On September 23, 2026, USD.AI announced a $128.9 million asset-backed financing facility, the largest loan originated through its platform to date, according to its press release on PR Newswire. The money buys computing: 32 NVIDIA GB200 NVL72 rack systems going into production in British Columbia, Canada, for an undisclosed publicly-listed GPU cloud provider, contracted under a multi-year agreement with what USD.AI calls a blue-chip investment-grade counterparty.
The dollar figure is the headline. The chip count is the story. Divide 32 racks by the 72 Blackwell GPUs NVIDIA wires into each NVL72 domain and you get 2,304 GPUs, precisely the number of NVIDIA B300 GPUs financed by USD.AI's previous record facility, a $98.1 million deal announced in June 2026 and covered alongside this one by Unite.AI. Two facilities, three months apart, identical silicon counts, 31 percent more debt. That coincidence, if it is one, is the clearest window yet into how AI compute is being repriced: not per flop, not per token, but per rack, as a collateral class.
The deal, as announced
The disclosed facts are thin by design, because the borrower and the counterparty are unnamed. What USD.AI did publish:
| Item | September 2026 facility | June 2026 facility |
|---|---|---|
| Facility size | $128.9 million | $98.1 million |
| Hardware financed | 32 GB200 NVL72 racks (2,304 Blackwell GPUs) | 2,304 B300 GPUs |
| Term | Not disclosed | 3 years |
| Location | British Columbia, Canada | Not disclosed |
| Borrower | Publicly-listed GPU cloud provider, undisclosed | Duos Edge AI (Duos Technologies Group subsidiary), operated via Hydra Host's Brokkr platform |
| Offtake | Multi-year, blue-chip investment-grade counterparty | GPU and offtake contracts pledged |
| Structure | Non-recourse, secured by GPU infrastructure only | Non-recourse, non-dilutive, off-balance-sheet |
Two of the borrower's attributes are doing most of the underwriting work here. The operator runs large-scale NVIDIA clusters as a managed service with deployments across North America, South America, and Europe. The revenue side is a named-but-unnamed investment-grade counterparty on a multi-year contract. In credit terms, this is project finance wearing a crypto-native label: the loan is serviced by an offtake contract, and the GPUs are the asset you seize if the contract fails.
The compute middle market thesis
USD.AI is a lending platform built by Permian Labs, and the transaction is a live exhibit for a thesis its co-founder and COO Conor Moore laid out in a July 2026 essay that Unite.AI summarizes: the "compute middle market," which Moore defines as AI infrastructure financings between $50 million and $250 million. His claim, and it is his claim, not an audited figure: over an eight-month span USD.AI saw $13 billion in deal flow across more than 200 neoclouds in that band, in a segment he estimates at $2 trillion of spending through 2030.
The structural protections Moore describes are worth reading like a software engineer reads a threat model, because they are one:
- A bankruptcy-remote special purpose vehicle with an independent manager sits between lender and sponsor.
- Customer contracts are assigned to the vehicle, and its bank accounts sit under a deposit account control agreement, so the lender controls cash before the borrower can move it.
- A first-priority charge over the GPUs plus a pledge of the vehicle's equity, perfected through a Uniform Commercial Code filing.
- Step-in rights to the colocation facility, including the right to sell the servers on default.
- A debt-service reserve account prefunded with one to three months of principal and interest.
Sponsors typically put 20 to 30 percent of the cost down in cash, Moore wrote, and he argues the resulting packages carry net coupons 400 to 600 basis points above investment-grade-backstopped deals while maintaining comparable structural protections. USD.AI's platform-level numbers, self-reported on its website: $609 million in total value locked, a $183 million active loan pipeline, more than 80 partnerships. Its capital side leans on stablecoin rails: a $100 million facility from the exchange platform Bullish announced in August 2026 and a $40 million revolving facility from K3 Capital, both listed in USD.AI's own release history.
The arithmetic of a rack-scale loan
Here is where the data scientist's pencil earns its keep. Every assumption below is labeled; the only inputs are the published deal sizes and NVIDIA's published hardware configuration.
Assumption 1: 32 GB200 NVL72 racks contain 32 by 72 = 2,304 Blackwell GPUs. NVIDIA's GB200 NVL72 specification page confirms each rack is a single 72-GPU NVLink domain pairing 36 Grace CPUs with 72 Blackwell GPUs, 13.4 TB of HBM3E at 576 TB/s, 130 TB/s of aggregate NVLink bandwidth, 2,592 Arm Neoverse V2 cores, and 17 TB of LPDDR5X.
Assumption 2: the facility finances the great majority of a deployment's cost, with the sponsor's 20 to 30 percent cash equity (Moore's stated range) on top.
