Nvidia’s Next Moat Is Made of Debt

Nvidia has delivered a masterclass in ecosystem building. It may have built a moat that has nothing to do with making better chips.

Two onchain lenders, USD.AI and GAIB, finance AI chips with dollars raised from stablecoin holders.

Their collateral lists reveal something awkward for Nvidia's competitors: they will lend against Nvidia's chips and no other chipmaker's.

That verdict has little to do with which processor is technically best. It records which one a lender believes it can recover money from if the borrower stops paying.

Nvidia's best-known advantage is software. Most AI programs are written for CUDA, Nvidia's own toolkit for programming its chips, so moving to a rival's chip means rewriting and retesting code. Lenders may be building it a second moat.

Banks and private credit funds have learned how to lend billions against Nvidia GPUs. They can:

Competing chips can be financed too, but often need a stronger customer contract or a manufacturer willing to stand behind the risk.

That creates a loop.

The more Nvidia chips are bought, rented and resold, the easier they become to finance. The easier they are to finance, the more customers can afford to buy them. A rival could make a perfectly good chip and still struggle to compete if its customers find it harder to borrow the money to install one.

It is a strange advantage for an asset that depreciates so quickly. In early 2024, at the height of the AI chip shortage, an hour on Nvidia's H100 could rent for about $8. By late 2025 a one-year contract cost about $1.70 an hour, and Silicon Data's index now puts an hour at about $2.90. Nvidia brings out a faster generation roughly every year. Yet lenders keep financing the chips: three of this year's biggest AI-chip loans alone came to about $46 billion. The largest of them was on Google's chips, and it needed the chips' co-designer to stand behind it.

The implications go beyond Nvidia's market share. Governments increasingly want AI infrastructure they control themselves. But if the easiest hardware to finance comes from one US company, does the credit market quietly push them towards the supplier they want to depend on less?

If a chip goes out of date that fast, what is the loan secured on?

CoreWeave rents AI chips to customers such as Microsoft. It began in 2017 as a crypto miner, using graphics chips (GPUs, the kind AI now runs on) to mine the cryptocurrency Ether, before turning to renting the chips out instead. By its own account, it was the first company to borrow against AI chips. In 2023 it lined up a $2.3 billion loan, secured on its Nvidia chips and its contracts with Microsoft, at an interest rate of about 14%.

Michael Intrator, its chief executive and a former hedge-fund manager, explained how the loans work on the All-In podcast in March. CoreWeave signs the customer first. Then it sets up a separate company, which he calls the box, and puts three things inside: the customer's contract, the chips and the lease on the data centre. The customer's money never passes through CoreWeave. In his words: "they pay the box. They don't pay me."

The box pays the data centre and the power bill first, then the lender. CoreWeave gets whatever is left. If the payments stop, the lender can take what is in the box.

Lenders have liked that enough to keep cutting the interest rate. Intrator says CoreWeave's cost of capital has fallen by six percentage points in two years.

In March, CoreWeave lined up $8.5 billion of loans, the fixed-rate part at about 5.9%. Rating agencies graded the debt investment grade, the label that lets insurers and pension funds buy it, and CoreWeave called it the first chip-backed loan to earn one.

In August it lined up another $2.6 billion at a floating rate that works out at about 9.4% today. This time the agencies graded the debt below investment grade.

The borrower was the same and so was the chip supplier. The contracts in the "box" were different. The March loan rested on a single contract with a customer CoreWeave described only as a leading AI company. The August loan rested on a mix of customers with contracts averaging about three years, for a loan lasting about five.

Then, in October, investors turned on CoreWeave itself. Its shares fell 15% in a single day, the yield on its ordinary bonds due 2031 rose above 16%, and the cost of insuring its debt against default hit a record. Those bonds have no chips behind them. The loans in the boxes were built to hold up when the company does not, and whether lenders keep trusting the boxes more than the company is now the thing to watch.

In June, Apollo and Blackstone, two of the world's largest investment firms, arranged $35 billion of debt to buy Google's own AI chips and lease them to Anthropic, the company behind the Claude AI models. Google's chips are not Nvidia's. Google publishes list prices for renting them on its own cloud, but I can find no independent market price for them.

Most of that debt was still privately rated investment grade, helped by Broadcom, the company that co-designs the chips. Broadcom promised to cover the senior lenders' losses if Anthropic stopped paying and the chips sold for less than was owed. Most of the protected debt pays 5.75%. The slice without the promise pays 8.5%.

