RentServersBy the RentServers team at Wells Software Consulting LLC · Updated
GB200 specifications
| Architecture | Grace Blackwell |
|---|---|
| GPUs per rack | 72 Blackwell + 36 Grace CPUs |
| Memory per GPU | ~186 GB HBM3e |
| NVLink domain | 72 GPUs, ~130 TB/s aggregate |
| Cooling | Liquid |
| Rack power | ~120 kW class |
Figures for the SXM/HGX version unless noted, from NVIDIA. Dense Tensor Core throughput; NVIDIA's headline numbers with sparsity are double.
Why rack-scale matters
In an 8-GPU server, NVLink stops at the server boundary and everything else goes over InfiniBand. The NVL72 keeps all 72 GPUs in one NVLink domain, which helps very large models that need heavy tensor and expert parallelism, such as trillion-parameter mixture-of-experts training and serving.
How it's priced
Public per-hour prices are rare, and GetDeploying doesn't track the NVL72 as a standard listing. Expect multi-month or multi-year commitments, quoted per rack, with delivery dates tied to data-center build-outs.
What to ask before reserving
- Delivery and acceptance-test dates, and remedies if they slip.
- Whether you get whole racks or a slice of a shared NVLink domain.
- Storage throughput and the fabric between racks.
- Who operates the software stack, and how failed nodes are replaced.
GB200 questions
Can I rent a single GB200 GPU?
Rarely. The NVL72 is designed to be used as a rack, and most providers lease whole racks or large fractions on reserved terms.
Do I need GB200 instead of B200?
Only if your model benefits from a 72-GPU NVLink domain. Most training and inference jobs run well on 8-GPU B200 or H100 servers.