Anthropic raises $60 billion in debt to rent Google chips: banks put the AI appetite to the test

By Zendoric · October 6, 2026.

Bank of America, Citigroup and Morgan Stanley have begun syndicating —the process by which a bank spreads a loan among other institutions to share the risk— a $60 billion debt package to finance Anthropic's leasing of Google chips, as revealed by the Financial Times. It is, according to the British newspaper's own reconstruction, the largest artificial intelligence chip financing deal done to date.

The money does not buy the chips: it pays for the lease of the TPUs (tensor processing units, the processor Google designs in-house to train and run inference on AI models, an alternative to Nvidia's GPUs) that Anthropic has on order for 2027. The lease payments will not begin until the hardware is deployed, which delays the cash pressure on Anthropic but does not eliminate the risk for whoever puts up the money now.

The structure reveals where the real risk lies. There is a $42 billion secured senior tranche backed by Broadcom's A- credit rating —Broadcom designs the TPUs alongside Google—, which could allow it to be sold on later in the private placement or investment-grade bond markets. And there is an $18 billion junior tranche without that backing, where Blackstone has already committed some $9 billion and is helping place the rest among other investors.

That junior tranche is what really exposes lenders to Anthropic's credit profile, a company that, like most generative AI labs, still burns more cash than it generates. According to the FT, the banks may wait to place it more widely until Anthropic completes the IPO that, per that report, it plans to carry out before the end of the year: a public offering would give lenders the audited accounts they currently lack. As an alternative route, Broadcom's investment terms allow Anthropic to issue up to $42 billion in convertible notes —debt the creditor can turn into shares— directly to Broadcom to cover the lease.

In general, the formula of financing AI infrastructure with structured debt instead of equity is nothing new: other big tech companies have already turned to private debt vehicles to pay for data centers without that liability showing up directly on their operating balance sheet. What is distinctive here is the scale —$60 billion for a chip lease contract alone— and the fact that the ultimate borrower is not an Alphabet or a Meta with tens of billions in annual profit, but a startup that still depends on funding rounds to survive. That is why the Financial Times describes the deal as a test: if Wall Street absorbs this risk without demanding disproportionate premiums, it confirms that debt markets are willing to finance AI's compute race at the same pace venture capital funds and the labs themselves have. If not, the cost of money will start to act as a brake before the technology itself does.

There is a second, more structural reading in the split between Google, Broadcom and Anthropic. That a custom silicon designer like Broadcom is willing to put its credit rating behind a cloud customer is a signal of how much it wants Google's TPUs to become a real alternative to Nvidia's GPUs, today dominant in training large models. The more deals like this close, the cheaper and more liquid it will be for any lab to finance compute outside the Nvidia ecosystem, and that pushes down the margins of the supplier that has so far set the price of frontier AI.

Our take: this deal is the financial flip side of the thesis we defend. In the short term, piling tens of billions of debt onto lease contracts for hardware that does not yet generate revenue is exactly the kind of leverage that, if AI demand slows or models take longer than expected to monetize, can turn into real financial strain —not unlike what the telecom sector went through when it overinvested in fiber in the late nineties, betting on demand that did arrive, but later. The risk is not borne by the AI labs: it is borne, ultimately, by the banks that place that junior tranche and the investors who hold that debt in their portfolios. In the long run, however, this kind of massive financing is precisely what makes it possible to build the compute capacity needed for frontier models to keep improving to the point of accelerating medical and scientific discoveries that today seem distant. The question is not whether that compute is worth financing, but whether the sector is building the financial architecture —with cross guarantees, well-segmented risk tranches and transparency ahead of an IPO— so that a stumble along the way does not drag the whole system down with it.