I started out assuming these loans were secured on the chip. On this evidence a lender wants two things: a customer whose payments it can count on, and something to fall back on if they stop. With Nvidia's chips the fallback is a market, because they can be re-rented or sold. With Google's, which have no such market, it was Broadcom's promise.

Those three deals add up to about $46.1 billion, and they are far from the year's total. A bigger one, about $60 billion of Broadcom-backed debt for chips Anthropic will lease, is being sold to lenders now.

That is where the moat sits. A liquid market in used Nvidia chips makes the lender's fallback easier to price, which can reduce the protection a lender needs from anyone else. Other chips can still be financed, but financing Nvidia's is becoming routine, which is what entrenches it.

Nvidia also helps pay for its own customers' expansion, which feeds that market directly.

In September 2025, CoreWeave disclosed an order with an initial value of $6.3 billion under which Nvidia would buy residual computing capacity that CoreWeave could not sell to other customers, through April 2032. The commitment is subject to delivery, availability and termination conditions. Nvidia was not promising to buy back the physical chips. It was agreeing to pay for some of the hours those chips could produce.

Then, on its August 2026 earnings call, Nvidia explained the model plainly. It provides a take-or-pay commitment on part of a cloud operator's capacity: a minimum revenue floor that helps independent lenders finance the project. In return, Nvidia gets a share of the rental revenue above that floor. Its chief financial officer, Colette Kress, said: "In this model, we get paid twice, once on the hardware sale and again through the share of rental revenue."

Nvidia sells the equipment, promises to buy some unused computing time if demand disappoints, and makes the project easier for somebody else to lend against. The resulting installations enlarge the pool of Nvidia hardware that can be rented and resold. The guarantee helps create the market liquidity that may make future guarantees less necessary. Kress made the lender's case on the same call too, describing Nvidia's computing as "fungible and durable" and able to be "redeployed to support other customers".

AMD shows the other side of the same strategy. In February, The Information reported that Goldman Sachs was providing a $300 million loan for Crusoe to buy AMD chips. AMD agreed to rent those chips itself if Crusoe could not find customers. The arrangement reportedly helped secure a rate of about 6%. Nvidia has no monopoly on the tactic. The question is which manufacturer can support its customers' financing, and eventually develop a market deep enough that lenders no longer need the same degree of support.

If Nvidia's chips also need Nvidia standing behind them, the moat may lie as much in its balance sheet as in its resale market, and a balance sheet is something a rival can bring. AMD's guaranteed loan priced at about 6%, below the 9% that USD.AI, the largest onchain lender, charges on its biggest Nvidia loan. Kress saw the objection coming: "we know some will call this circular financing."

Most companies that rent out chips cannot borrow like CoreWeave. Kony Kwong runs GAIB, one of the onchain lenders. He describes the typical operator as paying for growth out of its own profits or by selling shares. Borrowing from banks and investment funds is harder, he says, because few lenders understand the asset.

At the largest onchain lender, USD.AI, you put dollars in and receive a token for your share of a pool. The pool lends to companies buying chips and keeps whatever is not yet lent in PayPal's dollar stablecoin. The returns on both flow back to you. To get out, you join a queue that is paid out every 30 days, or sell your token to someone else.

The industry calls this GPU tokenisation, though what you own is a share of a pool of loans. It is a small pool: about $660 million at the end of September, much of it not yet lent out, against tens of billions in this year's big chip loans.

Does any of this need a blockchain? BlackRock's money reached AI chips without one: its funds were among the lenders in CoreWeave's first chip loan. The one thing BlackRock puts onchain is cash, in its tokenised money-market fund. USD.AI's loans are ordinary contracts enforced by ordinary courts. What the token changes is who can fund them: anyone holding stablecoins, a pool of savings with no easy route into private credit, which a small operator can now reach without a bank. Tokenisation here is distribution for money that is already onchain. The test is whether that money stays when chip loans pay depositors about 7% and Treasury bills pay just over 4%.

USD.AI publishes its prices, and they follow the same logic as CoreWeave's two loans: 7% to 9% when the borrower's customer is rated investment grade, rising to 12% to 15% when chips are rented out by the hour with no long contract. The documents say the rate depends on the customer's contract and the size of the loan relative to the price of the hardware, and that the borrower's own credit does not come into it.

