GPU Capacity.
Bare-metal NVIDIA GPU capacity billed hourly, for training, fine-tuning, inference, model serving, and BYOM workloads.
Choose your compute
$3.50
Per hour
H100
Designed for large-scale model training, fine-tuning, and high-throughput inference.
SXM5 · NVLink · 1 to 8 GPUs
$3.50
Per hour
$3.50
Per hour
H200
More memory for running large language models with longer context and larger batch sizes.
SXM5 · NVLink · 1 to 8 GPUs
$3.50
Per hour
Capacity
Dedicated servers, not spot.
Dedicated GPU servers allocated to the workload. Not spot capacity. Run your own models, containers, serving stack, or training jobs on Origon infrastructure.
NVIDIA HGX H100
- GPU Count
- 8
- GPU Ram
- 640 GB
- CPU (x2)
- Intel® Xeon® Platinum 8468
- System RAM
- 2,048 GiB
- NVMe Storage
- 61.44 TiB
NVIDIA HGX H200
- GPU Count
- 8
- GPU Ram
- 1,128 GB
- CPU (x2)
- Intel® Xeon® Platinum 8592+
- System RAM
- 2,048 GiB
- NVMe Storage
- 61.44 TiB
AMD MI300X
- GPU Count
- 8
- GPU Ram
- 1,536 GB
- CPU (x2)
- Intel® Xeon® Platinum 8568Y
- System RAM
- 2,048 GiB
- NVMe Storage
- 61.44 TiB
Spec
NVIDIA HGX H100
NVIDIA HGX H200
AMD MI300X
GPU Count
8
8
8
GPU Ram
640 GB
1,128 GB
1,536 GB
CPU (x2)
Intel® Xeon® Platinum 8468
Intel® Xeon® Platinum 8592+
Intel® Xeon® Platinum 8568Y
System RAM
2,048 GiB
2,048 GiB
2,048 GiB
NVMe Storage
61.44 TiB
61.44 TiB
61.44 TiB
Workloads
The full training-to-serving path.
Train, fine-tune, and serve your models on dedicated GPU infrastructure, from first run to production.
Training
Full training runs on dedicated GPUs.
Fine-tuning
Adapt base models to your own data.
Inference
Serve model predictions at scale.
Model serving
Host your own model endpoints.
Customer models
Run the models you bring.
BYOM workflows
Your stack on Origon capacity.
Service Pairings
Compose with the rest of AI Cloud.
Use GPU Capacity with AI Datastore, Speech, and Voice Network when the workload needs datastore, real-time speech, or telephony.
