Confirm condition, inspection evidence and commercial terms in the quote.

NVIDIA

Used NVIDIA A100 80GB Tensor Core GPU (PCIe and SXM4)

No photo

NVIDIA

Used NVIDIA A100 80GB Tensor Core GPU (PCIe and SXM4)

Market Price

Request Quote — No current market data

Request pricing for the required configuration and condition

Request Quote →

Contact for availability. Stock and condition require confirmation for a specific unit.

Get a quote for this partPricing and availability to be confirmed

Selling this part? Place an ask · Talk to a specialist · parts@caladansemi.com

✓Request unit-specific inspection and test records
✓Confirm included documentation and accessories
✓Agree delivery, payment and return terms in the quote

How much does a used Used NVIDIA A100 80GB Tensor Core GPU (PCIe and SXM4) cost in 2026?

Pricing for used Used NVIDIA A100 80GB Tensor Core GPU (PCIe and SXM4) units varies by condition and configuration. Contact us for current 2026 market pricing. Get a quote.

What should I confirm before buying used NVIDIA AI/GPU Compute?

Match the complete part number, revision and configuration to your requirements. Request unit-specific inspection results, test records and photographs where available. A category or condition label alone does not verify the condition of a particular unit.

What terms apply to a quote for Used NVIDIA A100 80GB Tensor Core GPU (PCIe and SXM4)?

Ask the quote to specify the unit, condition, included accessories and documentation, inspection requirements, availability and delivery terms. Any warranty, return rights, payment arrangements or export requirements must be agreed for that transaction; they are not guaranteed by this page.

Used NVIDIA A100 80GB: One GPU, Two Form Factors, a 46% Price Gap

The same GA100 die and the same 80GB of HBM2e ship in two packages, and on the used market they price like two different products. On September 9, 2026, US sellers on eBay were asking a median $18,500 for the A100 80GB PCIe card and $9,925 for the A100 80GB SXM4 module. The gap is not a bargain signal. A bare SXM4 module plugs into an HGX A100 baseboard and nothing else, so the buyers for it are people who already own the board with an empty socket. The PCIe card goes into any server with a free PCIe Gen4 x16 slot, a 300W power budget, and enough chassis airflow to cool a passive dual-slot card.

Sort out which one your hardware accepts first. Then read the prices.

Market Price (Observed): Used NVIDIA A100, September 2026

These are live asking prices from US-located eBay listings for single cards or modules, pulled through the eBay Browse API on September 9, 2026. Asks, not sold prices. We excluded lots, servers, brackets, cables, and listings flagged as defective. Where fewer than three US listings existed, we print "price on request" rather than a number.

Variant Market price (observed)
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 on Sep 9, 2026, too few to publish)

Two things in that data you should know before you anchor on the medians. All 15 US PCIe asks were secondhand stock: 10 listed as used, 4 as open-box, 1 as certified refurbished, none as new. The new-in-box asks in the wider set were all outside the US and ran higher. And one $12,500 US listing that would have been the cheapest used 80GB PCIe card was titled "ECC Error." We dropped it from the table. A seller who tells you about the ECC error in the title is doing you a favor; a seller who doesn't is the one this page is about.

The full global set was 22 PCIe asks from 17 sellers (high of $31,196, a new card from Canada) and 9 SXM4 asks from 8 sellers, including two Dell V40HX-branded 500W SXM4 modules from Poland at $13,803 and $14,070.

Datasheet Specs: A100 80GB PCIe vs SXM4

From the NVIDIA A100 Tensor Core GPU datasheet (June 2021) and the A100 80GB PCIe product brief. Compute is identical across the two form factors; memory bandwidth and power are not.

