AI Chip Financing Risks A Credit Trader Would Flag

I keep a small habit from an old trading floor. When a headline throws a giant number at me, I do not ask who raised it. I ask who is stuck with it if the story changes. This week the numbers around AI hardware have been loud enough to drown that second question. Billions moving off balance sheets. Tens of billions lent by a supplier to its own customer. A backstop large enough to sound like a floor under an entire funding machine. The coverage treats each figure like a scoreboard. A credit person hears something else. Who holds the residual, and do they know the shape of it?

There is a man who spent a career on that second question. He built the modern high-yield market, went to prison, and still managed to be early on a version of this problem. Technology, he argued years later, should fund growth mostly with equity. Stable cash flow can carry debt. A product cycle that eats itself every few years is a different animal. I have found that line more useful than most of the deal commentary circulating right now. It does not kill the AI story. It just refuses to let a letter grade do the thinking.

Why A Bond Desk Would Pause On GPU Debt

The popular memory of that era is simple and mostly wrong. A salesman pushed leverage on anyone with a pulse, the paper blew up, and the lesson was never to trust yield again. The record is messier. Public high-yield bonds outstanding climbed from a modest single-digit billions in the late 1970s to tens of billions by the mid-1980s and something near two hundred billion by the start of the 1990s. One desk underwrote a huge share of the new-issue dollars. The money built telecom, media, casinos, cellular networks, and financed raids on sleepy oil majors. By the end of the decade the annual gathering in Beverly Hills drew thousands. It looked like a permanent machine.

Then the dealer vanished. The firm filed in February 1990. Contemporary accounts described a staggering liquidity problem, not a pile of worthless issuers. It defaulted on a loan balance that was small next to the market it had created. Overnight, the one institution willing to take the other side was gone. Most of the companies were fine. Most of the paper was fine. What disappeared was the bid. I keep coming back to that distinction, because it is the part people skip when they retell the story as a morality play about bad bonds.

Two months later the man at the center pleaded guilty to six felony counts, paid a huge fine, and served less than two years of a ten-year sentence. Weeks after release he was diagnosed with prostate cancer. The next three decades went into research funding and faster paths from lab to patient, and a pardon arrived in early 2020. You can hold both facts. The legal ending was real. So was the credit insight that came decades earlier, in a library, not a courtroom.

A Student Who Trusted The Tape Over The Letter

Berkeley, 1965. An undergraduate runs into a dry monograph on corporate bond quality and investor experience, covering records from 1900 to 1943. The finding should have embarrassed a profession. Investors usually overestimated the risk in higher-yielding bonds. They treated the rating as the risk. The rating was wrong often enough that the paper everyone feared tended to beat its reputation. He later wrote that he was struck by the gap between theory and reality. Modest words for a sharp idea. A whole industry had confused a letter grade with a probability.

Most people who notice that write a paper. He studied capital structure, joined the firm that would become the high-yield house, and moved the desk to the West Coast in the late 1970s. The machine that followed was not magic. It was a willingness to price cash flows the rating agencies had flattened into a single notch. That is the part worth borrowing now. Not the legend. The habit.

Rating is not credit. Four words, and most of the current hardware-financing debate fails them.

He kept a short list of principles. The first is those four words. Debt is not good and debt is not bad. For some companies, close to zero debt is already too much leverage. In a 2009 essay he laid out factors that, when they point toward rising business risk, can make even a dollar of debt too much. A few years later he named the industry out loud. Stable, predictable revenue can carry borrowing. Technology should finance growth primarily with equity. Read that, then look at what is being sold against graphics processors this month. I suspect the man who built the junk market would hesitate before lending against a chip generation with a successor already on the calendar.

The Buyer Who Thought He Owned Bonds

The detail that belongs in this story is from 1987. Insurers, according to a government review, held more than 30 percent of junk bonds outstanding. A few years later California regulators seized a life insurer that had piled well over half of its general account into the stuff. The institution that failed was the marginal buyer. Long liabilities. A capital charge set by the rating rather than the asset. A board that believed it owned bonds.

That pattern is the one I cannot shake. Markets do not always break where the brochure says they will. They break where the buyer needed the paper to behave like something simpler than it was. Insurers in the late 1980s needed yield and a rating that fit a slot. They got both, until the slot and the asset parted ways. If that sounds familiar, it should. The structures being pitched now are built to look like corporate bonds and to behave, in part, like equipment leases with a technology kicker nobody can hedge cleanly.

