Training & fine-tuning
Custom training loops, experiment stacks, and frameworks you install yourself.
Who it's for
Self-hosted inference
vLLM, TensorRT-LLM, or your own serving layer on Linux with CUDA.
HPC & visual compute
Scientific simulation, 3D render, and visualization that needs real GPU memory.
Platform & agentic AI
Agent runtimes, pipelines, and multi-tenant GPU pools under your control.
What you get
Full root access
SSH in as root. Install drivers, frameworks, and serving stacks your way.
96 GB GDDR7
ECC memory on RTX PRO 6000 Blackwell Server Edition for large models and scenes.
CUDA-ready images
GPU Linux images with CUDA, or a base OS if you bring your own stack.
MIG up to 4
Partition one GPU into isolated instances with dedicated memory and compute.
Same stack as Compute
VPC, security groups, public and private IP, and the same bill as the rest of the platform.
Per GPU-hour
Hourly meter on the GPU host. Stop the instance, stop the charge.
NVIDIA RTX PRO 6000 Blackwell Server Edition
Universal AI and visual computing for the data center: the same Blackwell Server Edition silicon NVIDIA positions for agentic AI, scientific compute, render, and media.
Blackwell
Architecture
NVIDIA RTX PRO 6000 Server Edition
96 GB
GPU memory
GDDR7 with ECC
1,597 GB/s
Memory bandwidth
512-bit memory interface
Up to 4
MIG
Isolated GPU instances
From config to live
Select region
Miami (US1) at launch. Pick the datacenter for the workload.
Configure GPU instance
GPU plan, CPU/RAM bundle, network, security group: hourly cost before confirm.
Select image
GPU-ready Linux with CUDA, or a base OS if you bring the stack.
Deploy
Public and private IP. Manage it like any Cloud Compute instance.
Pricing locks with SKUs
Per GPU-hour rates and host bundles publish when the SKUs lock. Size the rest of the stack in the calculator meanwhile.
Short answers
Which GPU is this?
NVIDIA RTX PRO 6000 Blackwell Server Edition: 96 GB GDDR7 ECC, up to four MIG instances, positioned for universal AI and visual computing in the data center.
Do I get root access?
Yes. GPU Cloud is a full-control instance, not a managed endpoint. You SSH in, install what you need, and own the runtime.
How is this different from managed inference?
GPU Cloud rents you the silicon with root. Managed inference products serve models for you. Use GPU Cloud when you need custom training, fine-tuning, or your own serving stack.
When do rates publish?
Per GPU-hour rates and host bundles publish when the SKUs lock. Until then, size storage, compute, and networking in the public calculator.
Everything your stack needs
Everything else on the same console, the same network, and the same bill.
- Cloud ComputeInstances live in minutes, with hourly pricing before you deploy.
- Block StorageAttach high-performance NVMe volumes as your data grows.
- Object StorageS3-compatible buckets for media, backups, and datasets.
- BackupAutomated backups on a schedule, priced as a clear share of the instance.
- SnapshotsOn-demand point-in-time images before upgrades and risky deploys.
- VPC NetworksPrivate networks your instances join, keeping east-west traffic off the public net.
- IP AddressesReserved public IPs that stay with you when the instance changes.
- Security GroupsPlatform firewalls: ingress and egress rules across the fleet.
- DNSDomains and records next to Compute, balancers, and reserved IPs.
- KubernetesWe run the control plane. Live in Private Cloud, self-serve next.
- Managed DatabasesManaged databases with backups, patching, and failover.
- Load BalancersDistribute traffic across backends with health checks and TLS.
- Bare MetalDedicated hosts when virtualization is not enough.