The debt per GPU then works out cleanly:
| Metric | June facility (B300) | September facility (GB200) | Delta |
|---|---|---|---|
| Debt | $98.1M | $128.9M | +31.4% |
| GPUs financed | 2,304 | 2,304 | 0 |
| Debt per GPU | $42.6K | $56.0K | +$13.4K |
| Debt per rack (72 GPUs) | n/a | $4.03M | |
| Implied all-in cost per rack at 70 to 80% leverage | n/a | $5.03M to $5.75M | assumption-dependent |
The per-rack debt of roughly $4 million is not a hardware sticker price; a GB200 NVL72 deployment is rack, networking, liquid cooling, power delivery, and colo build-out. The interesting number is the leverage-adjusted band: at Moore's stated 20 to 30 percent equity, the implied all-in cost of this deployment is $161 million to $184 million for 2,304 Blackwell GPUs, call it $5 to $5.75 million per rack of installed, contracted, income-producing compute.
Now the engineering observation the table sets up. Both deals financed exactly 2,304 accelerators. That is a suspiciously round coincidence at first, but it is actually the unit economics of this lending class made visible. A lender writing non-recourse paper against rack-scale systems is underwriting in whole racks, and a 2,304-GPU block, exactly 32 NVL72 racks, looks like a standard ticket size: big enough to service a hyperscaler-scale offtake, small enough to fit inside a $100 to $150 million facility. When the same chip count reappears across two facilities three months apart, the most plausible reading is a template, and templates are what credit markets build before they build risk models.
What the collateral actually is
The uncomfortable question, from both sides of the ledger: what does the lender hold if the counterparty walks?
Not the silicon. This is the part the fintech framing obscures. The financed generation is already two hops from the front edge. NVIDIA's own GB200 NVL72 page now sits one section above the GB300 NVL72, and as we covered in our MLPerf v6.1 analysis, Vera Rubin NVL72 systems are already topping inference benchmarks at 3.7x the efficiency of the best Blackwell-era result. A five-year offtake contract against GB200 racks means the collateral depreciates along a capability curve that resets roughly annually. Step-in rights let the lender sell the servers on default, but selling last-generation rack-scale systems into a market where the next generation posts 3.7x results is a residual-value haircut the structure paper cannot draft away. This is the same depreciation treadmill we traced from silicon to tokens in the AI cost chain; the finance layer simply converts it into basis points. Moore's 400 to 600 basis point premium over investment-grade-backstopped deals is, on this reading, priced obsolescence, not just illiquidity.
What the lender actually holds is the contract. The deposit account control agreement, the assigned customer agreements, and the investment-grade credit behind them are the collateral; the GPUs are a liquidation backstop whose value trends toward scrap-plus-second-life on a two-to-three year clock. The engineering reason the GPU count matters more than the dollar count is that it tells you the underwriting template: same chip count, more dollars, because GB200's rack-scale integration, liquid cooling, and NVLink domain density pack more contracted revenue per financed unit than a tray of B300s. Higher capital intensity per rack is exactly why this market needs the middle-market lenders: the GPU value comparison we published shows how fast per-rack capital cost has climbed across generations, and a $5 million rack cannot be funded on a corporate credit card.
For a builder renting this compute, the read is practical rather than alarmed. Neocloud capacity funded this way arrives with the price of capital baked into hourly rates: a coupon 400 to 600 basis points above what a hyperscaler pays, per Moore's own framing, recovered from tenants. Discount pricing from a neocloud is therefore either older hardware, shorter contracts, or a sponsor subsidizing customer acquisition. All three are fine to buy from; they are just different products than the rack spec sheet advertises.
Outlook: what would prove the template
The compute middle market is either a genuine credit class, in which case it will grow up in public, or it is 200 neoclouds levered against contracts whose counterparties can renegotiate the moment newer silicon undercuts them. Three falsifiable signals separate the readings:
- Bank entry. If a regulated commercial bank prices a GB300 or Vera Rubin rack facility within 12 months at a spread under 400 basis points over USD.AI-style comps, traditional credit has validated the asset class and spreads will compress. Watch for UCC filings naming colo operators, not press releases.
- The first disclosed default. No structure of this kind has been tested end to end in the GPU era. A step-in event where a lender actually resells a 2,304-GPU block, and the realized price, will set the residual-value assumptions every subsequent deal uses.
- Ticket-size drift. If the next record facility finances 4,608 GPUs at roughly double the dollars, the per-rack template thesis holds. If it finances a wildly different chip count at a similar size, the "middle market" is being sized by available capital, not by underwriting units, which is a weaker foundation.
USD.AI's own pipeline, a self-reported $183 million, is barely more than one deal wide. The segment's biggest claim, $2 trillion of spending through 2030, belongs to one essay by one party to every transaction. But the arithmetic in this deal is real and public: two identical chip counts, 31 percent more money, and a security package borrowed wholesale from project finance. AI's infrastructure build-out now has its own shadow banking layer, and it is starting to behave, for better and worse, like every collateral market before it.