What The Tape Actually Showed This Week

Start with the cloud giant. Market reports described a pitch for a special purpose vehicle, a separate company set up to hold specific assets and borrow against them. The vehicle would own roughly eight billion dollars of already-installed high-end chips, raise debt against them, and lease the hardware back. The stated goal is a lighter balance sheet. The other effect is a rating. Investors expect the paper to print investment grade on the strength of a double-A parent. That is what opens the door to insurers and pension money. The company says each series of chips will last at least five years. The same reporting noted that the current generation will soon be superseded. Both facts sat in one story. I have not seen many people hold them in the same sentence.

A day earlier, a number surfaced from a model company’s filing. A chip designer will lend that customer up to forty-two billion dollars in convertible notes, roughly a third of a five-year commitment north of a hundred billion to lease custom accelerators the designer helps build. The filing itself flags a potential conflict and warns that certain defaults could make a substantial portion of the lease obligations immediately due. Separate reporting said banks are assembling about sixty billion for the designer’s chip financing, split between a senior piece and a junior piece, with a large alternative manager leading the junior. In August a graphics leader set up a platform with several of the biggest private-capital names to pull in more than five hundred billion of outside money, and its chief executive said the company holds an option to backstop a hundred and twenty-five billion of it.

The live experiment is the specialist cloud operator. In August it closed a loan of about two point six billion, roughly five-year money, against customer contracts that average closer to three years. A social giant’s huge data-center project was financed with more than twenty-seven billion of bonds through an outside vehicle, priced in the mid-sixes, rated a notch below the parent, and kept off the parent’s balance sheet. One asset manager reportedly took about eighteen billion of that deal. One firm. Two thirds of one issue.

All of this is being priced with the ten-year Treasury above 5.2 percent. It touched 5.34 percent at the start of October, the highest since 2002. A large bank expects something like 4.1 trillion dollars of AI-related debt to be issued through 2030. For scale, the entire junk market that peaked around the old high-yield boom was near two hundred billion. The new number is not a forecast I am endorsing. It is a measure of how much paper the street thinks it can place. Placement is not the same as a held risk that has been priced.

I put that table together because the deals blur if you only remember the dollar signs. Each row is defensible alone. Stacked, they describe a market that is manufacturing its own demand and then inviting long-liability money in through a rating door.

The Rating Is Borrowed, Not Earned By The Silicon

First pass, the way a trader would read it. The rating is borrowed. A vehicle that owns depreciating silicon is being rated off the cloud parent, not off the silicon. That works for as long as the parent pays the lease, and the lease is most of the trade. Strip the structure away and a buyer of that paper is long the parent’s credit at some spread over the parent’s own bonds, plus a residual bet on what a 2026 processor fetches in 2030. The first leg is easy to price. The second is hard to hedge. I am not aware of any real forward market in used current-generation boards. If the spread does not pay for the second leg, the buyer is being paid for parent risk and taking the chip risk for nothing.

Perhaps the most interesting aspect is how clean the marketing makes this sound. Investment grade. Five-year life. A name everyone already owns. The residual sits in the footnote, or it does not sit anywhere, because there is no screen for it. Old high-yield desks lived on footnotes. They priced the thing the rating had smoothed over. I would want the same habit here. What is the recovery if the lease ends and the boards are one generation stale? Who has a bid? At what discount to the number still sitting in someone’s model?

Useful life is not a vibe. One cloud operator trimmed the accounting life of some servers from six years to five in 2025. A well-known short seller spent the start of October arguing the real economic life is shorter still. I do not know the right number, and I doubt anyone does with confidence. What I do know is that the coupon is fixed and the collateral value is not. If the collateral falls short, the lender holds the difference. That sentence is the whole trade, once you stop admiring the structure.

Debt That Can Outlive The Thing It Finances

Second pass. The debt can outlive the asset. Five-year money on chips tied to three-year contracts, with a successor generation already announced, is a bet on renewal. Lenders to the specialist cloud said as much and described themselves as comfortable underwriting that renewal risk. Maybe that is a reasonable bet. It is closer to a venture bet than a term loan, and I would expect it to be priced like one. Comfort is not a spread.

Think about a car loan where the car is redesigned every thirty months and the loan runs sixty. You can still get paid if the driver keeps renewing and the resale market holds. You are not really secured by the car in the way a warehouse lender is secured by canned goods. You are secured by a relationship and by a fashion cycle in silicon. Relationships fray. Fashion cycles do not send a calendar invite.

