Hello,
This episode follows the money to the backbone of the AI boom, where David Choi of USD.AI is building the interest rate of intelligence, which is a debt market for the GPUs that everything else now runs on.
Banks have spent a century learning to lend against assets like houses, cars, office buildings, and even paintings. But they don’t give you a loan against one of the most valuable assets in the world right now: a GPU.
That is the gap David Choi is building into. He comes to chip financing from a background in lending against hard-to-price assets, from artwork to exotic real estate, at Deutsche Bank. The company started with his personal frustration when he couldn’t get a loan against his 300-unit server of H100s.
Why does this gap exist? Most assets became financeable because someone figured out how to turn the debt into something tradeable. Before Fannie Mae, you would roll over a home loan until the bank asked you to repay it. The fix was to rate it, package it, and sell it on. The primary lender is not stuck. That packaging machinery, the asset-backed securities market, is the engine that built modern finance. GPUs do not have one.
There are good reasons for that. A chip depreciates fast. It has a useful life of around six years, with newer generations landing every eighteen months or so. The traditional securitisation process, raising a warehouse facility from a bank, writing thousands of loans, going on a roadshow, tranching it, and selling it down, can take up to three years. That timeline is fine for a thirty-year mortgage. The value chain doesn’t know how to do it for an asset that becomes obsolete in six years. So banks pass, and neoclouds end up at private credit funds paying 20-25% all-in.
USD.AI compresses that machine onto a blockchain. A real-world use of blockchains, yay! The moment a loan is created, it is instantly folded into a single tradable instrument: sUSD.AI. Think of it as wrapped staked ETH, but for data centres. Because the debt is liquid, a lender is far more willing to fund you because you are not holding them hostage for years. The bet is to become the Fannie Mae for GPUs.
Instead of lending against the borrower, USD.AI lends against GPU chips. And no money is released until the hardware is installed and verified. The borrower does not own the chips while the loan is live, so USD.AI can take them back. The borrower has to keep a few months of payments in reserve. If the chips must be sold for less than the loan amount, Munich Re covers the shortfall. It helps that US data centres are metal boxes with high security and insurance, making them about the safest place to park a physical asset you intend to lend against.
The conversation spans several components of GPU financing, such as the mechanics of repossession, why Nvidia refuses to finance its own chips, how a gigawatt of power compares to fifteen Las Vegas Spheres, the trouble with buybacks as a token ritual, and why GPU-backed debt is not a niche.
Thinking about what backs the boom,
Saurabh Deshpande
