As Blackwell ships, the previous flagship GPUs are flowing onto the secondary market — and for a lot of real workloads, a refurbished H100 or H200 is the smarter buy than waiting months and paying a premium for the newest silicon. But 'used GPU' covers everything from a warranted, inspected enterprise unit to an ex-mining card with an unknown thermal history. This is the honest guide to what's happening to prices, where refurbished AI hardware makes sense, and how to buy it without inheriting someone else's problem.
What's happening to prices
Supply and demand have shifted. H100 cards that sold for around $40,000 in late 2023 now move on secondary markets for roughly $12,000-$22,000 used, with refurbished units around $21,000-$34,000 and new around $25,000-$40,000 — and Blackwell's general availability is expected to push H100 secondary pricing down a further 10-20% as fleets rotate. The H200, with substantially more memory and bandwidth than the H100, is increasingly available second-hand too. (All figures indicative and fast-moving — confirm on the day.)
The key nuance: refurbished holds value far more steadily than raw 'used'. Tracked data put refurbished H100s at roughly 84-85% of contemporaneous new pricing, precisely because provenance, inspection and warranty lower the perceived risk. You pay a bit more than the cheapest used card and get a unit you can actually deploy with confidence.
When refurbished H100/H200 is the right call
These are still extremely capable GPUs. For inference, fine-tuning, and most enterprise AI that isn't frontier-scale pre-training, an H100 or H200 does the job — and the H200's larger memory is a genuine advantage for serving big models. Refurbished makes sense when:
- •You're running inference or fine-tuning, not training trillion-parameter models from scratch.
- •Lead time matters — secondary stock can be available now while new Blackwell allocation is months out.
- •Budget is the constraint and the price surge has stretched it — refurbished sidesteps the newest-silicon premium.
- •You want proven, mature software support (CUDA, drivers, frameworks are rock-solid on Hopper).
How to buy it without the risk
The difference between a bargain and a liability is the channel. Buy refurbished AI hardware from an ITAD/reseller that does multi-point inspection, has clear provenance (ideally ex-enterprise or ex-hyperscaler fleet, not ex-crypto), provides a warranty, and confirms firmware and health. Insist on a burn-in/health report, check the warranty terms and RMA path, and be wary of prices that are 'too good' — a card with an unknown thermal history under sustained AI load is a false economy. The warranty and provenance are what you're really paying the refurbished premium for.
The verdict
For a large share of UK AI workloads in 2026 — especially inference and fine-tuning under budget and lead-time pressure — warranted refurbished H100 or H200 is the value sweet spot: most of the capability at a fraction of frontier cost, available now. Buy frontier Blackwell when the workload genuinely needs it; otherwise let someone else absorb the depreciation.
Servnet sources warranted, inspected refurbished GPUs and complete GPU servers, and will be straight with you about when last-generation is the right answer and when it isn't.