There is a fair counterpoint, and the old high-yield builder would probably make it himself. Anyone with a contract from the largest model labs or the largest clouds has absorbed every rate increase so far. One executive told a business network last month that fifty basis points would not stop a deal with a top lab. Probably not. My guess is that trouble, if it comes, starts two tiers down. Smaller specialist clouds that lenders have started turning away. Second-string chip lessors who priced their books when the ten-year was in the fours. Credit cycles have a habit of starting where few people are looking. The center looks fine until the edge has already repriced.

• Parent lease cash flow is the easy leg, and it can look investment grade.
• Renewal after the contract window is a separate underwriting, closer to growth equity.
• Used-chip value is the residual nobody can hedge in size.
• A backstop option is not the same thing as a standing bid.
• Concentration in one vehicle can turn a “diversified” buyer into a single-name holder.

The Marginal Buyer Looks Familiar Again

Third pass. The marginal buyer looks familiar. In 1987 it was the insurers. In 2026 it is insurers and pension money again, reached through investment-grade structures that look like corporate bonds and behave partly like equipment leases. A sharp quant put the point better than I can, back in March. The lion does not care if the ostrich is first loss or higher up in the capital structure. Tranching sorts out who takes the first loss. The size of any loss stays the same. To be fair, he has been careful not to claim he timed trouble in private markets, and I am not predicting a date either. I am describing a structure.

Structures are how risk changes clothes. A lease inside a vehicle, wrapped in a rating one notch off a famous parent, can sit in an account that is not allowed to buy the equity of the company actually taking the technology risk. That is legal. It can also be a category error. The account thinks it owns a bond. Part of what it owns is a depreciation schedule and a renewal negotiation. I have watched versions of this in other cycles. Aircraft. Shipping. Telecom switches. The wrapper was always cleaner than the metal.

In my experience, the dangerous moment is not the closing dinner. It is the first quarter when a holder needs to sell and discovers the buyer on the other side was the same club that structured the deal. Secondary liquidity is a rumor until it is tested. High-yield learned that in 1990 in the most expensive way available. The issuers did not have to be broken for the market to seize. The dealer only had to leave.

A Club Lending To Itself

There is one more pattern worth sitting with, and it is the one I think about most as a public-market investor. In 1989 a West Coast paper described the old high-yield network as a club whose members were lending to and borrowing from each other. Deals cleared because the same handful of balance sheets sat on both sides of the trade. It felt like a market. It was also a circle.

Look at the current circle without the logos. A chip designer lends to its largest customer so that customer can lease the designer’s chips. A graphics leader offers to backstop the debt its buyers use to buy its chips. Several of the sponsors in that platform also show up in the designer’s financing. Each deal is defensible on its own. Taken together, the suppliers have become a meaningful source of their own demand. The public is about to be invited into the arrangement through a model-company listing that people close to it have floated at a number large enough to reset records.

To be precise, I think AI demand is real, perhaps the most real thing in the economy right now. Usage is not a press release. Power contracts, land, and hiring say the build is physical. My concern is narrower. When the junk market seized in 1990, the companies were mostly sound. The dealer that made the market walked away, and nobody else would take the other side. For AI credit, the chips will almost certainly get used. What I would want to know is who makes a market in processor-backed paper in the quarter one of the large sponsors decides to sit out. Today, as far as I can tell, it is largely the sponsors themselves.

Demand can be genuine and the financing club can still be circular. Those are not contradictory sentences.

A credit habit, not a forecast

Circular demand is not fraud. Companies have vendor-financed customers for a century. The question is concentration and exit. If the same three balance sheets are the supplier, the lender, the backstop, and the reference customer, the system can look deep while it is actually narrow. Depth is the ability to sell to a stranger. Narrowness is the ability to sell to your cousin. Cousins are loyal until they are not.

People Are Still The Scarce Input

The part of that career I like most is the optimist. People are the scarce resource, he said. Not buildings, not printing presses, not factories. He got to chips too. For today’s computer chips, he argued, the silicon is a tiny slice of the cost. Brainpower has become the raw material for building companies. Hold that up against this week. Borrowed money is going into the factory at a risk-free rate above 5 percent, and very little of the headline flow is going to the people who have to make the factory pay.

I would guess the return on AI accrues mostly to the operating companies that take the compute and turn it into margin. Many of them are small and mid-sized, and many have not done it yet. They are sitting on the adoption gap. A leading model lab just committed a hundred million dollars to train ten thousand engineers to deploy its models inside customers. Read that as a signal. A model company is paying out of pocket to close the gap between buying compute and using it. The chips alone do not close it. The lenders are financing the supply side of a market whose bottleneck, I think, sits on the demand side, with people.

