$22 Billion, Secured by the Chips

Ten banks have lined up a $22 billion loan to finance Crux AI's purchase of Google tensor processing units, Bloomberg reported, in the largest chip-backed financing of the current buildout. The debt is secured by the chips and by Crux AI's customer contracts - not by the broader balance sheets of Alphabet or Blackstone, the two firms behind the cloud venture.

The named lenders include Goldman Sachs, SMBC, Barclays, BNP Paribas and Scotiabank. What makes the deal notable is not the size, which is now routine, but the structure. The borrowers are not pledging corporate credit. They are pledging the hardware, and the revenue agreements that turn the hardware into cash.

That is a lender accepting a depreciating asset as primary security. A TPU has a useful life measured in a few years, a resale market that depends on which generation it is, and a value that falls the moment the next generation ships. The contracts are what make the loan bankable - they convert an aging chip into a schedule of payments.

The buildout finances itself

This is the third structural pattern in the same story. Nvidia took up to $10 billion of a position in Anthropic's expected IPO, buying equity in the customer that buys its chips. Nscale raised against compute it had not delivered. Now a cloud venture has borrowed against the silicon itself, with a syndicate of ten banks accepting the arrangement.

Each step moves the financing further from a conventional credit judgment. A bank lending against a data center weighs the operator's credit and the site's value. A bank lending against the chips weighs a hardware generation and a contract term, and takes the position that the two together will outlast the debt.

One detail from the coverage is worth sitting with: the venture was announced in May and launched eight days before the financing was reported. Ten institutions arranged twenty-two billion dollars of non-recourse debt against an entity that young, on the strength of the hardware and the contracts it had already signed.

What it says about the asset

Chip-backed lending works only while the chips are scarce and the contracts are long. Both conditions currently hold. Compute is constrained, HBM supply is tight, and hyperscalers are signing multi-year capacity agreements that look, from a lender's desk, like the predictable cash flow of a utility.

The risk is that the collateral and the business depreciate on the same schedule. If a generation of accelerators loses value quickly, the loans secured against them reprice, and the entities holding them are the ones that bought at the top of the cycle with borrowed money. Ten banks have now placed that bet against a venture launched eight days earlier.

The Collateral Clock

Chip-backed debt has a maturity problem that data-center debt does not. A building holds or appreciates; an accelerator generation is worth what the next generation leaves it worth. Lenders have priced that by pairing the hardware with contracts long enough to cover the loan, which works if the contracts outlast the chips and fails if they do not. The ten banks behind this facility have made a judgment about which of those two clocks runs out first, on an entity that had been operating for roughly a week when the financing was reported. It is the most direct statement yet of how the buildout is being underwritten: not on the strength of the operator, and not on the strength of the sponsors, but on silicon and signatures.