Spec A100 80GB PCIe A100 80GB SXM4
Architecture NVIDIA Ampere (GA100) NVIDIA Ampere (GA100)
CUDA cores 6,912 6,912
Tensor cores (3rd gen) 432 432
GPU memory 80GB HBM2e 80GB HBM2e
Memory bandwidth 1,935 GB/s 2,039 GB/s
FP64 / FP64 Tensor 9.7 / 19.5 TFLOPS 9.7 / 19.5 TFLOPS
FP32 19.5 TFLOPS 19.5 TFLOPS
TF32 Tensor 156 TFLOPS (312 with sparsity) 156 TFLOPS (312 with sparsity)
FP16 / BF16 Tensor 312 TFLOPS (624 with sparsity) 312 TFLOPS (624 with sparsity)
INT8 Tensor 624 TOPS (1,248 with sparsity) 624 TOPS (1,248 with sparsity)
Max TDP 300W 400W
Multi-Instance GPU up to 7 MIGs at 10GB up to 7 MIGs at 10GB
Interconnect PCIe Gen4 64 GB/s; NVLink Bridge 600 GB/s for 2 GPUs NVLink 600 GB/s via HGX A100 board
Cooling Passive, dual-slot, needs chassis airflow Board-level heatsink on HGX baseboard
Host Any server with a PCIe Gen4 x16 slot and 300W budget HGX A100 4- or 8-GPU baseboard, DGX A100

The 40GB cards use HBM2 at 1,555 GB/s in both form factors, with a 250W TDP for PCIe. If a listing says "A100 80GB" and the photo shows a bare module with a bottom-mounted connector array and no PCIe edge, it is an SXM4 module regardless of what the title says.

Before You Pay: Three nvidia-smi Checks and Two Walk-Away Lines

Ask the seller to run these on the exact card, screenshot the output with the serial number visible, and send it. Every A100 supports them; a seller who cannot produce them either doesn't have the card in a working host or doesn't want you to see the result.

  1. Row remapper state. nvidia-smi -q -d ROW_REMAPPER reports remapped rows (correctable and uncorrectable) plus a pending flag. A100 HBM2e uses row remapping in place of the older page-retirement scheme. Walk away if "Pending" is Yes or "Remapping Failure Occurred" is Yes. A handful of correctable remaps with no pending flag is normal wear on a data-center pull.
  2. ECC error counters. nvidia-smi -q -d ECC shows volatile and aggregate counts. Walk away on any aggregate uncorrectable count above zero. Aggregate correctable counts in the low hundreds on a multi-year card are typical; thousands are a card that has been running hot or is on its way out.
  3. Throttle reasons under load. With a sustained load running (a gpu-burn pass or your own training step), nvidia-smi -q -d PERFORMANCE lists active clock throttle reasons. Thermal slowdown or power-brake throttling on a PCIe card at 300W in a properly cooled chassis means a failing heatsink interface or a seller testing it on a desk fan. Software power cap is fine; HW slowdown is not.

Also record nvidia-smi -q -d INFOROM (the InfoROM image version and ECC mode) and confirm the board part number on the label matches the listing: 900-21001-0020-xxx is the 80GB PCIe family, and two of the US listings printed it in the title.

Below roughly $13,000 for an 80GB PCIe card sold as working, you are underneath every US ask we saw except the one with the ECC error in its title. Treat that price as a question, not a discount.

Frequently Asked Questions

Q: How much does a used NVIDIA A100 80GB cost in 2026? A: On September 9, 2026, US eBay asking prices for a single A100 80GB PCIe card ran $14,950 to $24,999 with a median of $18,500 across 15 listings. The 80GB SXM4 module ran $8,300 to $12,200, median $9,925, across 4 US listings. These are asks, not sold prices, and they move weekly. Email parts@caladansemi.com for a current quote on tested units.

Q: Why is the A100 SXM4 module so much cheaper than the PCIe card? A: Because almost nobody can use it. The SXM4 module only installs on an NVIDIA HGX A100 baseboard (4 or 8 sockets) or inside a DGX A100. If you don't already own one of those with an open socket, the module is a paperweight, and the price reflects the small pool of buyers. The PCIe card fits any server with a Gen4 x16 slot, a spare 300W, and enough airflow for a passive card.

