The AI build-out has a money problem hiding in plain sight: the chips and data centres now cost more than even the most valuable companies can pay for out of pocket. This week brought the clearest answer yet to how that gets solved — and it looks less like a tech deal than a property one.

Apollo Global Management and Blackstone have lined up a private credit package of about $36 billion (reports put it in a $35–36bn range) to finance Anthropic's compute expansion. By most accounts, it is the largest private debt deal on record.

What the money actually buys

The debt isn't landing in Anthropic's account as a cash injection. It's being used to buy Google's custom AI chips — Tensor Processing Units, or TPUs — which Anthropic will then lease rather than own. The capacity is spread across five US data centres. In effect, the lenders fund the hardware, and Anthropic pays to use it over time.

Why three giants are tied into one deal

The structure is what makes it notable. Broadcom — which co-designs Google's TPUs — is backstopping payments on the largest tranches of the debt, and that brought the borrowing costs down. Google, for its part, is providing payment guarantees on the leased capacity at all five facilities. So a single financing now binds together Google (the chip designer and landlord), Broadcom (the chip partner and backstop), Anthropic (the tenant), and Wall Street's biggest private-credit firms (the lenders).

Why it matters beyond the headline number

For most of the software era, growth was funded with equity and the occasional venture round. Compute changes that. Frontier AI now needs the kind of capital that builds power stations and ports — and it's increasingly being raised the same way: structured private debt, secured against long-term lease commitments and de-risked by guarantees from the very companies that stand to benefit.

That has two implications. First, it deepens the interdependence at the very top of the industry — Google, Broadcom and Anthropic are now financially entangled, not merely commercial partners. Second, it shows the AI race is as much a contest of balance sheets and financing creativity as of model quality. Whoever can fund the most compute, on the best terms, gets to keep training.

It also sits alongside this week's other infrastructure moves — OpenAI's reported Ohio mega-campus, Meta's Indian data centre with Reliance, TensorWave's AMD-backed raise. The common thread: 2026's AI story isn't only about who has the best model. It's about who can pay to run it.

One caveat worth keeping: the headline figure has been reported in a $35–36bn range, and some specifics of the final terms come from sources close to the deal rather than official disclosure. The shape, though, is clear — and it's a template others will copy.