Capacity
Two ways to rent.
Every server is single-tenant bare metal: no virtualization, no shared GPUs.
Enterprise reserved
A dedicated cluster on a fixed term
- Whole 8-GPU servers, reserved for 12 to 36 months
- Fixed rate per GPU-hour for the full term
- Servers networked as one cluster for multi-node training
- Capacity held with a down payment at signing
On demand
B300 by the hour
- Rent single servers when you need them
- Pay only for the hours you run
- Root access to the full machine
- Move to a reserved term at any time
Hardware
Built for the largest models.
Memory
2.3 TB per server
288 GB of HBM3e on each of eight GPUs, so large models fit on fewer machines.
Interconnect
NVLink 5 inside the server
All eight GPUs share a single NVLink domain at 1.8 TB/s per GPU.
Operations
Racked by our own team
We install, network and support the servers ourselves, so you deal directly with the people who run them.
How it works
From inquiry to running jobs.
- Tell us the shapeGPU count, term and start date.
- Agree the termWe quote a fixed rate and confirm the delivery window.
- We rack and networkOur team installs and burns in the servers.
- You runRoot access to bare metal, with direct support from the people who built it.
Contact
Reserve capacity or ask a question.
Send your GPU count, term and target start date, and we will reply with availability and a quote.