The term sheet that reached credit investors this week carried a number more often associated with a troubled retailer than with a company selling subscriptions to a chatbot. For 60 billion dollars of borrowing to buy chips, Anthropic is offering a coupon above 10 percent.
The financing, code-named Polaris, is being arranged to fund the purchase of roughly 1.5 gigawatts of Google’s tensor processing units, according to pricing details reported on October 16 by the Financial Times and Bloomberg. A fund managed by Blackstone will commit 9 billion dollars, taking half of an 18 billion dollar subordinated loan at a 10 percent coupon. After original-issue discount and fees, the all-in yield to investors exceeds 10 percent.
The comparison that matters is in the public market. Borrowers rated BB, the top rung of the high-yield category, currently pay close to 7 percent. Anthropic is borrowing at a spread of roughly three and a half percentage points over that, and it is doing so in private credit rather than through a bond syndication.
Part of the premium is the structure. The debt is being raised against equipment rather than against the whole company, which means lenders are underwriting the resale value of accelerators that lose value quickly. A chip that is state of the art at delivery is competing with a better one three years later, and the useful life assumed in the model determines how much of the loan is covered if the borrower fails.
Part of the premium is the size. At 60 billion dollars this would be the largest financing ever raised for chips, and a deal that big has to be distributed to insurers, pension money and credit funds that each limit how much they will hold in one name.
Part of the premium is the counterparty. Anthropic is a private company whose revenue, while growing quickly, is a fraction of what it is committing to spend, and the leases behind the borrowing convert a flexible cost into a fixed one for years.
The involvement of Google changes how the deal reads. The chips are TPUs, made by a company that is also an investor in Anthropic and a cloud provider to it, and when a supplier helps finance a customer’s purchase the arrangement is called vendor financing. Telecom equipment makers lent to carriers in the late 1990s, booked the revenue, and then watched the loans sour when the carriers could not fill their networks.
The defense is that demand for computing has exceeded supply, that Anthropic’s contracts are with paying customers rather than speculative builders, and that the chips have alternative uses. The counterargument is that those uses are worth less when the borrower is in trouble, because everyone else is selling the same equipment at the same moment.
Private credit is where this kind of paper now goes. Banks face capital charges that make a 60 billion dollar exposure to one private company unattractive, while credit funds backed by insurance premiums can hold it. The trade-off is transparency: private loans are marked by the lender, so deterioration shows up in a valuation committee rather than in a price on a screen.
For Anthropic the calculus is straightforward. Compute is the input that determines what models it can train and how many queries it can serve. Renting capacity from a cloud provider preserves flexibility and costs more per unit; owning the equipment through a financing vehicle lowers the long-run cost and locks in supply, provided the company can carry the fixed charge.
For the market the deal does something else. It creates a reference price. The next company that wants to borrow against accelerators, whether a neocloud, a developer or a consortium building a campus, will be asked why its terms differ from Polaris. A benchmark set at 10 percent makes every subsequent deal more expensive than it would have been at 8, and it tells builders what the marginal lender thinks their business is worth.
There is a wrinkle in how such debt is accounted for. If the borrowing sits in a special-purpose entity that owns the chips and leases them back, the obligation may be disclosed as a purchase commitment rather than as debt on the balance sheet. Investors comparing Anthropic with a public company will have to read the notes rather than the headline figures.
The tenor is the other open question. Chips are usually depreciated over five to six years, and a loan of this size is unlikely to amortize that quickly. The gap is bridged by assuming residual value, refinancing at the end, or by extending the asset’s life in the model. Each of those is a judgment, and the lender’s judgment is embedded in the coupon.
What the pricing does not settle is whether the underlying bet is right. If inference demand keeps compounding, 10 percent is a bargain for the borrower and a good trade for the lender. If it plateaus, the equipment is worth scrap and the leases remain. The people buying the subordinated tranche are being paid three points above a BB bond to take the second loss on that question.
Analysts said the deal will be watched less for its size than for its covenants: how much cash the borrower must hold, when financial tests trigger, and what happens if a chip generation is delayed. Those clauses, not the coupon, will decide whether the structure survives a bad year.
For now the arrangement has done what Anthropic needed. It has converted a capital number it could not fund from operating cash flow into a yield someone else would accept, without a public rating exercise. The bill arrives annually, in interest, for as long as the chips are in the racks.