That is not an anti-capex sermon. Factories get built because someone believes the output will be worth more than the steel. The mismatch is in the capital structure. Equity is supposed to absorb the uncertainty of a new production function. Debt is supposed to clip a coupon off something that already throws cash with boring regularity. When you debt-finance the uncertain part and leave the adoption work to whoever can hire, you have inverted the old rule. The past is always triple-A, he liked to say, and the future is always single-B. He meant it as encouragement. I take it that way. The future is worth financing, as long as the people doing the financing know what they own.

Six Questions Before The Paper Clears

If I were sitting on a credit committee this month, I would not start with the total addressable market slide. I would start with six questions. They are not clever. They are the ones a high-yield desk would have asked before the circle got comfortable.

1. Does the debt pay down before the chips stop earning?
2. Should this business be financed with debt at all, or primarily with equity?
3. Whose credit is actually being rated, the chip’s or the guarantor’s?
4. Who owns the residual if a 2026 processor is worth a fraction of book in 2029?
5. Is the seller financing its own customer, and who makes the market if the dealer steps back?
6. Who turns the compute into margin?

The sixth is the one I spend my days on, and it is the only one where the answer is a company rather than a structure. Everything else is plumbing. Plumbing matters. It is also where people get hurt when they think they bought a story and received a depreciation schedule.

Walk the first question with a pencil. A five-year note against hardware whose competitive life might be three, with contracts that roll inside the loan, only amortizes cleanly if renewal is a right rather than a hope. Hope is not a covenant. If the documents give the lender a look at replacement contracts, good. If they give a parent guarantee that survives a technology miss, better. If they give a rating and a slide, you are underwriting a narrative. Narratives do not cure a shortfall.

The second question is the one the equity-versus-debt line was built for. Some of these borrowers have contracted revenue that really does look like a utility. Some have a pipeline and a logo. Mixing them in one “AI debt” bucket is how you get a false average. A utility can carry debt. A product cycle should not pretend to be a utility because the parent has a famous rating. I have found that committees hate this distinction because it slows the deal. Slowing the deal is the point.

Question three is where the borrowed rating lives. If the analysis would not survive the removal of the parent’s name, you are not analyzing the asset. You are analyzing the name and hoping the asset behaves. Question four forces a number on the residual. Even a wide range is better than silence. Question five asks whether you are inside a club. Clubs are fine until you need an outsider. Question six asks whether anyone downstream can turn watts into gross margin. Without that, the whole stack is a cost center with a coupon.

What Could Go Right, Because Something Might

A fair note has to leave room for the other side. Utilization could stay high enough that residual values hold. Power constraints could keep older chips relevant longer than the product roadmap implies, because the new boards cannot get electrons. Contracts could roll at prices that cover depreciation and the coupon with room to spare. The backstop options could be exercised in an orderly way, and secondary buyers could show up because the spread finally pays for the second leg. None of that is fantasy. It is a path. It is just not the path embedded in a rating that ignores the silicon.

There is also a macro path that helps the borrowers and hurts the lenders in a different way. If growth in model revenue is as steep as the builders claim, equity holders of the operators capture most of the upside, and the debt clips a fixed coupon that looks skinny in hindsight. That is a good problem for the economy and a dull one for the bond. The bad problem is the reverse. Growth arrives slower than the depreciation, renewal conversations get tense, and the paper discovers it was never just a bond. Both problems can exist in different parts of the stack at the same time. The top of the stack can be fine while the second tier is already negotiating.

Rates matter more than the deal announcements admit. A ten-year yield above 5 percent changes the hurdle on every leveraged hardware trade. The same cash flow that looked ample when the curve was lower has to clear a higher risk-free rate before it pays anyone for technology risk. I do not need a crash for that math to bite. I need the marginal project to be financed at a spread that assumes the old curve. Those projects are usually the ones that do not make the headline.

A simple desk checklist:
  Lease term versus loan term
  Parent support: guarantee, keepwell, or comfort letter
  Residual assumption, and who marks it
  Renewal: contracted, expected, or hoped
  Buyer of last resort if the sponsor steps back
  Concentration: one holder, one vehicle, one customer

That checklist is intentionally plain. Fancy structure is how simple risks hide. If a deal cannot answer those lines in a page, the page is the risk.

