Products / GPU Cloud
Run training, fine-tuning, and inference on NVIDIA GPU instances or bare metal GPU servers. Choose flexible terms or reserve capacity for an agreed period.
Access options
The difference is how much of the server your team controls. Commercial terms are a separate choice.
Access option
Virtualized GPU environments with a chosen GPU allocation, CPU, memory, storage, and software image. Start working without managing hardware.
Access option
Dedicated physical servers with no virtualization layer. Your team controls the operating system, drivers, and runtime on the whole host.
Commercial terms
Reserved capacity is a commercial commitment to an agreed quantity, configuration, term, and start date. We plan delivery and expansion around your expected usage. Flexible terms suit work that is still taking shape.
Deployment options and termsNVIDIA systems
GPU count, NVLink domain size, memory, and facility requirements belong in the same decision. We help you match them to your model and workload.
Blackwell Ultra · Server
8 GPUsper server
A server-scale building block. Clusters connect many of these servers over an InfiniBand or Ethernet network.
Grace Blackwell Ultra · Rack
72 GPUs36 Grace CPUs per rack
All 72 GPUs share one NVLink domain, which suits large models and reasoning inference.
Vera Rubin · Rack · Next generation
72 GPUs36 Vera CPUs per rack
NVIDIA’s next-generation rack-scale platform for large-scale pretraining, post-training, and agentic inference. We plan deployments with customers as availability is confirmed.
System names refer to NVIDIA platform families. Configuration, quantity, and delivery timing are confirmed for each engagement.
| System | Architecture and form | Typical fit |
|---|---|---|
| NVIDIA HGX B300 | Blackwell Ultra; 8 GPUs per server connected by fifth-generation NVLink | Modular servers for training, fine-tuning, inference, and scale-out clusters |
| NVIDIA GB300 NVL72 | Blackwell Ultra; 72 GPUs and 36 Grace CPUs in one fully liquid-cooled rack | Large models and reasoning inference that benefit from a 72-GPU NVLink domain |
| NVIDIA Vera Rubin NVL72 | Rubin (next generation); 72 GPUs and 36 Vera CPUs per rack with sixth-generation NVLink | Large-scale pretraining, post-training, and agentic inference; we plan deployments with customers as availability is confirmed |
Running on Hopper today? We can plan H100 and H200 configurations as part of a migration or expansion.
Configuration
Many performance problems on rented GPUs start somewhere other than the GPU, so we size the whole server.
Sized to model weights, activations, precision, context length, and batch size.
Enough headroom for data loading and preprocessing so GPUs aren’t left waiting.
Matched to how your jobs read datasets and write checkpoints.
Images, drivers, frameworks, access, and monitoring agreed before you start.
When jobs outgrow individual servers, move to a dedicated multi-node cluster.
Bring your model, job profile, expected duration, and target start date. We’ll recommend a system, an access option, and terms.
or email contact@arccloudai.com
