The clearest signal in the 2026 GPU market isn't the Blackwell launch — it's what's happening to the Hopper cards it's displacing. H100 cards that sold for around $40,000 in late 2023 are now trading at roughly $12,000-$22,000 used, according to UK reseller Servnet's own pricing read, quoted by GPU Smith in July 2026, with refurbished units running $21,000-$34,000. Multiple trackers put H100 SXM5 depreciation at 34% by January 2026 alone. For UK IT leaders eyeing AI infrastructure, the question isn't whether Hopper is cheaper now — it obviously is — but whether a discounted card with an uncertain service history is a bargain or a liability.
View the data behind this chart
| Used H100 | Refurbished H100 | New H100 | |
|---|---|---|---|
| Low | $k12 | $k21 | $k25 |
| High | $k22 | $k34 | $k40 |
The used Hopper flood in mid-2026
Blackwell has changed where Hopper sits in the pricing hierarchy. BuySellRam's April 2026 market note put Blackwell B200 boards at $30,000-$40,000 — a range that now sits above, not below, most H100 quotes, confirming Hopper is no longer the top-tier resale object it was through 2024 and 2025.
That repricing has been abrupt. GPU Smith, citing Exit Technologies auction data, recorded 34% depreciation on H100 SXM5 units by January 2026, with the slide directly triggered by the March 2025 Blackwell launch. Ex-hyperscaler fleets refreshing onto Blackwell are the supply behind this: decommissioned H100 and H200 cards are landing on broker desks, auction platforms and reseller shelves faster than demand for last-gen Hopper can absorb them at 2023-era prices.
UK pricing landscape: used, refurbished and new H100
The most UK-relevant benchmark comes from Servnet's own pricing, as quoted by GPU Smith in July 2026: used H100 cards at roughly $12,000-$22,000, refurbished units at $21,000-$34,000, against new H100 pricing of $25,000-$40,000. CloudZero independently confirmed the same $12,000-$22,000 used band in its July 2026 update, while also flagging much wider dispersion elsewhere in the market: deep-discount channels as low as $6,000-$15,000 versus lightly discounted stock at $21,000-$34,000.
Other trackers sit in adjacent but distinct bands, and the differences matter. BuySellRam's April 2026 figures show lightly used or certified H100 SXM units at $17,000-$28,000. GPU Smith's own July 2026 index shows used H100s anywhere from $8,200 to $25,000 depending on age and channel, with March 2026 auction clears as low as $8,200-$12,500 — the no-paperwork end of the market. Intuition Labs put the wider used range at $6,000-$22,000, describing it as an 85% decline from the 2023 peak of $40,000. Mercatus AI's July 2026 resale guide is more granular still: H100 SXM5 cards aged 18-30 months trading at 75%-85% of a $30,000 reference (roughly $22,500-$25,500), and refurbished or decommissioned units at $18,000-$22,000 typically carrying a 12-month limited warranty.
H200 sits above H100 throughout. BuySellRam puts the H200 NVL (141GB HBM3e) premium at 15%-40% over H100. CCIR's index — tracking executed sales rather than headline asks — shows H100 NVL 94GB units with a posted-ask median of $35,265 against an executed 90-day median of just $26,250, and H200 141GB with a posted-ask median of $36,499 against an executed median of $22,500. That ask-versus-executed gap is the single most useful number in this whole market: sellers quote one figure, buyers who negotiate hard pay meaningfully less.
PCIe vs SXM5 vs NVL: why your quote depends on the form factor
"Used H100" is not one product. PCIe air-cooled cards, SXM5 baseboard modules, and NVL variants carry different infrastructure requirements and depreciate differently on resale. CCIR's data specifically tracks H100 NVL 94GB and H200 141GB configurations, not a blanket "H100" figure — and even within that narrow scope, posted-ask prices sit well above what actually clears.
Before comparing any two quotes, confirm the exact SKU, memory capacity and cooling type in writing. A PCIe quote and an SXM5 quote are not interchangeable, and a seller happy to let you assume otherwise is one worth being cautious of.
Why a used H100 can still be the right call
At $12,000-$22,000 used against $25,000-$40,000 new (both GPU Smith, July 2026), a used H100 gets a UK buyer into serious training and inference capacity for roughly half the new-unit entry cost — capacity that's still entirely capable for workloads that don't require Blackwell-class density.
Rental economics have also moved against pure cloud consumption. SemiAnalysis tracked 1-year H100 rental contract pricing rising from $1.70 per GPU-hour in October 2025 to $2.35 per GPU-hour by March 2026. For teams running near-continuous workloads rather than short bursts, that trajectory makes owned hardware — even second-hand — increasingly competitive against opex-only cloud spend. Anyone weighing this trade-off should check current UK GPU cloud rental prices before committing capital either way.
The risks: depreciation speed, warranty gaps, grey-market provenance
The same 34% SXM5 depreciation swing that makes Hopper cheap today is a warning, not just a discount. A market that moved that fast once Blackwell arrived can move again as Blackwell supply normalises further — buying now assumes the discount holds, not that it deepens.
Warranty cover thins out fast as price drops, and it varies by channel as much as by price. New H100 units carry full OEM warranty cover from NVIDIA or its channel partners. Refurbished stock — such as Mercatus AI's decommissioned units — typically carries a reseller-backed 12-month limited warranty: defined, but a fraction of new-unit cover. The sub-$15,000 auction-grade cards GPU Smith documented in March 2026 commonly carry no warranty at all, leaving the buyer to absorb any failure risk in full. For UK enterprises treating the purchase as a recoverable capex asset rather than a disposable experimental cluster, that gap matters to finance teams as much as to engineers: no service record, no support contract, and no easy recourse if a card fails outside a short return window.
