest level since May 2002.

The editorial raises the question circulating at the heart of Wall Street: “What will break?” And it counters that what matters more than the first thing to break is how many other things nearly broke at the same moment.

This issue starts from another line in that editorial. Its diagnosis is that the boundaries between state and market, between financial and non-financial firms, between public and private markets, have buckled under the weight of this system. Turned into a question, it reads like this: in today’s AI infrastructure game, who is actually lending the money, and on whose balance sheet does the risk sit?

Nvidia No Longer Just Sells Chips

On August 11, Nvidia signed a memorandum of understanding with six firms — Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, and KKR. The plan is to pull in more than $500 billion in outside capital for AI infrastructure. The structure lets customers borrow against the hardware itself — GPUs, power, equipment — rather than paying for all of it upfront. What’s worth noting is that Nvidia itself can backstop2 up to $125 billion of that, roughly a quarter of the total. That means the company selling the chips is also taking on part of the credit risk of the customers buying them. The commitment amounts, interest rates, and execution timelines for each financial firm have not yet been disclosed.

Three days later, an even further-reaching structure came to light. Nvidia would guarantee neocloud3 operators a minimum GPU utilization rate, and if customer demand fell short, Nvidia itself would rent the idle capacity at a predetermined price. Sharon AI, an Australian firm, became the first case under a six-year, 40,000-GPU contract. It’s a structure in which a company that couldn’t borrow on its own credit alone can now raise money on the strength of Nvidia’s promise. If demand slips, the party that absorbs the loss first shifts from the lenders to Nvidia.

Seen this way, Nvidia now wears several hats at once. In The Economist’s words, it is the world’s most valuable company, one of Silicon Valley’s most important investors, and now also the guarantor of its own customers’ debt.

Nvidia isn’t the only one whose boundaries have blurred. Goldman Sachs, known for trading and M&A advisory, agreed to buy the ETF manager Innovator for about $2 billion late last year, and in August of this year agreed to acquire Neos for up to $2.25 billion. Assets overseen by Goldman’s asset and wealth management division topped $4 trillion at the end of the second quarter. Apollo, meanwhile, is a private equity and credit firm that owns the life insurer Athene. The long-term money raised through insurance premiums and the private credit4 it flows into now sit under the same roof.

There was a time when a balance sheet alone could tell you who was lending and who was borrowing. Now the same company sits simultaneously in the seats of supplier, investor, and guarantor, so no single ledger shows clearly where the risk has moved from and to.

The $2.4 Trillion in the Footnotes

So where does this money start? With promises made by a handful of giant tenants.

In late July, Bloomberg tallied the spending commitments held by Alphabet, Amazon, Meta, and Microsoft at roughly $2.4 trillion combined. The mix includes long-term chip purchase orders, power contracts, and data center leases running as long as 30 years. Alphabet’s share is about $900 billion — roughly 9 times what it was a year earlier — while Meta’s sits around $700 billion, about half of which is data center leases. These commitments aren’t booked as debt yet; they live in the footnotes of financial statements. That’s why they’re called off-balance-sheet5 obligations. None of this is illegal or hidden. It’s simply that, unlike corporate bonds or bank loans, these commitments don’t flow directly into debt-ratio calculations.

Flip the picture around, though, and these promises are someone else’s revenue trailer. According to reporting on Anthropic’s investor materials, the company has committed at least $518 billion in infrastructure obligations to six partners over the next 10 years, roughly 80% of which is either non-cancellable or owed regardless of actual usage. $111.1 billion of that is tied to Google, $110 billion to Amazon. For Google and Amazon, this money counts as future revenue — and on the strength of that future revenue, they go build more data centers and sign more commitments. One company’s footnote becomes another company’s backlog.

The most level-headed attempt to size all this up is a paper by Columbia Business School professor Stijn Van Nieuwerburgh, “Financing the AI Build-Out.” Presented at the Brookings Papers on Economic Activity (BPEA) fall conference on September 25, the paper estimates total U.S. AI investment from 2025 to 2032 at $10.3 trillion — an average of 3.63% of GDP per year. By that measure, this buildout is larger relative to the economy than past infrastructure booms like canals, railroads, or electrification.

What matters more is his diagnosis that the way this money is being raised has shifted. Financing that used to happen transparently on the balance sheets of large corporations is migrating into off-balance-sheet structures — joint ventures, private credit, securitization, special-purpose vehicles, lease commitments, loan guarantees. And whether these structures hold up ultimately comes down to whether the credit quality of a small handful of data center tenants stays strong.

