GPU Impairment: When Do You Write Down a Node Fleet?
The short version
Impairment and fair value answer different questions. Impairment asks whether the book is too high. Fair value asks what a buyer pays today. A fleet can pass the first test and miss the second by a wide margin. Lenders live on the second.
Four things force a write-down conversation: secondary prices fall, a new generation ships, a contract rolls off, utilization slips. All four are live now. NVIDIA's latest 10-Q says Vera Rubin has begun production shipments. B300 nodes ask 1.68x an H100 node. H100 rental rates sit roughly two-thirds below their 2023 peak (estimate).
The triggers
| Trigger | Our screen | Reading, Sep 2026 | Why it moves value |
|---|---|---|---|
| Secondary price decline | Comps or asks down 10%+ from last mark, or ask-to-executed spread widening | Refurbished H100 median ask $246K, 76% of the $324K new ask (Hashrate Index, Aug 31 print) | Comps set fair value directly. Asks lead trades. |
| New-generation launch | Next generation reaches volume shipments | Vera Rubin in production shipments (NVIDIA 10-Q). B300 median ask $544K vs H100 $324K | Buyers who wanted the frontier stop bidding on the old chip. |
| Rental rate decline | Index down 10%+ over 90 days | Silicon Data, Sep 28: H100 $2.72, H200 $3.34, B200 $5.78, B300 $6.78 per GPU-hour | Rate is the income approach's biggest input. |
| Contract rolloff | Contract ends within 12 months, no signed renewal | About $6,300 per node-year for every $0.10/GPU-hr (estimate, 90% utilization) | Contracted cash flow becomes market cash flow. |
| Utilization drop | Trailing utilization under 85% for two quarters | Debt-financed clusters break even near 70%, per third-party analysis | Idle nodes carry depreciation and interest, no revenue. |
| Credit event | Covenant breach, missed payment, customer default | Borrower-specific | Value shifts from fair value to forced-sale value. |
The 10% screens and the two-quarter window are our working defaults, not accounting rules. Set yours in the credit agreement.
New generation is the trigger that matters most. It doesn't wait for your reporting date. B300 rents at $6.78 against $2.72 for H100, about 2.5x (estimate from the Sep 28 Silicon Data readings), on a node price 1.68x higher. That gap is what reprices Hopper. It lands when buyers can get the new chip in volume, not on launch day. I covered how that played out for the H100 in What an H100 Is Worth Today. B300 is now on that clock, and NVIDIA's cadence is annual.
Rolloff is the quiet one. A contracted fleet is valued off the contract until the day it isn't. The math is 8 GPUs × 8,760 hours × 90% utilization × $0.10, about $6,300 per node-year (estimate). A fleet that re-lets 40 cents lower loses roughly $25,000 per node per year before anyone touches the discount rate.
Utilization is the same math run the other way. At $2.72 an H100 node grosses about $172K a year at 90% utilization and about $133K at 70% (estimate, before power, colo and opex). That is a $38K swing per node with no change in rate.
Impairment is not fair value
Your borrower's auditor runs a different test than you do.
Under US GAAP, ASC 360 is two steps. A triggering event starts a recoverability test: undiscounted cash flows over the asset group's remaining life against carrying amount. Only if that fails do they measure fair value and write down to it. Undiscounted is the catch. A fleet with five years of contracted revenue can pass while a buyer would pay well under book.
IAS 36 is one step. Carrying amount against recoverable amount, which is the higher of fair value less costs of disposal and value in use. Strong contracts support value in use, so the book holds.
Either way, a clean impairment test tells you the borrower's books are supportable. It does not tell you what your collateral is worth.
Lenders answer to a different rule. Under CECL, the collateral-dependent expedient measures the allowance off collateral fair value, less costs to sell if repayment depends on a sale. For regulatory reporting, the interagency policy statement says collateral-dependent loans use collateral fair value whether or not foreclosure is probable.
Book lives vary, which changes when the borrower's test even trips. CoreWeave uses 6 years, Lambda 5 and Nebius 4, per a neocloud economics summary. IREN's 10-K carries $1.49B of pledged GPU hardware on 5-year straight-line. The same filing shows $638.8M of fiscal 2026 impairment, mostly bitcoin mining hardware and related equipment, written down to estimated fair value. Hardware whose best use has changed gets marked to what buyers pay.
How fair value is measured for a fleet
| Method | Question it answers | Inputs | Where it breaks | Use it for |
|---|---|---|---|---|
| Cost (depreciated) | What did we pay, less a schedule? | Invoice, useful life | Ignores the market entirely. There is no sticker, only a bid. | Book value and sanity check. Not fair value. |
| Comparables | What do similar nodes trade for? | Executed sales, brokered lots, discounted asks. Adjust for memory, cooling, CPU, networking, warranty, lot size | Asks are not trades. Thin data on newer chips. | Primary method for H100/A100-class with active resale. |
| Income (DCF) | What will it earn? | Rate curve, utilization, power and opex, discount rate, end-of-term resale | Small changes in rate or discount rate swing the answer. Resale value carries a lot of weight. | Contracted fleets and newer chips with few comps. |
We run comps and income side by side and reconcile them. When they diverge by more than about 15%, the report explains why.
