GPU Residual Values 2026: The UK Depreciation Study

The State of the Used GPU Market in Mid-2026

By August 2026 the secondary market for enterprise GPUs has enough visible pricing infrastructure — CCIR's market index, Compute Exchange's listings, ServerBuyback's resale index, Clustermelt's pricing page and Hashrate Index's aggregated listings — that residual value can finally be observed rather than guessed from cloud rental rates alone. That maturity comes with a caveat: what's published is mostly asking prices, and asking prices are not selling prices.

CCIR's own trailing 90-day data makes the gap explicit: the H100 NVL 94GB's posted-ask median sits at 132% of its executed median. In plain terms, listed prices for this SKU are running almost a third above what buyers are actually paying. Clustermelt's 23 August 2026 pricing page shows top-tier AI accelerators clustering in the high-$20,000s to mid-$30,000s, broadly consistent with the CCIR and Compute Exchange figures below — but every one of these numbers is USD, and none of them include what it costs a UK buyer to actually take delivery.

Enterprise GPU Resale Values: H100, H200 and A100 Compared

The clearest dataset is on the two current Hopper-class chips. For the H200 141GB, Compute Exchange's indicative used-market page (20 August 2026) shows $30,000–$35,000, while CCIR's trailing 90-day figures put the posted-ask median at $38,248 against an executed median of $27,155 — a wide but informative spread between what sellers hope for and what actually clears. The H100 NVL 94GB shows the same pattern: CCIR's posted-ask median is $34,050 against an executed median of $25,750.

Lower down the Hopper PCIe stack, ServerBuyback's resale index (July 2026) lists used H100 PCIe 80GB units at $14,000–$25,000 against $25,000–$34,000 new, while Compute Exchange's separate used H100 PCIe page shows $16,000–$20,000 — a narrower band that overlaps ServerBuyback's lower end. CloudZero's August 2026 buy-price estimate for H100 more broadly spans $25,000–$40,000 depending on PCIe versus SXM5 configuration; this is a buy-price estimate, not a resale median, and shouldn't be read as the same metric as CCIR's figures.

Ampere-generation A100 pricing sits well below Hopper. Hashrate Index's tracked commercial listings (April 2026) show A100 40GB refurbished units from about $7,800 and A100 80GB PCIe units around $18,900. These are visible listings rather than confirmed sales data, but they establish A100 as the clear budget tier in this dataset, with H200 commanding a quote-based premium over H100 for higher-memory inference work, per CloudZero.

The Residual Value Curve: What Launch-Price Retention Really Shows

CCIR's retention figures give the first proper depreciation curve for this hardware class: the H200 141GB, introduced in 2024, has held 86.2% of its launch value on a posted-ask basis roughly two years in; the H100 NVL 94GB, from the prior generation, has slipped to 76.6%. That ten-point gap is the closest thing to a per-generation depreciation rate that currently exists in public data.

The important caveat for anyone writing down an asset register is that both retention figures are calculated on posted-ask prices, and CCIR's own executed-median data shows asks running well above what buyers actually pay for both chips. A finance team modelling realistic resale value should treat the 86.2% and 76.6% figures as ceilings, not settled outcomes, and reference the executed medians — $27,155 for H200, $25,750 for H100 NVL — as the more conservative basis for impairment testing.

New vs Used: A Worked Cost Comparison for UK Buyers

ServerBuyback's side-by-side range for the H100 PCIe 80GB gives a clean single-GPU comparison: $25,000–$34,000 new against $14,000–$25,000 used, a gap wide enough to change a cluster business case on its own. At the system level, CloudZero puts an 8x H100 build at roughly $250,000–$320,000 and states that new capacity runs about 70% more than equivalent used capacity — a server-level figure, not a per-GPU one, so it shouldn't be divided down to imply a single-card saving.

For a UK buyer, none of these USD figures are the real comparison point. Import VAT, duty, shipping and brokerage sit on top of the used-market ask before it lands on a UK invoice, and because most of this market prices in dollars, sterling movement between quote and payment can move the effective saving materially even when the US-dollar ask hasn't changed at all. Anyone sizing a cluster against this data should run the shortlist through a proper sizing exercise — you can calculate your AI GPU requirements before comparing landed used-market quotes against new list price.

Buying Used Enterprise GPUs: The Vetting Checklist

The spread between headline ask prices and what a card is actually worth makes vetting non-negotiable, particularly given how much of this market sits between fully warranted ex-cloud stock and unverified grey-market inventory.

Before agreeing a price, UK buyers should confirm provenance and remaining warranty status, run a sustained benchmark load rather than a brief power-on test, check thermal behaviour under load for signs of prior throttling or fan wear, verify firmware and driver compatibility with the target platform, cross-check seller reputation across the specific marketplace or broker used, and get an itemised landed-cost quote — VAT, duty, shipping, insurance — before comparing it against any USD ask price. Buyers weighing Hopper-class purchases specifically should consider buying second-hand Hopper GPUs only once provenance and thermal history are confirmed.

