Every NVIDIA H100 ownership model has one number that quietly decides whether the investment works: what the hardware resells for when you are done with it.
Most TCO spreadsheets treat residual value as a footnote, a flat "assume 25% in three years" plugged in without evidence. That assumption is doing more work than people realize. On an owned NVIDIA H100, moving the residual estimate from 15% to 40% changes the depreciation line of your effective cost by roughly 40%. It is the second-largest controllable variable in ownership economics, behind only utilization.
This is what H100s actually resell for in 2026, where the secondary market lives, how residual flows into your real cost per GPU-hour, and when to sell. For the rate-of-decline analysis, see GPU Depreciation. For the full ownership picture, see The Real Cost of an H100 and Total Cost to Own 100 H100 GPUs.
What a used H100 actually sells for
H100s have held value unusually well for compute hardware. The secondary market in 2026 reflects continued demand, constrained supply of current-generation cards, and a slow Blackwell ramp at the long tail.
A snapshot of H100 SXM5 secondary pricing in 2026, against a $30,000 OEM list reference:
| Card age | Residual (% of list) | Approx resale value | Notes |
|---|---|---|---|
| New OEM | 100% | $30,000 | Reference point |
| 6 to 12 months | 85 to 90% | $25,500 to $27,000 | Minimal decline, still in-demand |
| 18 to 30 months | 75 to 85% | $22,500 to $25,500 | Residual peak window |
| 36 months | 40 to 75% | $12,000 to $22,500 | Wide range, Blackwell-ramp dependent |
| 60+ months | 25 to 35% | $7,500 to $10,500 | Residual floor |
| Refurb / decommissioned | n/a | $18,000 to $22,000 | Channel price, 12-month limited warranty |
The standout feature is the 36-month range. That single row is where most of the financial uncertainty in GPU ownership concentrates, and the spread is driven almost entirely by how fast Blackwell reaches mainstream availability. A faster ramp pushes H100 36-month residual toward 40%. A slower one holds it at 65 to 75%. Most institutional financial planning should run both.
Why GPUs hold value better than ordinary hardware
Standard enterprise IT depreciates to near-zero in three to five years. H100s do not, and the reasons are structural rather than sentimental.
CUDA lock-in. The software ecosystem is built on NVIDIA. A used H100 drops into existing CUDA, library, and framework toolchains with no migration cost. That keeps demand deep across the entire installed base, not just first buyers.
Supply constraint. Current-generation NVIDIA supply has been allocated tightly. As long as new H100 and H200 capacity is hard to get at list, used cards have a real floor because they are a substitute for unavailable new inventory.
Software support runway. NVIDIA continues CUDA and framework support for H100 through at least 2028. A buyer purchasing a used H100 in 2026 gets a multi-year supported life, which is what makes the secondary purchase rational.
Continued inference utility. Even as H100 stops being the frontier training card, it remains a strong inference and mid-size training GPU. That second life is the floor under the 60-month residual. The card does not become useless; it moves down the workload stack. See B200 vs H100 for how Blackwell's arrival reshapes where H100 lands.
Where the secondary market actually is
"Resale value" only matters if you can realize it. The channels differ sharply on price and liquidity:
OEM and integrator trade-in. The lowest-friction, lowest-price option. Supermicro, Dell, and integrators will take fleet trade-ins against new orders. Convenient for rotation, but you give up margin versus selling direct.
Specialized hardware brokers. Brokers who move datacenter GPUs match decommissioning sellers with buyers. Better price realization than OEM trade-in, with a commission. The main channel for fleet-scale disposal.
Peer-to-peer between operators. Operators sell directly to other operators, often regionally. Best price, most effort, requires trust and logistics. Works best for nodes, not loose cards.
Decommission and liquidation auctions. When a datacenter or a failed venture liquidates, lots hit auction. This is where refurbished inventory originates, and where the $18,000 to $22,000 refurb pricing gets set.
A liquidity gradient runs through all of this. An intact 8-GPU HGX node is far easier to sell than eight loose SXM5 cards, because buyers want deployable units. Fleet rotations from a known operator place faster than a single card from an unknown seller. If resale is part of your plan, keep nodes intact and keep provenance documented.
How residual value moves your real cost per GPU-hour
This is where the resale number stops being abstract. Residual feeds directly into the depreciation component of your effective cost.
The formula:
// text
Effective depreciation $/GPU-hour =
(Capex − Resale_value) / (Lifespan_hours × Utilization)
Residual realization rate = Resale_value / CapexWorked example. Take a single H100 deployed in an 8-GPU node at $33,000 all-in capex per GPU, a 3-year horizon, and 70% utilization. That gives 3 × 8,760 × 0.70 = 18,396 useful GPU-hours. Now run three residual assumptions:
| Residual at 36 months | Resale value | Net depreciation | Depreciation $/GPU-hour |
|---|---|---|---|
| 40% | $13,200 | $19,800 | $1.08 |
| 25% | $8,250 | $24,750 | $1.35 |
| 15% | $4,950 | $28,050 | $1.52 |
Same card, same utilization, same horizon. The residual assumption alone swings the depreciation line from $1.08 to $1.52 per GPU-hour, a 40% range. Layer in power, colocation, and ops at roughly $0.40 per GPU-hour, and your all-in owned cost lands somewhere between $1.48 and $1.92 depending on a number most models guess at.