Then we add a second number: orderly liquidation value. In our anonymized example, the same H100 equipment came out at $214K a node on an income approach and sold for $121K. Both are correct. They answer different questions.
What the gap looks like on one node
Take an HGX H100 node bought mid-2023 at $340K, the midpoint of the $300K–$380K range we cited in our H100 piece. Three years on:
| Measure (per node) | $ | % of cost |
|---|---|---|
| Book, 6-year straight-line | $170,000 | 50% |
| Book, 5-year straight-line | $136,000 | 40% |
| Refurbished median ask, Aug 31, 2026 | $246,000 | 72% |
| Fair market value, example appraisal (income) | $214,000 | 63% |
| Executed sale, same example | $121,000 | 36% |
On this fleet, book sits below fair value. No ASC 360 loss. But the executed sale came in $15K to $49K per node below book. That gap belongs to the lender.
Now put a loan against it. Assume an 80% advance, $272K a node (hypothetical):
| Loan balance per node | Share of advance | LTV vs fair value ($214K) | LTV vs executed sale ($121K) |
|---|---|---|---|
| $272,000 | 100% | 127% | 225% |
| $204,000 | 75% | 95% | 169% |
| $136,000 | 50% | 64% | 112% |
All LTVs are estimates from the example above. Paydown is linear and repricing comes in jumps. Even at half the advance repaid, this loan is underwater against the sale price.
What an impairment review looks like
- Inventory the fleet. Serials, configuration, memory, cooling, warranty status, location. Nodes with different specs are different collateral.
- Group by cash flow. One contract, one cluster, one asset group. Don't blend a contracted B200 block with idle H100s.
- Screen the triggers. Run the table above. Log which fired and when.
- Test recoverability. If it is the borrower's ASC 360 test, undiscounted cash flows against carrying value. If you are the lender, skip to fair value.
- Measure fair value and liquidation value. Comps and income, reconciled. Costs to sell explicit.
- Document and set the next date. Compare to carrying amount and to the loan. Write down why the number moved since the last mark.
How often lenders should re-mark
Value moves quarterly in this market. The H200 premium over H100 went from 0.4% to 16.2% in twelve weeks this summer (Silicon Data). An origination appraisal is a guess with a date on it.
| Collateral state | Desk re-mark (indices and comps) | Full appraisal |
|---|---|---|
| Contracted, LTV under 65% of fair value | Quarterly | Annual |
| Contracted, LTV 65–85%, or under 12 months of contract left | Monthly | Semi-annual |
| Off-contract, LTV over 85%, covenant breach or watchlist | Monthly | Within 30 days, with liquidation value |
Any state, re-mark within 30 days of:
- a next-generation volume-shipment announcement
- a 10%+ index move over 90 days
- a contract termination or a renewal at a lower rate
- a customer default
- a sale below book
Get a number that holds up
GPU Appraisals issues USPAP-compliant appraisals of GPU infrastructure, with fair market value and orderly liquidation value in one report. For a directional number, try the quick estimate. For a defensible one, book 20 minutes.
Disclosure: GPU Appraisals is affiliated with GPU Capital LLC, a private credit fund that invests in GPU servers. As USPAP requires, GPU Appraisals discloses any current or prospective interest in the subject property or the parties to the client before accepting an assignment.
Sources
- Hashrate Index, Introducing the AI Hardware Price Index (Aug 31, 2026)
- Silicon Data indices, Sep 28, 2026: H100, H200, B200, B300; blog
- NVIDIA Form 10-Q, quarter ended Jul 26, 2026
- IREN Form 10-K, FY2026
- Deloitte DART, ASC 360 recoverability
- Sofer Advisors, ASC 360 triggering events
- IFRS Foundation, IAS 36
- PwC Viewpoint, CECL collateral-dependent expedient
- Federal Reserve, Interagency Policy Statement on Allowances for Credit Losses
- Zettabyte, neocloud unit economics (useful lives)
- Moduledge, neocloud business model (utilization break-even, citing American Compute)
- GPU Appraisals, What an H100 Is Worth Today and example appraisal
Rental and resale figures are third-party market observations, not appraisals of any specific equipment. Figures marked "estimate" are our arithmetic from the cited inputs. This article is market commentary, not an appraisal, accounting advice or investment advice.
Marek Omilian | Managing Director | Value Prism Consulting