Where Public Investors Actually Sit

Most people reading this will not buy the private vehicle. They will own the parents, the suppliers, the specialist operators, or a fund that quietly owns the paper. The exposure is indirect and easy to miss. A supplier that finances its customer has manufactured revenue and manufactured a credit exposure. If the customer stumbles, the supplier can lose the sale and wear the loan. A cloud parent that moves chips off balance sheet has not deleted the obligation if the lease is the economic reality. Accounting can be lighter. The cash can still be spoken for.

Specialist operators are the cleanest public window on the renewal bet. Their equity is a levered claim on contracts that may be shorter than their debt, plus whatever the hardware is worth if a customer leaves. That can work brilliantly if demand keeps outrunning supply. It can gap if one large tenant renegotiates. I would rather own the companies that turn compute into priced products than the companies that only rent the shovels, but that is a preference, not a law. Some shovel landlords will earn superb returns. The ones that will not are the ones who borrowed as if the rental rate were a bond coupon.

Funds are the quieter channel. A pension allocation to “investment grade private credit” can end up holding a slice of a chip vehicle because the rating fit the mandate. The mandate was written for corporate cash flows. The asset is a lease with a technology residual. Nobody has to have lied for that mismatch to happen. Mandates lag product innovation. They always have. The insurer in 1991 did not set out to become a junk-bond hedge fund. It set out to earn a little more than the liability cost, inside a rating bucket. The bucket did the rest.

A Note On Scale, So The Zeros Do Not Hypnotize

Large numbers numb. Four trillion through the end of the decade is a street estimate, not a destiny. Eight billion in one vehicle is a rounding error next to the cloud parents and a very large check next to the old junk market’s early years. Forty-two billion from a supplier to one customer is a relationship, not a diversified book. A hundred and twenty-five billion backstop option is an option. Options expire unexercised all the time. Holding those distinctions in your head is the whole job. The headline adds them. A desk separates them.

Compare the old peak, near two hundred billion of junk, with the new issuance dream, and you can talk yourself into a bubble story too quickly. The economy is larger. The buyers are larger. Some of the cash flows really are contracted. The useful comparison is not the total. It is the marginal buyer and the marginal asset. In the late 1980s the marginal buyer was an insurer reaching for yield inside a rating. The marginal asset was a leveraged issuer whose cash flow was less stable than the rating implied. Today the marginal buyer looks similar. The marginal asset is a chip whose value depends on a roadmap the lender does not control. Different costume. Similar posture.

I am not arguing that every dollar of this financing fails. I am arguing that a slice of it is being sold as something simpler than it is. That slice is where the old lesson lives. Rating is not credit. The club is not a market until a stranger will bid. Debt that outlives the earning life of the collateral is a hope with a coupon. People, not boards, close the gap between installed compute and earned margin.

How I Would Want To Be Paid

If I were buying the paper, I would want to be paid for both legs. Parent credit at a spread that makes sense against the parent’s own bonds, and an explicit premium for residual silicon risk, renewal risk, and club risk. If the spread only pays for the first leg, I am donating the second. Donating risk is a popular business in a hot theme. It rarely shows up in the tear sheet.

If I were buying the equity of a supplier that is also the lender, I would haircut the growth rate by the portion that is vendor-financed, and I would ask what happens to both the income statement and the credit book if the customer slows orders. If I were buying the equity of an operator, I would want the debt maturity inside the contract life, or a very clear reason it is not. If I were allocating to a fund, I would ask what share of “investment grade” exposure is actually a lease on depreciating hardware. Those are ordinary questions. They feel fussy only because the theme is loud.

This week a good number of buyers will own famous-company credit with some used-chip risk attached. I hope they are being paid for both. The chips will get used. That is the easy sentence, and it is probably true. The harder sentence is the one from 1990. Sound assets and a missing bid can coexist. A credit trader from that era would not need a new theory to see it. He would need the old one, applied to a faster product cycle and a larger club.

None of this is a call to hide from the build-out. Power, land, models, and the people who can deploy them are real. The financing is the part that can be mislabeled. Mislabeled risk is how quiet accounts end up holding a trade they did not think they signed. If the past is always triple-A and the future is always single-B, then the honest way to own the future is to know which one you bought. This month, a fair share of the new paper is a blend. Blends are fine. Pretending they are pure is how the marginal buyer gets surprised.

I will keep the small habit. Big number, then the second question. Who is stuck with it if the story changes? On AI chip financing, that question is still more interesting than the scoreboard.