HKCHL
NVIDIA H100 official key visual — Hopper GPU module on dark background
GPU Servers
80GB HBM3 Transformer Engine PCIe / SXM5 Refurbished · Tested · In Stock

NVIDIA H100 80GB

The production standard for serious AI: Hopper's Transformer Engine, FP8 precision, and 80GB of HBM3 — the card that trains and serves the models everyone else benchmarks against. Blackwell is here, but H100 supply has finally normalized and the software stack is boring-in-a-good-way mature. New and tested units ship from Hong Kong.

Request a Quote 2026 reference: PCIe new $25,000-$32,200 — INDICATIVE, small sample

Fully Tested & Documented  ·  Hong Kong Warehouse Stock  ·  Worldwide Shipping 3–7 Days  ·  1-Year Warranty

Hardware Gallery

The H100, Up Close

NVIDIA H100 official render — Hopper module with HBM stacks and die shot styling

Official NVIDIA Key Visual

NVIDIA's official H100 key visual: the Hopper package with HBM stacks flanking the GH100 die. The SXM5 module shown here is the 700W, 3.35TB/s variant; the PCIe card (350W, 2.0TB/s) shares the same silicon.

NVIDIA H100 Hopper die and package render — GH100 die surrounded by HBM stacks on the SXM module

Hopper Package, Up Close

Close render of the H100 SXM module: GH100 die with six HBM stacks. FP8 Transformer Engine and 4th-gen Tensor cores are what separate it from the A100 generation — roughly 3x AI throughput per card.

80

GB HBM3 (SXM5) / HBM2e (PCIe)

3.35

TB/s bandwidth, SXM5

1,979

TFLOPS FP8 tensor, SXM5

900

GB/s NVLink, SXM5

Key Specifications

Component Specification
ModelNVIDIA H100 80GB — SXM5 (900-21010-0000-000) and PCIe variants
ArchitectureHopper GH100, 132 SM (SXM5) / 114 SM (PCIe)
Memory80GB HBM3 3.35 TB/s (SXM5) / 80GB HBM2e 2.0 TB/s (PCIe)
Power700W (SXM5) / 350W (PCIe), passive
ComputeFP8 1,979 TFLOPS / FP16 989 TFLOPS (SXM5, with sparsity)
Transformer EngineFP8 training and inference, 4th-gen Tensor cores
NVLink900 GB/s GPU-to-GPU (SXM5); NVLink bridge (PCIe)
MIGUp to 7 instances (10GB each)

Two form factors, one GPU: SXM5 (700W, 3.35TB/s, full NVLink) for HGX training nodes, and PCIe (350W, 2.0TB/s) for standard 4U/5U servers. The step down from the H100 is the A100 40GB — roughly one-third the AI throughput at one-fifth the 2026 price; the step sideways for inference fleets is the L4. Deployment patterns: AI GPU server buying guide; full lineup: GPU server range.

Transparent Pricing

2026 Reference Pricing by Configuration

Tracked Sep 7, 2026 — four asking-price data points for the PCIe 80GB variant. The H100 market in 2026 is post-shortage but not cheap: supply normalized while HBM costs climbed, so prices have plateaued rather than fallen. SXM5 units typically quote $35,000-$40,000.

INDICATIVE RANGE

New PCIe 80GB

$25,000-$32,200

ITCT Shop guides $25,000-$30,000; GPU Vendors lists from $32,200 in stock; Jarvislabs pegs the market near $25,000 (Mar 2026). EU channel quote seen at EUR 50,000 (approx $54,000) via ANAFRA — wide regional variance.

Used / Pre-Owned

approx $28,000

eBay pre-owned PCIe 80GB (900-21010-0000-000) asking $27,999.99. Used H100 pricing barely undercuts new — demand still outruns secondary supply.

Asking prices, not transaction prices — volume deals typically close 5-15% below asking. Very small sample (n<5): INDICATIVE only. Sources: itctshop.com, gpuvendor.com, jarvislabs.ai, smicro.eu, ebay.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 H100 Ships

1

HBM3 Stack Verification

Full 80GB HBM exercised with ECC telemetry logged — six stacks on SXM5, five on PCIe; any error history is disclosed with the unit.

2

Tensor & FP8 Stress

Sustained Transformer-Engine FP8 load at rated TDP while logging clocks, power, and thermals — throttling or power-limited modules are rejected.

3

NVLink Fabric Test

On SXM5 units, NVLink lane validation against a reference baseboard; on PCIe cards, bridge link and bandwidth checks.

4

Firmware & VBIOS Check

Firmware and VBIOS verified against NVIDIA's current data-center branch — reflashed or miner-modified boards are screened out.

5

Serial & Provenance Check

Serial number and P/N verified against board markings; engineering samples and known-flagged serial ranges excluded. Export-control documentation prepared per destination.

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 H100 80GB cost in 2026?

New PCIe 80GB cards ask $25,000-$32,200, with one EU channel quote as high as EUR 50,000 and used units near $28,000 (four data points tracked Sep 7, 2026 — very small sample, INDICATIVE only). SXM5 modules typically quote $35,000-$40,000. Prices have plateaued this year: supply normalized, but HBM3 costs keep the floor high.

H100 vs A100 — is the upgrade worth it in 2026?

For training and 70B+ serving, yes: the H100's Transformer Engine, FP8, and 3.35TB/s HBM3 deliver roughly 3x the A100's AI throughput, and FP8 halves memory pressure. For inference of smaller models, the math is less kind — an A100 40GB at $3,150-$5,870 delivers far more tokens per dollar. Rule of thumb we quote: training or flagship serving, H100; budget inference fleets, A100 or L4.

Should I buy H100 PCIe or SXM5?

PCIe (350W, 2.0TB/s) drops into standard GPU servers and is the only option for most existing chassis. SXM5 (700W, 3.35TB/s, 900GB/s NVLink) requires an HGX baseboard but unlocks full bandwidth and 8-GPU NVLink fabrics — the training-grade option. If you are building new HGX nodes, buy SXM5; if you are upgrading existing PCIe servers, the PCIe card is the practical path.

What condition are HKCHL H100 units in?

New-sealed and tested pre-owned units, each with a full burn-in report (HBM ECC, FP8 stress, thermal log) and a 1-year warranty. Serials are verified and disclosed. We do not deal in engineering samples or reflashed boards.

Can you ship NVIDIA H100 internationally?

Yes, subject to export-control regulations for high-performance GPUs — these rules vary by destination and end use, and we handle the licensing paperwork before quoting. Express air-freight typically arrives in 3-7 days to eligible markets; HGX node orders ship assembled and factory-tested.

Get a Current H100 80GB Quote

Tell us form factor (PCIe or SXM5), quantity, and destination — we confirm allocation, handle export documentation, and ship worldwide from Hong Kong, usually within one business day of cleared paperwork.

Request a Quote