For AI inference companies
Used GPUs for inference.
Serving models comes down to cost per request, and last generation's GPUs still serve them well. Source used H100s and A100s, PCIe or SXM, and inspect them before you pay.
We reply within one business day.How buying works
Why Inference companies get deals done here
Lower cost per GPU
Used H100s and A100s cost far less than new, which lowers what every request you serve costs.
The form factor you need
PCIe cards for standard servers, SXM in HGX systems for larger models, and A100s with 40 GB or 80 GB.
Checked sellers
Every seller goes through KYC, and you inspect the units before any money moves.
What they buy
- H100 PCIe cards
- H100 NVL cards
- A100 80GB
- A100 40GB
- L40S cards
- HGX H100 servers
How it works
- 01
Tell us what you need
Model, quantity, configuration, and where the hardware will be used.
- 02
Get checked
KYC, confirming the end user, and export screening. Sellers go through KYC too.
- 03
Inspect, pay, ship
Inspect the units, pay, and the hardware ships to you.
Hardware we close deals on
NVIDIA
B300
- Architecture
- Blackwell Ultra
- Memory
- 288 GB HBM3e per GPU
- Form factors
- SXM (HGX)
NVIDIA
B200
- Architecture
- Blackwell
- Memory
- 180 GB HBM3e per GPU
- Form factors
- SXM (HGX)
NVIDIA
H200
- Architecture
- Hopper
- Memory
- 141 GB HBM3e
- Form factors
- SXM, NVL (PCIe)
NVIDIA
H100
- Architecture
- Hopper
- Memory
- 80 GB (SXM, PCIe), 94 GB (NVL)
- Form factors
- SXM, PCIe, NVL
NVIDIA
A100
- Architecture
- Ampere
- Memory
- 40 GB or 80 GB
- Form factors
- SXM, PCIe
And more
L40S, GB200 and GB300 NVL72 racks, AMD MI300X, and every other model.
See all GPUs
Plus whole servers and the networking that goes with them.
Questions people ask
Are used GPUs good for inference?
Yes. Serving a model takes far less compute than training it, so used H100s and A100s run most production models well. Memory matters most: 80 GB cards fit larger models.
Should I use A100s or H100s for inference?
H100s are faster per GPU. A100s cost less on the secondary market. Many teams run both and send each model to the hardware that fits it.
Do PCIe GPUs work for inference?
Yes. PCIe cards fit standard servers, so they're easier to deploy. SXM modules need an HGX board but have faster links between GPUs for larger models.
Can GPUs be sold to buyers outside the US?
Often, yes, depending on the destination and the end user. US export rules restrict where advanced GPUs can go, so every deal has a confirmed end user, screened against those rules, before any money moves.