Early access
Whole machines we build and host in Eugene, Oregon, never a slice of one. Tell us what you need and we match one.
| Node | GPU memory | Memory bandwidth | Between GPUs | Form factor |
|---|---|---|---|---|
| NVIDIA H200 NVL | 141 GB HBM3e per GPU | 4.8 TB/s per GPU | NVLink | 4U rack |
| NVIDIA RTX PRO 6000 Server Edition | 96 GB GDDR7 per GPU | 1.6 TB/s per GPU | PCIe | 4U rack |
| NVIDIA RTX 5090 | 32 GB GDDR7 per GPU | 1.8 TB/s per GPU | PCIe | Workstation tower |
Market rate: The median hourly price for this GPU across public listings on other marketplaces, polled live. It is context, not a B3IQ price.
Own one: NVIDIA H200 NVL · Own one: NVIDIA RTX PRO 6000 Server Edition · Own one: NVIDIA RTX 5090
| Detail | Figure |
|---|---|
| square feet | 28,000 |
| carbon-free grid | ~90% |
| redundant power | N+1 |
| monitoring and on-call | 24/7 |
Tour the facility · Security and compliance
Ask for up to 104 GPUs, in whole nodes of 8. Networking between nodes is arranged on request.
Serverless inference, for when you want tokens rather than a machine.
Point your client at our base URL. The rest of your code stays.
We don’t sell them or write them to storage.
A request starts with a short call. Your choices come with it. Book a call · Browse machines
Tell the team what you’re running. They’ll help you size it. Book a call