- Key insight: Lenders fronting the money for massive arrays of AI-friendly semiconductors have no visibility into their actual output. That could change with better reporting standards modeled after the energy industry.
- What's at stake: Even if a borrower consistently ran its GPUs at an 80% to 90% rate, there's no standard way for it to prove that to lenders under the current setup.
- Supporting data: Per a February Bloomberg report, private credit lending to AI has reached well beyond $200 billion, and the Bank for International Settlements expects this amount to hit between $300 and $600 billion by 2030.
I've never been able to trust anything I didn't or couldn't personally verify. This, which many would consider a character flaw, has served me well throughout my career.
However, since I co-founded an infrastructure technology company nine years ago, I've been working with millions of machines I've never laid my eyes on. As each was merely a number on my screen, I had to learn soon enough the difference between a number that has already been checked and a number that has not, or can't be checked.
As I see it, part of the
CoreWeave started the whole trend three years ago when the company borrowed $2.3 billion against Nvidia H100 GPUs. Led by Magnetar Capital and Blackstone, the funding round was the first time anyone had pledged H100s as security. A little more than a year later,
But let's be clear about one thing: controls. These are usually in place and aren't sloppy, meaning that collateral is audited, liens are perfected, and people who are very good at drafting covenants get to draft them.
Yet these controls all raise the same questions: Are the GPUs actually there? Who owns them? And who else has a claim on them? What we have never seen is a clause that requires the borrower to provide cold hard evidence on what each specific GPU actually did last month. Or how much money that work actually brought in.
As per a February Bloomberg report, private credit lending to AI has reached well beyond
With this
So, what does this say to
Read more:
Bitcoin's Overton window of opportunity is closing Maybe Wells Fargo's unions need a blockchain Iran built its own financial system, and has created a cyberfront in the war Fed meetings are about to get really awkward
And we haven't even mentioned utilization. You'd believe this would stand at a minimum of 70% to 80% with the current AI boom and the high costs of LLMs that led to a firm burning
The worst thing about this is that even if a borrower consistently ran its GPUs at an 80% to 90% rate, there's no standard way for it to prove that to lenders under the current setup.
By saying all that, I'm not here to bash anyone taking part in the GPU lending market. I'm working closely with many mid-market data center teams. They are some of the most measured people in the business, and many track their facilities with the extraordinary level of detail that satisfies even the most demanding lender.
The key issue is that they can't hand that record to their lenders in a way that the lender can easily verify the data, tied to specific GPUs and specific customer contracts. Lenders are left taking their word for it, which is a lot to ask in a soon-to-be $600 billion industry. It leaves lenders with less visibility into the credit risk than they'd have anywhere else.
The problem is there, so what can operators do to break this status quo?
Market players in power finance worked this out decades ago. Instead of borrowing on turbine blades, lenders advance against megawatt hours — all metered independently and settled through the grid.
This formula is super easy to implement in the GPU market. Here, the GPU delivers the computing power to customers under contract, and that output is the exact thing that repays the loan. And we even have this measurement ready, as every data center logs power draw, utilization, thermals, as well as uptime nonstop.
Now credit agreements have to be structured in a way that this data reaches the lender with standardized performance reporting and delivered GPU output matched against counterparty contracts. Rather than leaving every lender to work this out deal by deal, regulators can fast-track this process with common reporting standards, like in aviation or the power industry.
Lenders that require verified data in underwriting and ongoing monitoring on GPU utilization, uptime, and actual compute delivered under customer contracts will see exactly how the collateral is performing and what cash it's throwing off. And they'll be the ones best positioned to lead this fast-growing market.
