The GPU Mortgage: How USD.ai Finances AI's $10 Trillion Build-Out

Financing as the bottleneck

David Choi argues that AI's demand problem comes down to financing. "There's not enough compute in the world," he says, adding that operators who want to scale "are buying it with all cash." He compares this to housing: if everyone had to pay cash for a home, he says, the industry "wouldn't really exist."

He says demand comes from the $10 trillion that will be spent on compute hardware over the next five years. By his account, 70% of a data center's cost goes to the GPU and 30% to the shell and land, the parts that can be securitized.

How the products work

USD.ai issues USDai, a US dollar stablecoin overcollateralized by PayPal's PYUSD, and sUSDai, a yield-bearing version whose return is backed by compute revenue. Choi describes sUSDai as a perpetual instrument, much like perpetual futures, which lets things "be continuous."

Choi says a loan takes under one week to arrive, and that USD.ai is the only lender in the market that will consistently go down to $300,000. Its largest single deal is a facility of up to $500 million for Sharon AI to expand GPU capacity across Australia and Asia-Pacific. Every month, he says, "we've had our largest loan ever done."

A bet against banks

Choi predicts that banks "will never catch up" to blockchain-based GPU lending, which he calls "an infinite machine." He rejects the idea that GPU lending is a niche form of equipment financing and calls it "a very large category."