✓Buyer Guides✓Technical Articles✓Industry Insights✓Equipment Tips
Buying Guides10 min readBy Caladan SemiUpdated: 2026-09-09

Used NVIDIA H100/A100 GPU Buying Guide 2026

Used H100 and A100 GPUs: observed US asking prices by form factor (Sep 9, 2026), datasheet specs, nvidia-smi checks that catch bad cards, PO clause.

This guide is for: the infrastructure lead who has budget approval for eight used GPUs, a spreadsheet of eBay and broker asks that disagree by 2x for what looks like the same card, and a training run that starts in six weeks whether the hardware arrives or not.

Eight H100 80GB modules at this week's US median SXM5 ask of $19,550 come to $156,400. The same eight GPUs as PCIe cards at the $32,000 median come to $256,000. The $99,600 you saved on paper disappears the afternoon someone opens the box and finds eight bare modules with no PCIe edge, no server in your rack that can take them, and a six-week clock that is now a five-week clock. The sentence to carry into the budget meeting: on the used market, form factor is the price, and the cheap form factor is the one you probably can't plug in.

H100 vs A100 Used Prices: What US Sellers Were Asking on September 9, 2026

We pulled every single-card and single-module listing for the H100 and A100 from the eBay Browse API on September 9, 2026, dropped lots, servers, brackets, cables, and anything flagged defective, and kept US-located sellers. These are asks, not sold prices. Where a variant had fewer than three US listings we print "price on request" rather than a number we don't trust.

GPU Market price (observed)
H100 80GB PCIe $28,000 to $32,399 · median $32,000 · 3 US listings · last checked Sep 9, 2026 · asking prices, not sold
H100 80GB SXM5 $14,000 to $22,550 · median $19,550 · 3 US listings · last checked Sep 9, 2026 · asking prices, not sold
H100 NVL 94GB PCIe $29,500 to $39,999 · median $37,750 · 4 US listings · last checked Sep 9, 2026 · asking prices, not sold
A100 80GB PCIe $14,950 to $24,999 · median $18,500 · 15 US listings · last checked Sep 9, 2026 · asking prices, not sold
A100 80GB SXM4 $8,300 to $12,200 · median $9,925 · 4 US listings · last checked Sep 9, 2026 · asking prices, not sold
A100 40GB PCIe $3,800 to $13,196 · median $7,750 · 22 US listings · last checked Sep 9, 2026 · asking prices, not sold
A100 40GB SXM4 price on request (1 US listing)

Read the H100 rows with the sample size in mind. Three US asks is a thin market, and the wider set showed why: the six new-in-box H100 PCIe listings were all overseas (China and Canada) at $27,884 to $52,950, so a US buyer paying in the low thirties for a used card is paying roughly what a new one costs to import before duty and freight. The A100 80GB PCIe row is the deepest and most trustworthy number on this page: 15 US asks, all secondhand (10 used, 4 open-box, 1 certified refurbished), clustered tightly around the median.

One more data point that didn't make the table. A US seller listed an A100 80GB PCIe at $12,500, below every other US ask, with "ECC Error" in the title. That is the honest version of the card you are trying not to buy at $16,999 from someone who left the ECC error out of the title.

Datasheet Differences That Change the Price

From the NVIDIA A100 datasheet (June 2021), the NVIDIA H100 datasheet, and the H100 NVL product brief. The compute gap between generations is large; the gap between form factors within a generation is mostly memory bandwidth, power, and the socket.

Spec A100 80GB PCIe A100 80GB SXM4 H100 80GB PCIe H100 80GB SXM5 H100 NVL 94GB
Architecture Ampere Ampere Hopper Hopper Hopper
GPU memory 80GB HBM2e 80GB HBM2e 80GB HBM2e 80GB HBM3 94GB HBM3
Memory bandwidth 1,935 GB/s 2,039 GB/s 2 TB/s 3 TB/s 3,938 GB/s
FP64 9.7 TFLOPS 9.7 TFLOPS 24 TFLOPS 30 TFLOPS see NVIDIA NVL brief
FP16/BF16 Tensor (with sparsity) 624 TFLOPS 624 TFLOPS 1,600 TFLOPS 2,000 TFLOPS see NVIDIA NVL brief
FP8 Tensor none none 3,200 TFLOPS* 4,000 TFLOPS* yes
Max TDP 300W 400W 350W 700W 400W
Interconnect NVLink Bridge 600 GB/s (2 GPUs), PCIe Gen4 64 GB/s NVLink 600 GB/s on HGX A100 NVLink 600 GB/s (bridge), PCIe Gen5 128 GB/s NVLink 900 GB/s on HGX H100 NVLink bridge 600 GB/s, PCIe Gen5
Cooling / form Passive dual-slot PCIe HGX module Passive dual-slot PCIe HGX module Passive dual-slot FHFL 10.5"
Host you need Any Gen4 x16 slot, 300W HGX A100 4/8-GPU board or DGX A100 Any Gen5 (or Gen4) x16 slot, 350W HGX H100 board or DGX H100 Gen5 x16 slot, 16-pin power at 400W

*With sparsity; halve for dense.

There is no "H100 40GB" and there is no "A100 SXM5." If a listing uses either phrase, the seller is guessing at their own inventory, and you should assume the rest of the listing is guessed too.

PCIe Card or SXM Module: The Decision Before the Price

Buy SXM only if you already own the baseboard. An A100 SXM4 module installs on an HGX A100 board or inside a DGX A100; an H100 SXM5 module installs on an HGX H100 board or inside a DGX H100. Nothing else. The reason the module asks are 46% (A100) and 39% (H100) below the PCIe asks is that the pool of buyers who can use one is small, and the sellers know it. If you have an HGX system with an open socket, the module is the best deal on this page. If you don't, it is a very expensive doorstop.

