ft.com web signal

Google Wires $200B Wall Street Machine to Fund Anthropic Chips

TL;DR

  • An FT investigation puts Google's interlocking contracts to bankroll Anthropic's compute at roughly $200 billion, with about $150 billion tied to AI chips.
  • Apollo and Blackstone anchored a $35 billion SPV that bought about one gigawatt and one million TPUs and leased them to Anthropic.
  • Google is backstopping about 2.4 gigawatts of data-centre capacity built by former Bitcoin miners TeraWulf, Cipher and Hut 8.

The most surprising thing in the FT's investigation into how Anthropic is actually paying for its compute is that the answer isn't really Anthropic. Google, Broadcom, Apollo, Blackstone and Morgan Stanley have stitched together roughly $200 billion in interconnected contracts to underwrite the buildout, with about $150 billion of that tied to AI chips headed for Anthropic. Google guarantees the data centres, Broadcom commits to buying and financing the chips, and Apollo and Blackstone put up the private credit that buys the hardware before leasing it back.

The reason the structure has to look like this, per the reporting, is that Anthropic does not possess a credit rating. So the first transaction was arranged as a special-purpose vehicle that bought roughly $35 billion of hardware, about one gigawatt and one million TPUs, and leased them to Anthropic, with the lease payments servicing the debt. Broadcom's own filings show $128 billion in purchase commitments sitting behind the pipeline. On top of that, Google is backstopping lease and power obligations at data centres built by former Bitcoin miners TeraWulf, Cipher and Hut 8, covering about 2.4 gigawatts of capacity. Morgan Stanley has been the packaging shop, including a $15 billion financing under discussion for a Texas campus and a $3.2 billion construction bond it took to market in October.

Why any of this matters if you don't underwrite AI hardware for a living: the AI capex cycle stops being a story about hyperscaler capex and becomes a Wall Street credit story. Once the counterparties are private-credit funds, chip vendors and investment banks, a slowdown in Anthropic's revenue ramp does not just dent Google's cloud backlog. It ripples through leveraged loan books, residual-value guarantees on chips whose secondary market has never really been tested, and data-centre developers that repriced themselves as AI landlords. Jefferies' Jonathan Petersen put the macro risk plainly to the FT: if their appetite to invest decreases, all of it sees a slowdown.

The honest caveat is that the $200 billion is a directional sum of a lot of moving contracts, some of them still being syndicated, and the reporting does not spell out how Google's data-centre guarantees are treated on its own balance sheet, or what residual-value assumptions Broadcom is using when it guarantees senior tranches. Take the specifics as reported, not settled. The forward-looking part worth watching is whether the same template, Google and Broadcom taking the risk that a model lab cannot carry on its own, gets copied for OpenAI, xAI or the next lab that needs multi-gigawatt compute without an investment-grade rating. If it does, the AI infrastructure boom is already less a tech story than an infrastructure-finance one.