• •Confirm ex-cloud or OEM provenance and any transferable warranty
• •Run sustained benchmark and stress testing, not just a boot check
• •Inspect thermal performance and fan/VRM wear under load
• •Verify firmware, BIOS and driver compatibility before purchase
• •Assess seller reputation on the specific marketplace or broker used
• •Request a full landed-cost quote (VAT, duty, shipping) before agreeing price

Selling Strategy: Reading the Ask-Executed Spread

Sellers listing at headline ask prices should expect meaningful negotiation: CCIR's data shows the H100 NVL 94GB's ask sitting at 132% of its actual trailing 90-day executed median, and the H200's ask-to-executed gap ($38,248 versus $27,155) tells the same story. Pricing at the published ask and expecting it to clear at that level is not supported by this data.

The same two-tier dynamic shows up in rental pricing, which offers a useful secondary signal even though it's not a resale-price metric. GPUaaS reports a hyperscaler quoting H200 capacity at $10.60 per hour in July 2026, while a wholesale buyer paid $3.79–$4.54 per hour for the same chip — a gap of the same character as the hardware ask-executed spread. Wccftech's reporting of Silicon Data's index put H100 hourly rental just shy of $2 in January 2026, useful context for utilisation trends but not a substitute for hardware resale data. Sellers timing an exit should watch executed medians, not list prices, as the real leading indicator.

Channels, Cloud Alternatives and the Road Ahead

The dataset behind this study comes from several distinct channel types, each with different strengths for a UK buyer. CCIR functions as a market-index provider, publishing trailing medians and retention metrics rather than individual listings. Compute Exchange and Clustermelt operate more like marketplaces, showing indicative asking prices that need verifying against executed data. ServerBuyback and Hashrate Index behave more like resale-index and listings-aggregation services respectively — useful for benchmarking a quote, less useful as a guarantee of what a specific card will fetch.

Geographically, every pricing source in this study — CCIR, Compute Exchange, ServerBuyback, Hashrate Index, Clustermelt and GPUaaS — is a US-facing index, marketplace or reseller quoting in USD, and none of them represent confirmed UK-based stock. In practice this makes the used enterprise GPU market a cross-border one for UK buyers: sourcing a used H100 or H200 typically means shipping from an overseas seller rather than drawing on local inventory, so shipping time and customs clearance sit alongside VAT, duty and FX exposure as factors that widen the gap between a headline USD ask and the price and lead time a UK buyer actually experiences.

None of the sources in this dataset provide forward guidance on how the next architecture cycle will move today's Hopper-class resale values, so this study won't guess. What the data does show is that the ask-executed spread is already wide before any new product cycle lands, which suggests the market is pricing in some anticipated repricing risk now rather than waiting for an announcement. Buyers assessing whether to hold, sell or buy into this generation should treat trailing executed medians — not static ask prices — as the metric to monitor, and should weigh used purchase against renting capacity or understand the refurbished GPU market before committing capital to owned hardware.

Methodology

This study compiles published secondary-market pricing for NVIDIA A100, H100 and H200 GPUs from market-index providers, hardware marketplaces, resale-value indices and cloud-cost research published between April 2026 and August 2026, with figures cross-checked against their original publisher pages in August 2026. Sources include CCIR's trailing 90-day market data, Compute Exchange's listed hardware pages, ServerBuyback's resale-value index, Hashrate Index's tracked commercial listings, Clustermelt's secondary-market pricing page, CloudZero's cost research and GPUaaS's rental-market reporting, supplemented by Wccftech's reporting of Silicon Data's rental index.

Each figure is presented with its original scope preserved: posted-ask prices are never merged with executed medians, marketplace indicative pricing is kept separate from resale-index bands, and rental-rate data is treated as supporting context rather than a direct residual-value input. Where two sources describe similar hardware over different time windows or with different pricing bases, both are shown distinctly rather than averaged or combined into a single figure.

Sources

Every figure in this article traces to the sources below.

• •Compute Exchange — used H200 and H100 PCIe indicative pricing
• •Compute Exchange — used H100 PCIe listing page
• •CCIR — trailing 90-day medians and launch-value retention for H100 NVL and H200
• •CloudZero — H100/H200 buy-price estimates and new-vs-used system cost
• •ServerBuyback — GPU Resale Value Index for H100 PCIe 80GB
• •Hashrate Index — visible commercial listings for A100 40GB and 80GB
• •Clustermelt — secondary-market pricing page for AI accelerators
• •GPUaaS — hyperscaler vs wholesale H200 rental pricing
• •Wccftech (reporting Silicon Data) — H100 rental rate context