Van Nieuwerburgh himself draws a clear line. It’s too early, he says, to conclude that AI infrastructure already carries systemic risk on the scale of past credit booms. His priority isn’t sounding alarms — it’s measurement. In off-balance-sheet structures, interlocking risks stay hidden until a downturn hits, so the industry should build transparency now — before its capital structure hardens into something harder to see through.

The People Who Lent Money Against a Promise

Money that starts with a tenant’s promise travels through several hops before it lands somewhere entirely different. A developer sets up a special-purpose vehicle to build the data center, and the lending syndicate extends credit based on a lease signed by a Big Tech firm or Oracle. Sitting inside that syndicate are private credit funds, and sitting in the limited-partner seats of those funds are life insurers.

The numbers on the insurer side have been getting heavier since this spring. By Barclays’ analysis, U.S. life insurers’ private credit holdings grew more than 20% in 2025, bringing them to roughly 10% of total assets — and at private-equity-affiliated insurers like Athene and Global Atlantic, that figure tops 15%. Fitch’s tally put the U.S. private credit default rate at an all-time high of 6.0% in April 2026, and the U.S. Treasury has assembled a team specifically to monitor insurer exposure. Not all of the trouble in private credit originates from data-center loans. But The Economist’s point is that this pressure happens to be converging right where the largest pool of long-term capital for AI infrastructure is supposed to come from.

ai boom blurred lendersSignals from the edges are piling up one by one. On September 24, Oracle sent a force majeure6 notice to Stack Infrastructure, the Blue Owl-affiliated developer of the “Project Jupiter” data center in New Mexico. The reason given was the possibility of delayed power availability, and reports followed suggesting Oracle wanted to push back payments if the project slipped. Oracle said it was still on schedule, and Blue Owl said its financial covenants were unchanged. Even so, this campus is the facility built for OpenAI, backed by $18 billion in syndicated bank loans. This is exactly the point The Economist zeroes in on: power delays aren’t some unforeseeable shock in this industry — they’re a risk everyone already knows about — and yet it’s being labeled force majeure.

Five days later, on September 29, smart-ring maker Oura postponed a Nasdaq listing that could have raised up to $2.2 billion. Demand was strong, but the stated reason was uncertainty in the IPO market, and reports suggested the offering price was shaping up to land at the low end of its range. By Renaissance Capital’s count, seven IPOs were postponed in Q3, up from four in Q2. The index is sitting near its highs, yet the people who actually have to set the price are hesitating.

Issue No. 192 argued that leverage hadn’t vanished — it had simply migrated to the periphery. This month, that periphery is getting names attached to it, one by one: the chipmaker standing behind the guarantees, the tenant looking to delay payment, the insurer holding the private credit.

Where Does Korean Money Sit in This Chain

Korea isn’t just watching from the sidelines. SK Hynix’s $26.5 billion ADR made the Economist’s list of record-breaking deals. And as I covered in Issue 175, when the AI hedge fund rumored to be a cornerstone investor in that offering went through a leveraged unwind in July, the shock transmitted directly into Korea’s domestic stock market.

In terms of hard numbers, financial institutions’ exposure still looks limited. According to the Financial Supervisory Service, as of March, Korean insurers’ exposure to private credit funds stood at roughly ₩28.5 trillion (~$20.4 billion) — the largest among institutional investors — but that’s only about 2.2% of insurers’ total assets. Combined exposure from the National Pension Service and the Korea Investment Corporation comes to around ₩18 trillion (~$12.9 billion). The FSS has said it will review insurers’ private credit holdings case by case and require immediate loss recognition whenever a default is reported. In April, S&P estimated that major Korean institutions’ exposure to U.S. private credit amounted to roughly 2.1% of their investable assets, concluding that the direct impact on the financial system was limited. Still, it flagged that some mutual aid associations carry exposure averaging around 10%, which could weigh on their returns.

So for Korea, the question worth asking isn’t the amount — it’s position. Is the fund holding our money sitting close to Big Tech itself in this chain, or closer to the edge, where the capital is backed by nothing more than a promise?

Oswarld’s Lens

These days I’m traveling to various regions for advisory, design, and due-diligence work on AI data center projects. When a review request comes in, I check two things before I even look at site size or GPU count: whether there’s a confirmed tenant, and whether the lending syndicate is moving on that tenant’s credit.

There’s a notable difference in the contracts, too. Looking at CPU-based service terms for CDN operators or telecom companies, most assume no changes during the operating period. But in neocloud contracts running 5-6 years, a clause requiring mutual agreement whenever the equipment generation changes is often baked in by default. [Editorial note: needs author confirmation] In effect, both sides acknowledge from the outset that the nature of the collateral can change within the contract term.