Q: What is the difference between A100 80GB PCIe and SXM4 specs? A: Same 6,912 CUDA cores, 432 Tensor cores, and 80GB of HBM2e. The SXM4 module runs its memory at 2,039 GB/s and a 400W TDP; the PCIe card runs 1,935 GB/s at 300W. Interconnect differs too: SXM4 gets full 600 GB/s NVLink across the HGX board, while a PCIe card gets 600 GB/s NVLink only to one partner card through a bridge, otherwise PCIe Gen4 at 64 GB/s. Source: NVIDIA A100 datasheet, June 2021.

Q: What should I check before buying a used A100? A: Three nvidia-smi outputs from the actual card: ROW_REMAPPER (no pending remaps, no remapping failure), ECC (zero aggregate uncorrectable errors), and PERFORMANCE under sustained load (no HW thermal or power-brake throttling). Match the board part number on the label to the listing, and get the serial number in the screenshot. Caladan runs these on every unit we ship and includes the output with the quote.

Need a tested A100 80GB with the diagnostics attached? Email parts@caladansemi.com with quantity, form factor (PCIe or SXM4), and the host system it's going into. We'll tell you which of this week's listings we would actually buy, and at what price.

Deciding between the card and the module? Read A100 80GB PCIe vs SXM4: why the used SXM module asks $8,575 less, built from the same September 9 comps.

Frequently Asked Questions

How much does a used NVIDIA A100 80GB cost in 2026?

On September 9, 2026, US eBay asking prices for a single A100 80GB PCIe card ran $14,950 to $24,999 with a median of $18,500 across 15 listings. The 80GB SXM4 module ran $8,300 to $12,200, median $9,925, across 4 US listings. These are asks, not sold prices, and they move weekly. Email parts@caladansemi.com for a current quote on tested units.

Why is the A100 SXM4 module so much cheaper than the PCIe card?

Because almost nobody can use it. The SXM4 module only installs on an NVIDIA HGX A100 baseboard (4 or 8 sockets) or inside a DGX A100. If you don't already own one of those with an open socket, the module is a paperweight, and the price reflects the small pool of buyers. The PCIe card fits any server with a Gen4 x16 slot, a spare 300W, and enough airflow for a passive card.

What is the difference between A100 80GB PCIe and SXM4 specs?

Same 6,912 CUDA cores, 432 Tensor cores, and 80GB of HBM2e. The SXM4 module runs its memory at 2,039 GB/s and a 400W TDP; the PCIe card runs 1,935 GB/s at 300W. Interconnect differs too: SXM4 gets full 600 GB/s NVLink across the HGX board, while a PCIe card gets 600 GB/s NVLink only to one partner card through a bridge, otherwise PCIe Gen4 at 64 GB/s. Source: NVIDIA A100 datasheet, June 2021.

What should I check before buying a used A100?

Three nvidia-smi outputs from the actual card: ROW_REMAPPER (no pending remaps, no remapping failure), ECC (zero aggregate uncorrectable errors), and PERFORMANCE under sustained load (no HW thermal or power-brake throttling). Match the board part number on the label to the listing, and get the serial number in the screenshot. Caladan runs these on every unit we ship and includes the output with the quote. Need a tested A100 80GB with the diagnostics attached? Email parts@caladansemi.com with quantity, form factor (PCIe or SXM4), and the host system it's going into. We'll tell you which of this week's listings we would actually buy, and at what price. Deciding between the card and the module? Read [A100 80GB PCIe vs SXM4: why the used SXM module asks $8,575 less](/blog/nvidia-a100-80gb-pcie-vs-sxm4-used-price-2026), built from the same September 9 comps.

Related Parts

Quick Info

ConditionConfirm in quote
CategoryAI/GPU Compute

Get a Quote

Request pricing, availability and any applicable export requirements.

Email parts@caladansemi.com

Or call us during business hours

Secondary Market

Place a Bid or Ask

Submit your buying or selling requirements for Used NVIDIA A100 80GB Tensor Core GPU (PCIe and SXM4).

Want to Buy
Want to Sell

Confirm export requirements and payment terms before any transaction · parts@caladansemi.com