The weaker the contract, the more of the loan rests on the chip. USD.AI's underwriting paper says its main protection is what it can recover from the hardware. That goes a long way to explaining its rule on collateral: only Nvidia's enterprise chips, the ones with a deep market to re-rent or sell into. Of all the lenders in this piece, the onchain ones have the least to fall back on besides the chip, which makes their collateral list the plainest evidence of the moat.

Its largest loan so far, announced in September, is $128.9 million to an unnamed listed company for 2,304 recent Nvidia chips in British Columbia, with a customer USD.AI describes as investment grade. USD.AI says the rate is 9% over three years once the money is drawn, and 6.5% while it waits in escrow.

CoreWeave's investment-grade loan in March cost about 5.9%. Both rest on strong customers. Lambda, another chip-rental company far smaller than CoreWeave, paid 6.78% this month with investment-grade customers behind its loan, so much of the three points between CoreWeave and USD.AI looks like the cost of borrowing outside the banks rather than of being small. That need not last: three years ago CoreWeave was itself paying a rate in the teens.

Some of this lending exists because the big operators turn customers away. Intrator says he turns down customers who want chips for a single year, because that is too short to pay off the hardware. A smaller operator that takes such a customer is the borrower USD.AI's most expensive tier is priced for.

Shrink that $128.9 million loan to a single chip and it comes to about $56,000 per GPU. Repaid over three years at 9%, that is roughly $2.40 for every hour of those three years, rented or idle. USD.AI has not published this loan's schedule, so treat that as a model. As a rough guide, Silicon Data puts a different Nvidia chip, the B200, at about $6 an hour. A chip rented all the time pays 40 cents of every dollar of rent to the lender. At 60% use the loan takes two-thirds, before power, staff or the data centre. Rents could fall 60% before a fully rented chip stopped covering its loan, and the H100's fell by roughly three-quarters from its 2024 peak.

Depositors do not receive the full 9%. A warranty provider, Barkr, backed by the reinsurer Munich Re, takes 1.5 points a year; if repossessed chips sell for less than a value fixed when the loan was made, it covers the gap. USD.AI keeps a share of the interest too. What is left is around 7%, roughly what DefiLlama, a site that tracks these pools, shows depositors earning across the whole pool.

USD.AI also charges borrowers a fee when a loan is made, so the borrower's true cost is above 9%. That fee is where most of USD.AI's own income has come from so far: $4.7 million in the second quarter, when the loans paid $1.3 million of interest in all.

A lender that earns most of its income when a loan is made has a reason to make more loans, good or bad. I have seen nothing to suggest USD.AI has loosened its standards, and its largest loan went to its best grade of customer. It is still the incentive I would watch.

In the third quarter, only about a third of what the pool earned came from chip loans. Most of the rest was interest on dollars waiting to be lent, the idle-dollar income I wrote about recently.

That piece also points to where the depositors come from. The GENIUS Act, the US stablecoin law, bars issuers from paying holders interest, so holders who want a return have to look elsewhere, and onchain private credit is one of the few places to find it without leaving the blockchain. Some of the appetite for chip loans may owe more to that rule than to chips.

I could not find a single default on an onchain chip loan. The loans are young, and the biggest was announced in September.

By 2022, bitcoin miners, the companies that run machines to earn bitcoin, owed somewhere between $2 billion and $4 billion on loans secured by those machines. When bitcoin's price fell, the machines earned less and were worth less at the same moment. Stronghold, a listed miner, handed about 26,000 machines back to its lenders to cancel $67 million of debt. Several lenders ended up running mining machines themselves.

A used machine is worth what it can still earn. When the borrower's income and the collateral's value hang on the same price, the collateral is worth least at the moment the lender needs it. AI chips share that weakness.

Nvidia's rental guarantees add another version of that risk. If rental demand weakens across the market, the operators may be less able to fill their machines at the same time that Nvidia is obliged, under its qualifying take-or-pay contracts, to buy some of the spare capacity. That can protect lenders and operators while shifting part of the downside to the chipmaker. It does not make the risk disappear. Investors have started asking how far that support goes: on 15 October, as CoreWeave's bonds slid, Huang said CoreWeave did not need a backstop from Nvidia. Cisco's vendor financing during the telecom boom offers a warning: helping customers buy equipment can accelerate adoption, but it can also leave the supplier exposed when customers' economics deteriorate.

