Hardware

Dedicated NVIDIA GPUs, down to the metal.

H100, H200, B200, and B300 servers handed over whole. Scale-out clusters can use 400 Gb/s InfiniBand or 800 Gb/s NVIDIA SN5610 Spectrum-X. No hypervisor, no shared lanes, no virtualization tax.

Request capacity
The lineup

Four GPU node families. Eight GPUs each.

Available now

H100 Node

Configured to workload
GPUs
8× NVIDIA H100
GPU memory
80 GB HBM3 per GPU
Interconnect
NVLink (intra-node)
Fabric
InfiniBand or Spectrum-X
Handoff
Single-tenant bare metal
Best fit
Training, inference, and research
Available now

H200 Node

Configured to workload
GPUs
8× NVIDIA H200
GPU memory
141 GB HBM3e per GPU
Interconnect
NVLink (intra-node)
Fabric
InfiniBand or Spectrum-X
Handoff
Single-tenant bare metal
Best fit
Larger memory-bound AI workloads
Available now

B200 Node

SYS-A22GA-NBRT
GPUs
8× NVIDIA HGX B200
GPU memory
180 GB HBM3e per GPU
Interconnect
5th-gen NVLink (intra-node)
NICs
NVIDIA ConnectX-8 SuperNICs
Power draw
~14.5 kW per node
Form factor
Air-cooled rack node
Available now

B300 Node

SYS-822GS-NB3RT
GPUs
8× NVIDIA B300 (Blackwell Ultra)
Interconnect
5th-gen NVLink (intra-node)
NICs
NVIDIA ConnectX-8 SuperNICs
Power draw
~19 kW peak / ~14.5 kW sustained
Density
4 nodes per 80 kW cabinet
Power feed
415 V 3-phase PDUs
Networking

One node or a hundred.

Single nodes run entirely on internal NVLink — no external fabric, no extra cost. When a job needs to span racks, nodes are stitched together with 400 Gb/s InfiniBand or 800 Gb/s NVIDIA SN5610 Spectrum-X so a large cluster behaves like one machine.

Scale-out fabric is provisioned only for multi-node clusters — you don't pay for networking a single node never uses.

  • Intra-nodeNVLink, all 8 GPUs (5th-gen on Blackwell)
  • InfiniBand400 Gb/s non-blocking fabric
  • Spectrum-X800 Gb/s NVIDIA SN5610 Ethernet
  • On-boardFabric NICs included (ConnectX-8 on Blackwell)
  • TopologySized to your training job
Power & cooling

Built for sustained GPU draw.

GPU servers don't idle. GlassPearl deploys into power-dense colocation engineered to feed and cool them at full tilt.

80 kW
Per cabinet
415 V
3-phase PDUs
4
B300 nodes / cabinet
Air
Cooled, no retrofit
Included by default

What every server includes.

Fabric NICs included

Fabric NICs come in the node — Blackwell nodes ship with NVIDIA ConnectX-8 SuperNICs on board. No add-on networking card to buy or wait for.

Full root access

Bare metal, end to end. Bring your own OS, drivers, CUDA, and orchestration. Nothing sits between your stack and the silicon.

Burn-in before handoff

Nodes are racked, cabled, and stress-tested in power-dense colocation. Fabric is validated before you ever log in.

Tell us what you're training.

We'll size the nodes, fabric, and cabinets around your workload.

Request capacity