No middleware and no virtualization overhead. These are the current published hourly prices per GPU, together with the complete hardware configurations.
Configurations range from one to eight GPUs. Eight-GPU systems use 3.2 Tbit/s InfiniBand and support distributed training across nodes.
A next-generation training platform. Its 180GB of VRAM accommodates long contexts and large models without splitting model weights.
A balanced option for training and inference. Its 141GB of high-bandwidth memory is especially useful for high-throughput large-model inference.
The industry-standard model, with mature ecosystem and framework support. It is well suited to teams with established training workflows.
These are the current published starting prices for one GPU, in USD per GPU per hour. Actual pricing depends on availability, contract term, and location.
| Model | Hourly price from | GPU memory | CPU platform | Available configurations | System memory | Cluster network |
|---|---|---|---|---|---|---|
| NVIDIA GB200 NVL72 | Pre-order | TBD | TBD | Rack-scale NVL72 | TBD | NVLink + InfiniBand |
| NVIDIA B200 | $3.00 | 180GB SXM VRAM | Intel Emerald Rapids | 1× or 8× B200 | 224 or 1,792 GB DDR5 | 3.2 Tbit/s InfiniBand |
| NVIDIA H200 | $2.30 | 141GB SXM VRAM | Intel Sapphire Rapids | 1× or 8× H200 | 200 or 1,600 GB DDR5 | 3.2 Tbit/s InfiniBand |
| NVIDIA H100 | $2.00 | 80GB SXM VRAM | Intel Sapphire Rapids | 1× or 8× H100 | 200 or 1,600 GB DDR5 | 3.2 Tbit/s InfiniBand |
Reserve the GB200 NVL72 for access to next-generation AI performance. This rack-scale system uses NVLink to operate the rack's GPUs as one compute unit, supporting trillion-parameter model training and large-scale inference.
Storage is billed only for capacity used, with no extra charges for uploads, downloads, expansion, or migration. Capacity uses binary units: 1 GB = 2^30 bytes (GiB).
Confirm the model, GPU count, expected training duration, multi-node interconnect needs, data residency, and compliance requirements.
We provide the actual unit price and delivery date based on the contract term. For terms of 3 months or longer, we also compare reserved options.
After signing, we initialize accounts, segment the network, configure firewall rules, mount storage, and grant access.
Before handover, we validate node and interconnect performance and provide the test results. The service then moves into 24×7 operations and incident reporting.
Tell us the model, quantity, and contract term, and we will provide available configurations and pricing.