View the data behind this chart
| Price / Rate | Warranty | Best fit | |
|---|---|---|---|
| Used H100 SXM (Servnet… | $12k-$22k | Limited/varies | Steady long-run training |
| Refurbished H100… | $21k-$34k | 12-month typical | Production workloads |
| New H100 (GPU Smith) | $25k-$40k | Full OEM warranty | Compliance workloads |
| Cloud rental, 1-yr… | $2.35/GPU-hr (Mar26) | N/A - opex only | Short-burst scaling |
| H200 Capacity Block… | $4.98/GPU-hr | N/A - opex only | On-demand H200 access |
Your due-diligence checklist before you buy
None of the pricing above tells you whether a specific card is sound. Before wiring funds — especially to US auction or broker channels shipping into the UK — work through this list:
- •Confirm the exact SKU (PCIe, SXM5 or NVL) and memory capacity match what's quoted — don't assume from a generic "H100" listing.
- •Request original deployment history or hyperscaler decommission paperwork where available.
- •Ask for burn-in or sustained-load stress-test logs, not just a power-on/idle check.
- •Get ECC memory error counts and thermal history from the vendor if they can supply it.
- •Establish precisely what warranty period, if any, transfers with ownership — 12-month limited cover is the going norm for warrantied refurbished stock, not a given for auction-grade units.
- •Get all terms, condition claims and warranty commitments in writing before payment, particularly on cross-border deals.
TCO worked example: buy used vs buy new vs rent
Take Servnet's UK-quoted used H100 band of $12,000-$22,000 against SemiAnalysis's March 2026 rental rate of $2.35 per GPU-hour. Run that rate continuously for a year (8,760 hours) and it works out to roughly $20,586 per GPU — illustrative arithmetic from the stated rate, not a separately sourced figure, but it lands squarely inside the used-purchase price band itself. For a workload that keeps a GPU busy close to around the clock, buying a used card can pay for itself in rental-equivalent terms within about a year, before power, cooling and support costs are counted on either side of the ledger.
The same maths is starker for H200. Cast AI's figures put AWS's H200 Capacity Block rate at $4.98 per GPU-hour ($39.80/hour for 8 GPUs), which annualises to roughly $43,600 per GPU on continuous use — nearly double CCIR's executed 90-day median of $22,500 for H200 141GB units. New-unit and refurbished purchases still carry the OEM or vendor warranty that rental never needs, but they also carry the up-front capital outlay, VAT, import handling and shipping costs that a US-sourced auction or broker deal brings with it — costs that don't show up in any of the headline dollar figures quoted above and need budgeting separately for a UK deployment. Because grey-market or auction-sourced cards typically won't come with the service record or support contract that enterprise finance teams expect for a capitalised asset, UK buyers treating the purchase as a recoverable capex item — rather than a short-life experimental cluster — should weigh VAT and import handling into the landed cost before comparing it against the headline dollar prices quoted by US trackers. Power and cooling infrastructure is a second overlooked line: an SXM5 baseboard system draws substantially more sustained power than typical rack servers were designed for, so rack power, cooling capacity and UPS headroom all need re-verifying before a used card goes into service, not after. None of the purchase-price or rental figures above capture ongoing maintenance or failure risk on a used card; the due-diligence checks above — burn-in logs, ECC error counts and thermal history — are the closest available proxy for estimating that risk before it turns into an unplanned support cost rather than a line item in a TCO model.
Sourcing channels and the verdict for UK buyers
For sourcing, this brief points to a handful of channel types rather than a single best answer: UK reseller quotes such as Servnet's, US broker and market-tracker outlets like BuySellRam and GPU Smith, auction data via specialists like Exit Technologies, refurbished specialists such as Mercatus AI offering warrantied stock, and index services like CCIR that let you check any asking price against what's actually clearing before you negotiate. Using CCIR-style executed-versus-ask benchmarking on any quote you receive is the single cheapest piece of due diligence available.
The verdict: a used H100 makes strong financial sense for steady, continuous, non-cutting-edge training or inference workloads, where a 12-month-warrantied refurbished unit in the $18,000-$34,000 band beats climbing 2026 cloud contract rates. Sub-$15,000 auction-grade cards without paperwork or warranty are a false economy for anything you plan to run in production. New H100 or Blackwell-class hardware only earns its premium where compliance, full OEM warranty, or peak density genuinely matter — worth weighing against the fuller H100, H200, B100, B200 comparison and against refurbished servers for AI as a lower-risk route into second-hand capacity.
Sources
Every figure in this article traces to the sources below.
- •BuySellRam — H100/H200/Blackwell secondary market pricing, April 2026
- •GPU Smith — NVIDIA AI GPU price index and trends, July 2026
- •CloudZero — H100 GPU cost analysis, July 2026
- •SemiAnalysis — GPU rental capacity pricing, March 2026
- •Mercatus AI — H100 resale value guide, July 2026
- •Intuition Labs — Data centre GPU pricing 2026
- •Cast AI — GPU cloud pricing (AWS Capacity Blocks)
- •CCIR — GPU hardware pricing index (posted-ask vs executed medians)