Buying NVIDIA H100 and H200 GPUs on the Secondary Market: What to Check
By All Chip Supply · Published
When new allocation is tight, many teams look for NVIDIA H100 and H200 GPUs through independent suppliers, system resellers or companies retiring hardware. That can shorten lead times. It also brings risks the original channel handles for you: the wrong variant, missing parts, unclear condition or a card that won't work in your server. This guide lists what to check before you pay and after delivery. It applies to any supplier, including us.
1. Decide which variant you actually need
"H100" and "H200" each cover several products that are not interchangeable. NVIDIA publishes different specifications for each:
- H100 PCIe: a dual-slot, passively cooled PCIe 5.0 card with 80 GB of HBM2e and about 2.0 TB/s of memory bandwidth. See our H100 PCIe listing.
- H100 NVL: a PCIe card with 94 GB of HBM3, designed to be bridged in pairs with NVLink. See the H100 NVL.
- H100 SXM5: a module for HGX H100 baseboards, with 80 GB of HBM3 and 3.35 TB/s. See the H100 SXM5.
- H200 NVL: a PCIe card with 141 GB of HBM3e at 4.8 TB/s. See the H200 NVL.
- H200 SXM: a module for HGX H200 baseboards with the same 141 GB of HBM3e. See the H200 SXM.
An SXM module cannot be used as a PCIe card, and a PCIe card cannot go on an HGX baseboard. Confirm the form factor first, because it decides everything else.
2. Confirm the server will accept it
These GPUs depend on the server around them:
- Passive cooling: the PCIe cards have no fans and rely on the server's directed airflow. They are not meant for desktop workstations.
- Power: check the server's GPU power cables and power supply capacity against the card's maximum power, which NVIDIA lists as configurable for each model.
- Qualification: server vendors publish which GPUs each model supports. A card that is not on the list may not be supported by the vendor, even if it fits.
- Platform for SXM: SXM modules are installed in HGX platforms built by system vendors. Buying modules on their own only makes sense if you already have, or are replacing parts in, a matching platform.
3. Match part numbers and contents
Ask the supplier for the exact part number, and compare it with the manufacturer or OEM documentation. Board partner kits such as PNY's add their own part numbers, and OEM server vendors use theirs. Check what is included: some listings are GPU only, without power adapters or NVLink bridges. The H200 NVL listing on our site, for example, states that it is the GPU only.
4. Get the condition in writing
The secondary market carries new, refurbished and used hardware. A listing should state the condition clearly, and the quote should repeat it. Ask:
- Is the item new in sealed packaging, new without packaging, refurbished or used?
- If used, what was it used for, and for how long?
- What warranty applies, and who provides it? Manufacturer warranties are often tied to the original purchaser or the OEM system, so ask how warranty service would work in practice.
- What happens if the item fails inspection on arrival?
On our site, products are new unless a listing clearly states otherwise, and price and availability are confirmed before any order is accepted.
5. Check the card with nvidia-smi after installation
NVIDIA's `nvidia-smi` tool, included with the data center driver, reports most of what you need to confirm a card is what it claims to be and is healthy. Running `nvidia-smi -q` shows, among other details:
- Product name and serial number, which you can match to the label and the paperwork.
- VBIOS version, useful when comparing cards in one server.
- Total memory, which should match the variant: 80 GB class for H100 PCIe and SXM5, 94 GB class for H100 NVL, 141 GB class for H200.
- PCIe link generation and width, which should match what your server slot supports. NVIDIA notes the current link generation may be reduced while the GPU is idle, so check it under load.
- ECC error counts and remapped rows, which show whether the memory has recorded errors.
Run a sustained workload or a burn-in test and watch temperatures, clocks and power, which `nvidia-smi` also reports. Thermal throttling in a properly cooled server can point to a cooling or hardware problem.
6. Ask for documentation
Useful documents include the original invoice or a supplier certificate of conformance, packing lists with serial numbers, and any test reports. Not every lot comes with full paperwork. Ask before you order, not after delivery, so you can decide whether the documentation available is enough for your use. Our guide on avoiding counterfeit chips covers documentation and inspection in more detail.
7. Plan for software and licensing
Hardware is only part of the picture. Check driver and CUDA support for your operating system, and confirm whether the software you plan to run needs separate licenses. NVIDIA describes its data center software stack on its product pages, so check there for what applies to your deployment.
8. Compare total cost, not unit price
A lower unit price can be offset by shipping, duties, a missing bridge or power adapter, or a server that needs changes. For international orders on our site, shipping, duties and taxes are quoted before your order is confirmed, so you know the full delivered cost before you commit.
Checklist before you request a quote
- The exact variant and form factor (PCIe, NVL or SXM).
- The server or HGX platform it will go into.
- Quantity, and whether you need bridges or power adapters.
- Required condition, warranty and documentation.
- Delivery country.
For background on choosing between these GPUs, read our data center GPU buying guide. You can browse AI and data center hardware, see every NVIDIA product we list, or request a quote with your configuration.