The practical takeaway: if you are building an ownership case, the residual estimate deserves as much scrutiny as your utilization assumption. A buy-versus-rent decision that looks marginal at 25% residual can flip in either direction once you pin the number to real secondary-market data. See Buy vs Rent GPUs and GPU ROI for where this slots into the full decision.
When to sell: the residual peak and the timing decision
The resale data points to a clear retention peak at 18 to 24 months, where cards still hold 75 to 85% of list. After that, the 36-month cliff opens up, with its width set by Blackwell. This creates a real timing decision.
Hold and amortize (the default). Run the fleet through the full 3-year horizon and sell at end-of-life. Simple, predictable, and right for most teams with steady utilization. You accept the 36-month residual wherever it lands.
Pre-emptive sale at peak (the option). Sell at the 18 to 24 month mark while cards still retain 75 to 85%, then rotate to current-generation hardware. This minimizes total depreciation cost, but only works if you have the capital to rotate, the operational capability to migrate, and a workload plan that actually closes. Done wrong, you eat transaction costs and downtime for nothing.
The variable that decides which is right is Blackwell ramp speed, the same factor driving the 36-month residual range. If Blackwell ramps fast, the pre-emptive sellers look smart because the cliff is steep. If it ramps slow, holders win because the floor stays high. You cannot know in advance, which is why most teams should default to hold-and-amortize and treat pre-emptive sale as a play for sophisticated operators only.
Track current H100 and H200 secondary and cloud pricing through GPU Index before timing any disposal.
What this means for buyers and owners
If you are buying new: do not accept a flat residual assumption in your TCO model. Pull real secondary-market comparables for the card and age you are modeling, and run both a fast-Blackwell and slow-Blackwell residual case. The gap between them is your real risk exposure on the ownership decision.
If you already own H100s: your residual is an asset with a peak. Know where you are on the curve. If you are past 24 months and watching Blackwell ramp, model the disposal decision explicitly rather than defaulting into a full hold by inertia.
If you are buying used: refurbished H100s at $18,000 to $22,000 with a 12-month warranty are a reasonable path for research, development, and inference workloads. For production, weigh the warranty gap against OEM-fresh hardware. The card has years of supported life left, but you are buying someone else's utilization history.
Frequently Asked Questions
What is GPU residual value?
Residual value is what a GPU is worth on the secondary market at a given age, expressed either as a dollar figure or as a percentage of original list. It is the recovery side of depreciation, and it directly reduces the net cost of owning the hardware over its life.
What does a used H100 sell for in 2026?
H100 SXM5 cards 18 to 30 months old resell at roughly 75 to 85% of original OEM list, or about $22,500 to $25,500 against a $30,000 reference. Refurbished or decommissioned H100s sell for $18,000 to $22,000, typically with a 12-month limited warranty. At 36 months the range widens to 40 to 75% depending on Blackwell availability.
Where can I sell used H100 GPUs?
Four main channels: OEM and integrator trade-in (lowest friction, lowest price), specialized hardware brokers (best for fleet-scale disposal), peer-to-peer operator sales (best price, most effort), and decommission auctions. Intact 8-GPU nodes sell faster and for more than loose individual cards.
Do H100s hold their value?
Better than typical enterprise hardware. CUDA ecosystem lock-in, constrained new supply, NVIDIA software support through at least 2028, and continued inference utility keep demand deep across the installed base. The 60-month residual floor sits at roughly 25 to 35%, well above where ordinary IT hardware lands.
When should I sell my H100s?
The retention peak is 18 to 24 months, where cards hold 75 to 85% of list. Most teams should hold and amortize through the full 3-year horizon. Pre-emptive sale at the peak minimizes total depreciation cost but only works for operators with the capital, migration capability, and a workload plan to rotate cleanly. Blackwell ramp speed is the deciding variable.
Are refurbished H100s worth buying?
For research, development, and inference, refurbished H100s at $18,000 to $22,000 are a reasonable value, especially with a 12-month warranty. For production workloads, weigh the shorter warranty and unknown utilization history against the premium for OEM-fresh hardware.
How does residual value affect total cost of ownership?
Significantly. On a $33,000-per-GPU H100 at 3 years and 70% utilization, moving residual from 15% to 40% changes the depreciation component of effective cost from about $1.52 to $1.08 per GPU-hour, a 40% swing. It is the second most important controllable variable in ownership economics after utilization.
Will Blackwell crash H100 resale value?
It is the dominant risk to H100 residual, but a crash is not the base case. A faster Blackwell ramp pushes 36-month residual toward 40%; a slower one holds it at 65 to 75%. Continued inference demand and software support through 2028 put a floor under the decline. Model both scenarios rather than assuming either.
Methodology
Secondary-market pricing in this article is derived from Mercatus tracking of used and refurbished H100 transactions across broker, OEM trade-in, and auction channels in 2026. OEM list reference pricing reflects public quotes from Supermicro, Dell, and HPE. Residual scenarios reference GPU Depreciation (40 to 75% at 36 months depending on Blackwell ramp; 25 to 35% floor at 60+ months). The effective-cost worked example uses the ownership framework from The Real Cost of an H100. Live H100 and H200 pricing is tracked through GPU Index. Last verified: 2026-06-23.
Stop guessing at residual value. Mercatus GPU Index tracks live H100 and H200 cloud and secondary-market pricing across 30+ providers, so you can plug real numbers into your ownership model instead of a flat assumption.