In-situ vs ex-situ GPU value
What a GPU server earns installed against what it fetches removed and sold, per 8-GPU node.
Installed vs removed, per 8-GPU node
| Server | Income value in place | New price | Ex-situ FMV | Orderly liq. | Forced liq. | Parted out | Income / FMV |
|---|---|---|---|---|---|---|---|
| HGX H100 | $216K* | - | $195,000 | $173,000 | $126,000 | ~$188,000 | 1.11x |
| HGX H200 | $586K*† | $340K | $348,500 | $309,000 | $225,000 | - | 1.68x |
| HGX B200 | $646K† | $415K | No used sales yet. Ex-situ 59-80% of the price new; income value 156% | ||||
| HGX B300 | $779K† | $425K | No used sales yet. Ex-situ 78-89% of the price new; income value 183% | ||||
| HGX A100 8x 80GB | $63K* | - | $95,200 | Installed is worth less than sold | 0.66x | ||
*Limited sample. Income value in place: for Hopper and A100 servers, the Rack Report's rental-implied value (today's rent net of hosting, at 70% utilization over the remaining life to six years from launch, discounted at 12% a year; the Rack Report itself warns this approach will likely overestimate value); for B200 and B300, the present value of a 3-year contract at the managedgpus.com preset rate, net of colocation only, plus a conservative-band resale value at the end. †Above the cost new: the excess is the value of the rent or contract, not of the equipment. New price: the Rack Report new-price midpoint. Ex-situ values are gross, before de-installation, shipping and data sanitization; orderly and forced liquidation are 89% and 65% of fair market value. Parted out: the node value less the premium a whole node fetches over its parts. Indicative, not an appraisal.
What the table shows
In-situ value is what a GPU server earns installed where it sits; ex-situ value is what it fetches removed and sold. The in-situ figures here are income values, rent or contract included, so they can exceed what the equipment costs new. For an HGX H100 the gap is 1.11x this month: about $216K in place on the rent it earns (a limited sample), against $195K sold as a node and $126K in a forced sale. For an A100 it runs the other way, 0.66x.
Updated October 17, 2026. Rack Report October 2026, data through August 2026.
Definitions
| Premise | Definition | For GPU servers |
|---|---|---|
| In-situ: in continued use (FMV-CU) | Fair market value assuming the equipment stays installed and in use as part of an operating business, with the cost of installation and its contribution to the whole considered. | The case for a cluster that keeps running in place: same site, network fabric and offtaker. It drops away if the colocation contract or the offtake ends, which is when a lender needs the number. |
| Ex-situ: fair market value (removed, in exchange) (FMV) | The price at which the equipment would change hands between a willing buyer and a willing seller, neither under any compulsion to buy or sell, both with reasonable knowledge of the relevant facts, as of a stated date. | For GPUs the evidence is completed arm's-length sales, not dealer asks or rental rates. Asks for used H100 SXM GPUs ran 1.16x the realized price in the October 2026 Rack Report. |
| Ex-situ: orderly liquidation value (OLV) | The gross amount that could typically be realized from a liquidation sale given a reasonable time to find buyers, with the seller compelled to sell, as-is and where-is, as of a stated date. | A GPU resale usually takes about 30 days once remarketing materials, diagnostics and refurbishment are done; our residual report recommends 30 to 60 days, because a rushed sale fetches less. |
| Ex-situ: forced liquidation value (FLV) | The gross amount that could typically be realized from a properly advertised and conducted public auction, with the seller compelled to sell with a sense of immediacy, as-is and where-is. | There is no central auction house for data center GPUs. A forced sale runs through brokers and resellers in parallel, and parting out a cluster cuts the buyer pool to those who can rebuild it. |
Definitions paraphrase the American Society of Appraisers machinery and equipment definitions. In-situ and ex-situ are the shorthand for two of its premises: in continued use (installed) and in exchange (removed and sold). The ex-situ fair market values on this page are the Rack Report's listing-based Complete systems values, an indicative market level, not a median of completed sales.
Income value is not equipment value
Under the continued-use premise, equipment is normally worth no more than it costs to replace new and install. An income value above that, as for the HGX H200, HGX B200 and HGX B300 this month, includes the value of the rent or contract: it holds only while the customer keeps paying and the site keeps the servers running, and a buyer of the equipment alone would not pay it.
The rent matters as much as the method. The Rack Report's $4.67 H200 rent is a limited-sample market rate. At the managedgpus.com $3.25 3-year contract rate, on the contract method the Blackwell rows use, the same node has an income value of about $512K, still above the cost new because the contract is in it, against $586K at market rent.
Older servers can go the other way: an HGX A100 8x 80GB node earns $1.82 per GPU-hour, which the Rack Report values at $63K in place against $95K sold. When rent no longer covers what the hardware fetches, selling beats renting.
What sits between the two
Moving a server from in place to sold costs money that none of the ex-situ figures above deduct: de-installation, shipping (about $2,500 a server), data sanitization and a selling period. The node value also excludes the cluster network, which costs about $50-70K a server for an 800G Ethernet fabric (managedgpus.com ballpark, October 2026).
The in-situ premium lasts while the site, power, network and customer stay in place. A default tends to end at least one of them, which is why appraisals report liquidation values and why lenders arrange a backup operator who can keep a cluster earning in place.
Methodology
How the values are built
Ex-situ fair market value is the October 2026 Rack Report's Complete systems value: observed listings for whole nodes, adjusted to a reference configuration, not realized sales. Orderly and forced liquidation follow the GPU Value Guide method. In-situ values are income values, not fair market value in continued use: for Hopper and A100 servers the Rack Report's rental-implied value, for Blackwell servers the implied-value calculation behind our GPU price pages. Both are computed from the data at build time and move with each monthly issue.
Related
- GPU server values: what every complete server is worth this month
- HGX H100 server value: value, residual curve, earnings and payback
- HGX H200 server value: value, residual curve, earnings and payback
- HGX B200 server value: value, residual curve, earnings and payback
- HGX B300 server value: value, residual curve, earnings and payback
- GPU appraisals: third-party values on the premise your deal needs
- Are 6-year GPU depreciation schedules realistic?: useful lives against the resale curve
Discuss a transaction
Need a value on a specific premise for a loan or lease? Send us the equipment list and where it is installed. We will come back with in-place and liquidation values at the dates that matter.
Prefer email? hello@amcompute.com
Questions
- What is the difference between in-situ and ex-situ GPU value?
- In-situ value assumes the server stays installed and in use, with its site, power, network and customer; ex-situ value assumes it is removed and sold. The table above has both for each server.
- Which value do lenders use for GPU collateral?
- Ex-situ values, usually orderly or forced liquidation. The in-situ premium lasts only while the site, power, network and contract stay in place, and a default tends to end at least one of them.