This is also where Professor Van Nieuwerburgh’s paper lands. Whether the off-balance-sheet structure holds up depends on the credit of a handful of tenants.

So I think the question to ask before “what might break” is “how many times is the same promise being counted.” The $2.4 trillion in Big Tech footnotes is lease income to the developer, collateral to the lending syndicate, and private-credit fund returns to the insurer. There’s only one promise, but multiple ledgers are each counting it as their own asset.

Nvidia’s guarantee is a point worth examining separately in this picture. That’s because the supplier is layering its own credit onto a project where the lending syndicate won’t move on tenant credit alone. The fact that a guarantee was attached also means the tenant’s credit alone wasn’t enough there. Going forward, I think where this guarantee shows up, and how fast it grows, will tell us more than any index will.

Closing

The Economist’s editorial closes by noting the scene where the market is on the verge of an all-time high even as Oura talks about “uncertainty,” and ends with the ironic suggestion that as long as the machine keeps running, it might be best to keep running with it.

Two things are true at once: the machine is running, and the machine’s creditworthiness rests on a handful of names. Once Anthropic’s prospectus is made public, the counterparty and terms behind the $518 billion commitment are worth checking against the original document, not just the news coverage.

This piece is not investment advice regarding any specific stock or asset. The figures cited are based on publicly available materials and news reports as of publication, and please keep in mind that the reporting on Anthropic’s prospectus concerns a document that has not itself been made public.

💬 Reader, which do you think is the weakest link in this chain? Pick one — Big Tech, Nvidia, private credit, or insurers — and leave a comment.

📨 If you have a colleague working in data centers or infrastructure finance, please forward this piece to them.

Your take shapes the next issue

What resonated most in this issue, or where has your experience been different?

Any registered reader can comment for free.

References & Further Reading

Primary sources

Background

Past issues worth reading alongside this one

  • Issue 192, “Where Did the Railroad King’s Dividends Come From?” ··· The starting point for the diagnosis that leverage has migrated to the periphery. Today’s piece traces the names that are now attaching themselves to that periphery.
  • Issue 198, “The Contract Google Can Exit in 90 Days” ··· A single contract that shows just how fragile a tenant’s promise can be.
  • Issue 175, “Liquidated Even When You’re Right” ··· The path by which SK Hynix’s ADR and the unwinding of AI hedge funds crossed over into Korea’s domestic stock market.

Illustrated portrait of Kwangseob Ahn (Oswarld)

The author is Oswarld (Kwangseob Ahn). Current roles: Adjunct Professor at Sejong University, Strategy Consultant at INLEVEL9. Career, research, books, and recent work are kept current on the About page. Latest · July 2026: HEMA-2: A Consolidation-Aware Tri-Memory Architecture with Multi-Channel Scheduling for Lifelong Conversational AI.

📝 Glossary

Footnotes

  1. ADR (American Depositary Receipt): A certificate issued by a U.S. bank that holds a foreign company’s shares on deposit, allowing them to be traded on U.S. exchanges. For American investors, it has the effect of buying foreign-company shares in dollars. ↩

  2. Backstop: A promise to absorb losses as a last resort if things go wrong. In lending, it describes a structure where a guarantor takes on a portion of the obligation if the borrower can’t repay. ↩

  3. Neocloud: A new breed of cloud provider, like CoreWeave, that specializes in renting out GPU compute for AI. Because these firms are smaller and carry less credit standing than major cloud providers, borrowing the money to buy GPUs comes with tougher terms. ↩

  4. Private credit: A market in which funds—rather than banks—lend directly to companies. Because it isn’t traded on public markets, it isn’t priced daily, and trouble tends to show up on the books later than it would elsewhere. ↩

  5. Off-balance-sheet commitments: Future payment obligations that haven’t yet been recorded as liabilities on the balance sheet but are instead noted in the footnotes. Leases that haven’t started yet or long-term purchase commitments are typical examples. ↩

  6. Force majeure clause: A clause allowing a party to delay or be excused from contractual obligations when an event beyond its control occurs—war, natural disaster, and the like. Whether a given event qualifies is often contested based on the exact contract language. ↩

Related issues

Issue #299

Google Ads Has Been Dying for 25 Years

Twelve obituaries came and went, revenue crossed $294 billion—so what does that curve leave out?

Business
Issue #298

541,000 EV Chargers Later, Drivers Still Ask: Does It Work?

Korea counts chargers by the unit; drivers judge infrastructure by whether one plug actually works and what it costs per kWh.

Business
Issue #297

They Stole an Nvidia Trailer. It Was 18 Tons of Sand

Two trailers bearing the Nvidia logo vanished, and when thieves pried the doors open, all they found was sand.

Business