Private credit's $200bn AI book rests on utilization the borrower has no standard way to prove

A GPU-backed loan is secured on a machine and repaid by what the machine sells. The controls test the machine. Collateral is audited and liens are perfected ^[6]. The author, who says he co-founded an infrastructure technology company nine years ago ^[13], writes that he has never seen a clause requiring a borrower to show what a specific GPU did last month, or how much money that work brought in ^[7].

So diligence spend goes on asset counts and lien perfection, and the utilization line in the model stays an assertion.

CoreWeave's $2.3 billion borrowing against Nvidia H100s, led by Magnetar Capital and Blackstone, was the first time anyone pledged H100s as security ^[5]. Set against a book the column puts well beyond $200 billion ^[3], that founding deal is about 1.15% of the total ^[15]. The Bank for International Settlements expects $300 billion to $600 billion by 2030 ^[4], and the midpoint of that range, $450 billion, is 2.25 times today's figure ^[16].

The two utilization ranges in the piece bound the problem. It says you would expect at least 70% to 80% in the current boom ^[18], and that a borrower consistently running at 80% to 90% has no standard way to prove it ^[2]. A fleet reported at 90% and actually running at 70% delivers about 22% less compute than the lender priced ^[17]. The instrumentation is already there: every data center logs power draw, utilization, thermals and uptime nonstop ^[10].

Power lenders advance against megawatt hours, metered independently and settled through the grid, not against turbine blades ^[8]. The remedy the column proposes is credit agreements carrying standardized performance reporting, with delivered GPU output matched against counterparty contracts ^[9], and it says regulators could fast-track common standards as aviation and the power industry did ^[14]. Mid-market operators keep that detail already; they just do not hand it to a lender tied to specific GPUs and specific customer contracts ^[11]. "Lenders are left taking their word for it, which is a lot to ask in a soon-to-be $600 billion industry," the author wrote ^[12].

I would want contracted output per counterparty, not the utilization percentage, because a fleet at 95% on a contract priced below cash cost still misses a coupon, and the column's own remedy points there in asking for output matched against contracts ^[9]. The counter-thesis is that utilization is beside the point: if losses in GPU credit arrive through the resale value of the hardware, a monthly output report does nothing for recovery, and the lender's exposure was always the secondhand market. What would settle that is loss experience, and the column does not cite a default, loss or writedown ^[20]. The $200 billion figure also reaches the reader secondhand, as the column's citation of a February Bloomberg report ^[3].