Anthropic wants $60 billion for chips. The lender wants 10%.

Training and running frontier models takes an enormous number of chips, and the bill is now big enough that AI labs are borrowing for it like airlines borrow for planes.

The latest example, first reported by the Financial Times on October 16 and citing people familiar with the matter: Anthropic is putting together a roughly $60 billion debt package, called Project North Star, to pay for about 1.5 gigawatts of Google's custom tensor processing units (TPUs). It would be the largest chip financing to date.

What's reported.

• About $42 billion of the package is senior debt guaranteed by Broadcom, which works with Google on the chips.
• Another $18 billion is junior debt, backed by Anthropic's five-year lease payments for the chips.
• Blackstone-managed funds are expected to take $9 billion of that junior piece, with a 10% coupon. After issuance discounts and fees, the total return would be higher than 10%.
• For comparison, borrowers with a similar credit rating pay around 7% in public bond markets, according to the reporting.
• Four months ago, Anthropic raised about $4.4 billion in a similar deal, Project Big Sky, at 8.5%.
• KBRA rated Anthropic one notch below investment grade in June, and the new junior debt is expected to get a similar rating. The debt could become tradable in the institutional bond market once it's drawn down in 2027.

Blackstone declined to comment and Anthropic didn't immediately respond, per Investing.com's write-up of the FT story. None of the terms have been publicly confirmed by either company, and the remaining $9 billion of junior debt is still to be offered to other investors.

Our take

The joke in our cartoon is the scale: "some chips" now means a power plant's worth of hardware. The serious part is the price. Going from 8.5% to more than 10% in four months says lenders want more cushion for funding AI compute, during a week when AI and chip stocks sold off on exactly that worry.

For people who just use these tools, the takeaway is simpler: compute is the biggest cost behind every AI subscription, and it's increasingly paid for with borrowed money. That's worth remembering the next time a plan's price or usage limits change.

What we did and didn't do: this story is based on the FT's reporting as relayed by Investing.com, Mint and Il Sole 24 Ore. MODEL hasn't seen the deal documents. Not investment advice.

MODEL Editors