A-TIER resource
CoreWeave debt explained — how GPU-backed loans fund the neoclouds
100k ft — What is this?
CoreWeave debt explained — how GPU-backed loans fund the neoclouds
50k ft — Why care?
An ex-Morgan Stanley CDO structurer explains the neocloud stack (landlords, vertically integrated players like IREN and Crusoe, CoreWeave, Nebius) and why CoreWeave's GPU-backed asset financing spooked markets. Advance rates hit 90-95% against off-taker credit, and loan, contract and GPU useful life are nearly the same length, leaving a thin tail.
10k ft — Key takeaways
• GPU loans have a short tail because asset life barely exceeds the contract
• Neoclouds can borrow 90-95% of capex against strong off-takers like Meta
• Post-contract hourly GPU rental is the bull case for the debt
5k ft — Tools & techniques
Anthropic, OpenAI
asset-backed financing, project finance
1k ft — Deep-dive / try it
In case you missed it, there’s been turmoil in AI stocks, sparked by concern over Coreweave’s debt. Here’s exactly what you need to know about the neoclouds.
A few things I couldn’t fit in the video time limit of 3 min:
1. The advance rate, meaning how much neoclouds can borrow relative to the amount of money they need to spend on Capex (to buy chips and equipment) can get QUITE high, in some cases 90-95%+ depending on the credit of the “off-taker”, so from Meta for example.
2. Asset backed financing and project financing (one of the groups I worked in at Morgan Stanley) isn’t new. What is unique about these GPU backed financings are the much shorter contract length; in the power world you often see power purchase agreements of 15-20 years and assets that have a much longer useful lif…
Transcript excerpt
we need to talk about what's going on with CoreWeave. It plunged 15% and is taking a bunch of the other AI stocks down with it. I happened to work in CDO structuring at Morgan Stanley in 2008. I'm going to explain exactly what's going on in under three minutes. Let's begin with where we are within the AI ecosystem. We're in layer three, the infrastructure, which at its core can be thought of as the companies who sell compute. Now compute is the raw processing power needed to train AI models. And Wall Street splits those companies into two categories. First there are the hyperscalers, household names like Amazon's AWS or Microsoft's Azure. And second is the neoclouds, where the poster children are CoreWeave and Nebius. To understand the neocloud business, let's run through what NVIDIA CEO Jensen Wong calls an AI factory. First you need energy, second you need chips, and third you need the building or powered shell. Many of these neoclouds started as crypto miners, but pivoted their business when they realized AI was a better opportunity. The output of the AI factory is compute, but companies make money in different ways. Applied digital and core scientific are landlords. They own the powered shells and make money renting the real estate without taking on the risk of buying the chips. Of the companies that sell compute, vertically integrated players like Iran own their own powered shells. Crusoe got there from the other direction. It started with power generation and then…