NVIDIA A100 40GB
The workhorse that taught the industry AI: Ampere with 40GB of HBM2, 1,555GB/s of bandwidth, and MIG to slice one card into seven. Production ended in 2024, which is exactly why 2026's secondary market is where the value is — same silicon, a fraction of the launch price. Tested PCIe and SXM4 units ship from Hong Kong.
Fully Tested & Documented · Hong Kong Warehouse Stock · Worldwide Shipping 3–7 Days · 1-Year Warranty
Hardware Gallery
The A100 40GB, Up Close
Official PCIe Card Render
NVIDIA's official render of the A100 PCIe card: dual-slot, passively cooled, 250W. The PCIe variant drops into standard GPU-capable 2U/4U servers — no SXM baseboard required.
SXM4 Module (Variant)
The same GA100 GPU in SXM4 mezzanine form — 400W, NVLink-enabled, for HGX baseboards. We stock both form factors; tell us your platform and we quote the right one.
40
GB HBM2 with ECC
1,555
GB/s memory bandwidth
312
TFLOPS FP16 tensor
7
MIG instances per card
Key Specifications
| Component | Specification |
|---|---|
| Model | NVIDIA A100 40GB PCIe (900-21001-0000-000); SXM4 variant available |
| Architecture | Ampere GA100, 6,912 CUDA cores, 432 Tensor cores (3rd gen) |
| Memory | 40GB HBM2 ECC, 1,555 GB/s |
| Interface | PCIe Gen4 x16, dual-slot passive (250W); SXM4 400W |
| Compute | FP16 312 TFLOPS / TF32 156 TFLOPS / FP32 19.5 TFLOPS |
| MIG | Up to 7 instances (5GB each) |
| NVLink | PCIe: 2-link bridge; SXM4: 12 links, 600 GB/s |
| Status | Production ended Feb 2024 — secondary market supply |
The A100 40GB is the budget door into serious AI silicon: identical GA100 compute die as the 80GB, with half the HBM. If your models fit — LoRA/QLoRA fine-tunes and quantized inference up to ~13B — you keep 100% of the compute at roughly half the card price. For 70B-class serving or long context, the answer is the 80GB or the H100. For pure inference fleets, compare the far cheaper L4. More context in our AI GPU server buying guide and the GPU server range.
Transparent Pricing
2026 Reference Pricing by Configuration
Tracked Sep 7, 2026 — seven data points from eBay's live A100 40GB market plus a secondary-market tracker. This is a post-production GPU: there is no new retail channel, so condition grading and seller reputation drive the spread.
eBay Secondary Market
$3,150-$5,870
Six live asking prices: $3,149.99, $4,250, $4,575 (China, new claim), $4,977, $5,677.84 (UK), $5,869.99. Condition ranges from pulls to claimed-new; verify serials and hours.
Dealer / Refurbished Band
$5,500-$7,900
Clustermelt's secondary-market tracker (Aug 1, 2026 snapshot) puts dealer and refurbished A100 PCIe 40GB at $5,500-$7,900 with warranty. The premium over eBay buys testing and a return path.
Asking prices, not transaction prices — volume deals typically close 5-15% below asking. Small sample per tier (n<5): indicative only. Sources: ebay.com, clustermelt.com. Collected Sep 7, 2026. Full dataset: Used Server Price Index.
Origin
Hong Kong Warehouse
Free port — full export documentation
Transit
3–7 Days Air Freight
Tracked, insured, worldwide
Warranty
1-Year Warranty
Diagnostic report in every box
Quality Process
What We Test Before an A100 40GB Ships
HBM ECC Scrub
Full 40GB HBM exercised with ECC error counters logged — memory health is the first thing that degrades on ex-datacenter cards.
Tensor Core Stress
Sustained FP16/BF16 GEMM load at rated TDP while logging clocks, power, and thermals — throttling cards are rejected.
NVLink / PCIe Link Test
PCIe Gen4 x16 link width and, on SXM4 units, NVLink lane validation under bandwidth test.
MIG Partition Check
MIG enabled and all seven instances created and destroyed cleanly — a quick proof the GPU is fully functional.
Serial & Provenance Check
Serial number verified against NVIDIA's format and board markings; known-stolen and engineering-sample serial ranges excluded.
Report in the Box
Every unit ships with its per-machine diagnostic report and warranty. Full process: inspection guide.
Frequently Asked Questions
How much does an NVIDIA A100 40GB cost in 2026?
The secondary market runs $3,150-$5,870 on eBay (six live asking prices tracked Sep 7, 2026) and $5,500-$7,900 through dealers with warranty, per Clustermelt's tracker — small samples, indicative only. Production ended in February 2024, so every A100 is secondary-market now; grading and warranty are what you are really buying.
A100 40GB vs A100 80GB — which one do I need?
Same GA100 die, same 6,912 CUDA cores and FP32 compute — the difference is memory: 40GB HBM2 at 1,555GB/s versus 80GB HBM2e at 1,935-2,039GB/s. Pick the 40GB for LoRA/QLoRA fine-tuning and quantized inference on models up to ~13B parameters; pick the 80GB to serve 13B-70B models, run context beyond 32K tokens, or hold weights in full precision. Per dollar, the 40GB is the better buy whenever the model fits.
A100 40GB or H100 80GB for a new cluster?
The H100 delivers roughly 3x the AI throughput with FP8 and the Transformer Engine — but at $25,000-$32,000 per PCIe card in 2026, it costs 4-6x the A100. For inference-heavy, budget-constrained clusters, A100 40GB cards still win on throughput per dollar; for training or for serving 70B+ models at production latency, the H100 pays for itself. We stock both and can quote mixed fleets.
What condition are HKCHL A100 cards in?
Graded secondary-market stock — pulls and dealer-refurbished units, each with a burn-in report (HBM ECC scrub, tensor stress, thermal log). Every card carries a 1-year warranty, and we disclose ECC error history before you pay.
Can you ship NVIDIA A100 to the US, Europe, or Asia?
Yes, subject to export-control rules for high-performance GPUs — we handle documentation and confirm destination eligibility before quoting. Express air-freight typically arrives in 3-7 days to eligible markets; multi-card orders can ship pre-installed and tested in a server chassis.
Get a Current A100 40GB Quote
Tell us quantity, form factor (PCIe or SXM4), and workload — we confirm live graded stock and ship worldwide from Hong Kong, usually within one business day.
Request a Quote