Nvidia is restructuring how it finances the buildout of AI data centers: instead of customers buying B200 chips outright, Nvidia and partners including Goldman Sachs and BlackRock are structuring deals where the GPUs themselves back the debt, the same way an aircraft backs a leveraged lease. The chips sit in a leasing vehicle, cloud operators pay to rent capacity, and the lenders get repaid from that rental stream. It's a bet that a five-year-old B200 will still be worth renting in 2031, which matters because the last generation of AI chips didn't hold value that long. Nvidia's own H100, launched 2022, is already being discounted by resellers as newer Blackwell-generation hardware displaces it in top-tier clusters.

The reason this matters beyond Nvidia's balance sheet: asset-backed lending only works if the collateral has a resale market, and GPU clusters don't have one the way office buildings or airplanes do. A defaulted data center in Suzhou or Hsinchu can't easily sell its Blackwell racks to a buyer who wants last-generation compute, because the entire industry is racing to the next chip generation on a roughly two-year cycle. If lease payments from cloud tenants slow (Poolside's $12 billion reverse-acquihire into Nvidia last week is itself a sign that even well-funded AI labs are running short on independent cash), the test of this structure comes at the first missed payment, not before.