NVIDIA Data Center GPU H20
The NVIDIA H20 is a specialized data center accelerator based on the Hopper architecture, specifically designed to comply with United States export control limitations for enterprise delivery in mainland China. Equipped with 96GB of ultra-fast HBM3 memory delivering 4.0 TB/s bandwidth and 900 GB/s NVLink interconnect throughput, it addresses large language model inference, dense retrieval-augmented generation (RAG), and parameter-efficient fine-tuning (PEFT). While its raw dense tensor compute is artificially throttled to remain below export thresholds, its massive memory capacity and bandwidth make it highly capable for multi-node LLM serving. Buyers of grey-market or surplus units should verify firmware serialization and HGX baseboard compatibility.
Also written as: NVIDIA H20 SXM5, H20 96GB, NVIDIA HGX H20
Specifications
| Brand | NVIDIA |
|---|---|
| Family | Data Center GPU |
| Form factor | SXM5 module or PCIe |
| Released | Mar 2024 |
| Gpu memory (GB) | 96 |
| Tdp (W) | 400 |
Generations
- Previous generation: NVIDIA Data Center GPU H800 SXM5 80GB
Reference prices (last 14 days, updated 2026-10-27)
New: USD 13,500 (1 sighting, Oct 2026)
Used: USD 10,500 (1 sighting, Oct 2026)
All sightings
| Seller | Condition | Price | Note | Seen |
|---|---|---|---|---|
| serversupply.com | new | USD 13,500 | New OEM SXM5 module | 2026-10-27 |
| ebay.com | used | USD 10,500 | Tested pull from 8-way HGX cluster, fully working | 2026-10-27 |
Prices seen on public listings on the dates shown, before tax and shipping. Not an offer.
Equivalent models from other brands
Buying a used H20: what to check
Confirm NVLink communication links across all 8 nodes in an HGX baseboard, verify that vendor firmware has not locked out regional access, check thermal paste/pad condition on cold plates, and ensure proper 400W/500W power delivery.
FAQ
Why was the NVIDIA H20 created?
It was engineered to comply with revised US export controls for China, reducing peak compute density while providing high 96GB HBM3 memory bandwidth for AI inference.
What is the memory bandwidth of the H20?
It delivers 4.0 TB/s of memory bandwidth across 96GB of HBM3 memory.
Can the H20 be used in standard HGX 8-GPU chassis?
Yes, it is designed for SXM5 HGX 8-way carrier boards as well as dual-slot PCIe server configurations.
Buying or selling H20?
New, refurbished or used. Tell us the quantity and condition and we connect you with a vetted reseller in your region, usually within one working day.
Sources
- NVIDIA Hopper Architecture Overview, checked 2026-10-27