Chips do have two things mining machines lacked: customers on contracts, and more than one use. A mining machine can only mine. An AI chip can train models, run them, or be rented to someone who wants to do either. The H100's rent is up by roughly half from its low late last year (not a sentence I expected to write about a chip launched in 2022).

So the risk gathers at one moment: when the customer's contract ends before the loan does. On its contract-backed loans, USD.AI asks for at least two years left on the customer's contract, for a loan of three to five years.

On a three-year loan, the shortest USD.AI makes, about a third is still owed after two years of equal repayments. The loan was at most 80% of the chips' price to begin with, so that third is roughly a quarter of what the chips cost. The bet is that a two-year-old Nvidia chip can be sold for more than that, or re-rented for enough to pay it off, with a reserve of about three months of payments to buy time. For loans made this year, that moment arrives around 2028.

The little evidence there is says the bet holds. Carmen Li, who runs Silicon Data, one of the firms that publish chip rental prices, said in June that an H100 in its second year was reselling for about 85 cents on the dollar, and that her car loses value faster. Silicon Data's figure dates from late last year, when rents were near their low, and has not yet been tested by a lasting glut. Hyperbolic, which rents out spare chip capacity, treats a 45% annual fall in rents as a 1-in-200 worst case. The H100's fell faster than that between early 2024 and late 2025.

A blockchain records who owns the token. It cannot check that the chip is in the rack, switched on and not pledged to someone else. USD.AI's answers: the money sits with a trust company until the hardware is installed and verified, its legal claim on the chips is publicly registered, and the machines are monitored and insured. The blockchain is how the money is raised; what protects it is ordinary law and paperwork.

The loans run for three to five years, and depositors can ask to leave every 30 days. USD.AI pays them from loan repayments and cash it has not yet lent. In a bad month, the queue gets longer. That has not been tested either.

GAIB draws the same line as USD.AI: Kwong lists the Nvidia models it will finance and stops there. If a borrower stops paying, the lender has to re-rent or sell the chips, and that takes a market with a visible price. Several of Nvidia's chips have a published hourly price. Only one of AMD's does. For newer designs, such as the open-standard processors Tenstorrent sells, I could find no price index, no resale market and no lender.

That leaves an awkward rule for any borrower without a strong customer contract, a manufacturer backstop or a state guarantee: the less familiar the hardware is to the resale market, the harder it is to borrow against. Nvidia's installed base makes its chips easier to underwrite, and easier credit enlarges the installed base.

Even the onchain market's sovereign deal followed the rule. An early GAIB deal, $30 million for Siam.AI in Thailand last year, was billed as Asia's first sovereign AI cloud. Siam.AI is an Nvidia cloud partner, and the cloud runs on Nvidia chips.

My prediction, with fair confidence: at the end of 2027, the only AI chips USD.AI and GAIB accept will still be Nvidia's. If the list widens, I expect AMD before any open design. The open question is whether rivals can build a resale market and lender confidence without matching Nvidia's balance-sheet commitments. Nvidia faces a test of its own: if rental rates fall sharply, how much unsold computing time will its guarantees require it to buy?

Two places will show whether I am wrong. One is USD.AI's rule-making, where holders of its separate voting token decide which hardware qualifies. The other is the list of chips with a public hourly price. CME, the futures exchange, has planned futures on two chips so far, each one a price for a month of one chip's time. Both are Nvidia's. They were due to start trading on 5 October, but the regulator has extended its review to 9 November.

Anyone offered a yield "backed by GPUs" can ask what the lenders ask. Who is the customer? How long have they signed for? Does the contract outlast the loan? And if the customer stops paying, who takes the chip?

Tokenisation can bring new lenders into AI infrastructure without changing which chips those lenders trust. More capital need not mean more competition. So far it has mostly financed more Nvidia hardware.

There is one borrower that can change the calculation: a government. A state does not need a lender to believe in the resale value of its chips. It can sign a long contract to buy the computing, and a lender can lend against that instead. In July the European Union invited private consortia to build AI gigafactories, with the EU's supercomputing body and a group of member states lined up to buy computing time from them. Bids close on 12 November.

If a government's promise can make the machines bankable, could a country fund its AI from its own savings, perhaps through a tokenised bond, and use that credit to choose chips the private market will not finance? That is next week's piece.

For now, the easiest AI hardware to finance is made by one company and sold abroad subject to one government's export rules. Nvidia's software moat is well known. Its emerging credit moat may be just as difficult for a rival to cross.

See you next week.

James Smith