A $36 Billion Loan to Buy Anthropic Its Chips

Two of the biggest names in private credit are putting together a financing deal that says a lot about how the AI boom is now paid for. Apollo Global Management and Blackstone are arranging roughly 36 billion dollars of debt to buy Google's custom AI chips and lease them to Anthropic, according to Bloomberg. If it closes as planned, it would rank among the largest deals of its kind.

The structure is the interesting part. The borrowed money flows through a special-purpose vehicle that actually buys the hardware, Google's tensor processing units, and then leases that hardware to Anthropic for use in data centres in New York, Texas, Louisiana, and Indiana. Anthropic gets the compute it needs without parking tens of billions of dollars of chip debt on its own balance sheet. The lenders, in return, get a loan backed by physical assets rather than an unsecured bet on a company that does not yet turn a profit.

How the risk is sliced

Bloomberg's reporting puts the debt in three layers: about 6 billion dollars of senior A1 notes, 25 billion of A2 notes, and 4.5 billion of riskier B notes, with the figures still subject to change before closing. Sitting behind much of it is Broadcom, the company that helps Google build the TPUs in the first place. Broadcom is providing a residual-value guarantee on roughly 31 billion dollars of the senior debt. In plain terms, if Anthropic stops paying its lease and the used chips do not fetch enough on resale to cover the loan, Broadcom eats the difference.

That backstop is what makes the whole thing bankable. It turns a wager on one AI lab's future into something closer to an asset-backed loan, which is exactly the sort of paper big credit funds want to hold.

Anthropic's chip diet

The deal also underlines how far Anthropic has moved beyond a Nvidia-only setup. Google's TPUs have become a serious alternative supply of compute, and locking in that capacity through a lease lets the lab scale without raising the cash to own the silicon outright. It comes at a busy moment for the company, which recently filed confidentially for an IPO at a reported 965 billion dollar valuation, with a revenue run rate it has put near 47 billion.

None of the parties has commented on the record, and the terms could still shift as more investors are pulled in. The bigger signal is structural. The cost of frontier AI has grown so large that buying the chips now looks less like a capital expense and more like a financing problem for Wall Street to engineer. Watch whether this leasing-through-an-SPV model becomes the template for the next round of build-outs.