Whole-server handoff
You get the entire machine — bare metal, full root, no hypervisor tax and no neighbors sharing your PCIe lanes. What the hardware can do, you can do.
Bare-metal GPU infrastructure
Whole servers. Not slices.
Raw Metal hands you dedicated NVIDIA Blackwell servers — full root, no hypervisor, no noisy neighbors. Committed capacity on 36-month terms, billed at 100%. The instance roulette ends here.
- 8×
- GPUs per node
- 100%
- Billed utilization
- 36-mo
- Committed terms
The model
Compute you own the whole of.
Most GPU clouds rent you a slice of a machine and bill you for the privilege of guessing. Raw Metal does the opposite — the whole server, a fixed price, for as long as you need it.
Billed at 100%
Flat, committed pricing on the full server. No per-instance metering, no utilization surprises, no scramble when spot capacity evaporates mid-run.
Committed for the long run
36-month terms give your training and inference fleets a stable home. Reserve capacity once; stop re-bidding for GPUs every quarter.
Built for AI workloads
Latest-generation NVIDIA Blackwell, InfiniBand fabric for multi-node clusters, and power-dense cabinets engineered for sustained GPU draw.
Hardware
NVIDIA Blackwell, whole and dedicated.
Eight GPUs to a node, ConnectX-8 SuperNICs on board, and InfiniBand fabric for clusters that need to act as one. No virtualization layer between your code and the silicon.
Available now
8 GPU B200 Node
- GPUs
- 8× HGX B200
- Memory
- 180 GB HBM3e / GPU
- Power
- ~14.5 kW / node
- Chassis
- SYS-A22GA-NBRT
Flagship
8 GPU B300 Node
- GPUs
- 8× B300 (Blackwell Ultra)
- Cooling
- Air-cooled, 80 kW cabinets
- Power
- ~19 kW peak / node
- Chassis
- SYS-822GS-NB3RT
InfiniBand fabric
Single nodes run on internal NVLink. Multi-node clusters are stitched together with non-blocking InfiniBand — copper or optical — so large training jobs scale across racks without a network bottleneck.
How it works
From spec to root access.
- 01
Scope your fleet
Tell us the GPU count, node type, and cluster topology your workload needs. We size the cabinets and fabric around it.
- 02
Reserve capacity
Lock in dedicated servers on a 36-month term at a flat, committed rate. Capacity is yours from day one — no queue, no spot bidding.
- 03
We rack & burn-in
Nodes are deployed, cabled, and stress-tested in power-dense colocation. InfiniBand fabric is validated before you ever log in.
- 04
Keys handed over
You receive bare-metal access to the whole server. Bring your own stack — root is yours, top to bottom.
Colocation
Own the hardware already?
If you'd rather bring your own GPUs, Raw Metal brokers power-dense colocation — sourcing the cabinets, power, and InfiniBand-ready space your cluster needs, and matching you to the right facility.
Talk colocation- Power-dense cabinetsUp to 80 kW per cabinet for sustained GPU draw.
- Fabric-readySpace provisioned for InfiniBand clustering.
- Right facility, right termsWe match capacity to your deployment timeline.
Request capacity
Reserve your bare metal.
Tell us what you're training. We'll come back with hardware, topology, and a committed quote — no instance roulette, no surprises.