By Viktor Yoshimura – 23th October 2026
Billions are lent against GPUs on the assumption they hold their value. Few have priced what it takes to get them back.
The AI build has a credit problem, and most of the argument is about one number: how long a GPU keeps its value. That’s the starting point, not the answer. What a lender can actually recover also depends on whether it can take possession, how quickly the hardware can be removed and sanitised, what it costs to test and remarket, and whether a buyer can legally acquire it. A depreciation curve captures none of that.
The stress showed up in the bond market in mid-October. Of the $80 billion in data centre bonds issued in 2026 to date, $37 billion is rated below investment grade, according to Forbes on 16 October; Barclays put 2025 at $8.6 billion of $36 billion. Bonds financing data centres leased to one large neocloud averaged a yield of about 11 percent that week, against about 7 percent for BB+ bonds. That’s the price of risk on the buildings. A growing share of the risk sits on what’s inside them.
Four values, one of which matters in a default
GPU hardware carries at least four different values, and the debate tends to blur them. Accounting value is what sits on the owner’s books. Market value is what a willing buyer pays in an orderly sale. Collateral value is what a lender is prepared to recognise when it writes the loan. Recovery value is what’s left after default, enforcement, delay and a distressed sale. This piece is about the last one.
GPUs have been financed this way for a while. In August 2023 a neocloud took a $2.3 billion debt facility led by Magnetar and Blackstone, collateralised by Nvidia chips including H100s, an early example of the structure. In August this year Nvidia announced partnerships with six financial firms, including Blackstone, KKR and Goldman Sachs, to mobilise more than $500 billion of third-party capital for AI infrastructure over time, with Nvidia able to backstop up to $125 billion. That’s a financing platform, not $500 billion already lent against GPUs, but it shows where the industry wants the money to come from.
Market value has moved faster than the financing. Industry estimates put H100 rental rates at roughly $7 to $10 an hour at launch and $2 to $4 by late 2025, varying by provider, term and configuration. Indicative prices for used eight-GPU H100 servers were around $150,000 to $180,000 in early 2026, which is an asking price, not a liquidation price. Published depreciation schedules among listed neoclouds run from four to six years, and hyperscalers use five to six. Those are accounting lives. The economic life is whatever the secondary market says it is on the day you sell.
Lenders model the value on a spreadsheet. They’ll recover it on a loading dock.
What recovery actually involves
A forced recovery of a GPU fleet isn’t a sale. It’s a decommissioning project run under time pressure, and each step can take something off the number the lender modelled.
| Step | What it involves | What can cut the recovery |
|---|---|---|
| Access | Getting into a colo or campus whose lease and power bill may also be in arrears | Landlord claims, unpaid power, access disputes; rights vary by jurisdiction and contract |
| De-installation | Where racks are liquid-cooled, draining, purging and capping first; a GB200 NVL72 compute rack weighs about 1.36 tonnes | Crews, lifts and procedures many ITADs don’t have yet |
| Data | Customer data, credentials and potentially sensitive model artefacts across the fleet | Serious buyers want sanitisation evidence; some media or contracts will require destruction instead |
| Test and grade | Burn-in, firmware, health checks; sorting working GPUs, complete systems, repairable parts and scrap | Time per thousand units; each category has a different buyer |
| Legal routing | Classification of the specific equipment, destination and end user under export controls and transit rules | A narrower pool of buyers allowed to take it |
| Sale | A secondary market absorbing thousands of units at once | Other distressed sellers and hyperscaler rotations selling into the same market |
Table 1. Basis: author’s analysis of decommissioning steps; rack weight from published GB200 NVL72 specifications.
The collateral unit matters too. A GPU, an eight-GPU server and an integrated rack aren’t interchangeable, and the buyer for one may not be the buyer for another. Recovery planning has to establish what the lender actually holds security over, who can lawfully access and remove it, and whether it can be sold intact or has to be tested, reconfigured or broken down into components.
Why Asia-Pacific makes it harder
This region adds a layer the US conversation mostly skips. On 14 July 2025, Malaysia made exports, transshipment and transit of high-performance US-origin AI chips subject to a Strategic Trade Permit, with a 30-day advance notification applying under its catch-all conditions. In Singapore, prosecutors charged another individual in April this year in a case involving alleged false representations about the end user of servers later sent to Malaysia; authorities had said the servers might contain Nvidia chips. As I set out in the paper trail piece, enforcement here increasingly runs through the documents.
So the markets building the most new AI capacity in Southeast Asia, Johor and Singapore among them, are also where moving recovered accelerators gets the closest look. Before a recovered unit is sold or shipped, someone has to establish its classification, any licence it needs, its destination and its end user. That narrows the buyer pool, slows the sale and pushes value toward whoever can process and remarket in-market with clean records.
What a lender should ask now
My inference, not anything a lender has reported: if a large forced sale coincides with a wave of hyperscaler fleet rotation, recovery capacity could matter more to the realised haircut than the depreciation schedule. How much more depends on the hardware, the security package and the market on the day.
The useful underwriting question is net recovery under a stressed scenario, not a single resale price. That means asking a recovery provider, while the loan is still performing, to put numbers against each line below.
| Variable | What to ask for |
|---|---|
| Gross market value | Recent comparable sales by model, configuration and condition |
| Access and removal | Time and cost to gain access, disconnect, de-install and transport, including liquid-cooled racks |
| Sanitisation | Per-unit cost, turnaround and the evidence a buyer will accept |
| Testing and grading | Expected pass rate, cost, and what happens to failed units |
| Legal and logistics | Classification, destination and end-user checks, permits where they apply, freight and insurance |
| Sale execution | Realistic time to sell, buyer capacity, and the discount for selling in bulk |
| Net recovery | Proceeds less every cost above |
Table 2. The inputs to a stressed recovery estimate. Basis: author’s analysis.
If you lend against GPUs in this region: do you know who would get your collateral out of the rack, how long it would take, and who would be allowed to buy it?
Bond, financing and price figures are as reported by the sources below, as of mid-October 2026. The view that recovery capacity could outweigh depreciation in a stressed sale is the author’s analysis.
Sources
Forbes, “Nearly 50% of 2026 data center bonds are junk bonds”, 16 October 2026 (2025 comparison per Barclays; yields as reported).
Blackstone, CoreWeave $2.3bn debt facility announcement, August 2023; Reuters on collateralisation by Nvidia chips including H100s.
NVIDIA Investor Relations, financing partnerships with six financial institutions, 10 August 2026; Capacity Media coverage.
Hashrate Index, used GPU market pricing and depreciation, 8 April 2026 (rental estimates, indicative used prices, published schedules).
MITI Malaysia press statement, 14 July 2025: Strategic Trade Permit for US-origin high-performance AI chips; catch-all 30-day notification.
Malay Mail, Singapore AI chip fraud charges, 2 April 2026.
NVIDIA GB200 NVL72 specifications (compute rack about 1.36 t).