Buy PCIe when the GPU is going into a general-purpose server. Both the A100 and H100 PCIe cards are passive, so the chassis has to move air through them; a 2U server with GPU-rated fans is fine, a workstation with a side panel off is not, and a seller who tested the card on a bench with a desk fan has not tested the card. The H100 NVL wants a 16-pin PCIe power cable strapped for 450W or 600W to run at its full 400W; on an 8-pin adapter it caps at 310W, per NVIDIA's product brief.

Walk away if a seller offering an SXM module cannot show it running in a host. There is no other way to test one.

Three nvidia-smi Outputs That Catch Bad Cards

Ask for these from the exact unit, with the serial number visible in the screenshot. Both generations support all three.

  1. Row remapper. nvidia-smi -q -d ROW_REMAPPER. A100 and H100 HBM handle bad memory rows by remapping them. Walk away if "Pending" is Yes or "Remapping Failure Occurred" is Yes. A few correctable remaps with nothing pending is normal on a data-center pull.
  2. ECC counters. nvidia-smi -q -d ECC. Walk away on any aggregate uncorrectable error count above zero. Correctable aggregates in the hundreds on a multi-year card are typical; thousands mean the card ran hot for a long time.
  3. Throttle reasons under sustained load. Run gpu-burn or your own training step for 20 minutes, then nvidia-smi -q -d PERFORMANCE. "SW Power Cap" is fine. "HW Slowdown," "HW Thermal Slowdown," or "HW Power Brake Slowdown" on a card in a properly cooled chassis means a bad heatsink interface or a power delivery fault.

Match the label to the listing too: 900-21001-0020-xxx is the A100 80GB PCIe family, 900-21010-0000-xxx is the H100 80GB PCIe, and 699-21010-0210-xxx is the H100 NVL per NVIDIA's product brief. Two of the US A100 asks and two of the US H100 asks printed those numbers in their titles.

Below $13,000 for a working A100 80GB PCIe you are under every US ask we saw except the one with the ECC error in its title; below $25,000 for a working H100 80GB PCIe you are under every US ask, full stop. Treat either price as a question about the card.

The PO Clause

Paste this into the purchase order. It has no price in it and every seller with working cards can sign it.

"Seller will provide, for each unit and prior to shipment, the output of nvidia-smi -q -d ROW_REMAPPER,ECC,INFOROM and nvidia-smi -q -d PERFORMANCE captured after no less than 20 minutes of sustained load, with the unit serial number visible. Units showing pending row remaps, remapping failures, any aggregate uncorrectable ECC errors, or hardware thermal or power-brake throttling are non-conforming and may be rejected within 10 business days of receipt for a full refund including return freight."

Frequently Asked Questions

Q: How much does a used H100 cost in 2026? A: On September 9, 2026, US eBay asking prices for a single H100 80GB PCIe card ran $28,000 to $32,399 (median $32,000, 3 listings); the H100 80GB SXM5 module ran $14,000 to $22,550 (median $19,550, 3 listings); the H100 NVL 94GB ran $29,500 to $39,999 (median $37,750, 4 listings). Asks, not sold prices, and thin samples. Email parts@caladansemi.com for a quote on tested units.

Q: How much does a used A100 80GB cost? A: US eBay asks on September 9, 2026: $14,950 to $24,999 for the 80GB PCIe card (median $18,500, 15 listings) and $8,300 to $12,200 for the 80GB SXM4 module (median $9,925, 4 listings). The 40GB PCIe card ran $3,800 to $13,196, median $7,750, across 22 listings. Full detail on the used NVIDIA A100 80GB page.

Q: H100 SXM5 vs PCIe, which should I buy used? A: SXM5 only if you own an HGX H100 or DGX H100 with an open socket; the module doesn't install anywhere else. PCIe for any server with a Gen5 or Gen4 x16 slot, a 350W budget, and real airflow. SXM5 gives you HBM3 at 3 TB/s and 900 GB/s NVLink; the PCIe card gets 2 TB/s and 600 GB/s NVLink to one partner card through a bridge.

Q: What should I check before buying a used H100 or A100? A: Three nvidia-smi outputs from the exact card: ROW_REMAPPER (no pending, no failure), ECC (zero aggregate uncorrectable), and PERFORMANCE after 20 minutes of load (no HW throttling). Match the 900- or 699- board part number on the label to the listing, and get the serial in the screenshot.

Q: Is a used A100 still worth buying in 2026? A: For inference, fine-tuning, and HPC that fits in 80GB, an A100 80GB PCIe at the $18,500 US median is a little more than half the H100 PCIe median of $32,000 with about half the FP16 tensor throughput and no FP8. If your framework uses FP8 and you're training, the H100 gap is worth paying for. If you're serving models that already run on Ampere, it usually isn't.

What to Do Next

  1. Write down the host: server model, free slot generation, power budget per slot, and whether you own an HGX baseboard. That's free and it removes half the listings.
  2. Price off the observed table above, by form factor, and treat any ask under the low end as a question about the card, not a discount.
  3. Put the nvidia-smi clause in the PO before you discuss price.
  4. Or email parts@caladansemi.com with the quantity, form factor, and host system. We'll tell you which of this week's listings we would buy and what the diagnostics on them look like.

Related reading: A100 80GB PCIe vs SXM4: Used Prices & Spec Diff | NVIDIA A100 vs H100 on the Used Market | Used NVIDIA H100 80GB

Related Parts

Caladan stocks used and refurbished parts referenced in this article — tested, inspected, and ready to ship.