Real hardware.
Root access.
Real GPU nodes with root access and no hypervisor between you and the silicon. Choose a single-tenant cluster on its own fabric, or a multi-tenant option to start smaller, and get full hardware performance either way.
Deploy in as little as 4 weeks / window in writing
- Tenancy
- Single or multi-tenant
- Access
- Root, bare metal
- Fabric
- InfiniBand / RoCEv2
- GPUs
- GB300 / B300 / MI355X / MI350P
- Scale
- 8 GPUs to multi-MW
- Deploy
- In as little as 4 weeks
- Facilities
- Tier II+ and Tier III
B300 and GB300 first.
Every platform is available single-tenant or multi-tenant. Every cluster is built to order and can be handed over in as little as 4 weeks.
NVIDIA GB300 NVL72
Rack-scale Blackwell Ultra: 72 GPUs and 36 Grace CPUs in one NVLink domain, liquid cooled. Built for frontier training, reasoning, and very large models.
NVIDIA HGX B300
8-GPU Blackwell Ultra nodes with 288 GB HBM3e per GPU. The workhorse for large-scale training, fine-tuning, and high-throughput inference.
AMD Instinct MI355X
8-GPU AMD nodes with 288 GB HBM3E per GPU on ROCm. Large memory per GPU for training and serving big models.
AMD Instinct MI350P
AMD MI350-series accelerators in a PCIe form factor on ROCm, suited to inference and fine-tuning in standard air-cooled servers.
What you get.
Dedicated nodes, provisioned to your image
You choose the OS, the CUDA and driver track, and the base image. We provision it, keep it on a tested version track, and hand you root. Nothing is virtualized and nothing is oversubscribed.
Non-blocking GPU fabric
InfiniBand NDR/HDR or RoCEv2 in a rail-optimized, non-blocking topology sized to the cluster. Full-cluster NCCL all-reduce results are part of the acceptance test you sign off on.
Storage that keeps the GPUs fed
Local NVMe scratch on every node plus a shared high-throughput parallel filesystem sized to your dataset and checkpoint pattern, on a separate storage network.
Private networking and isolation
Dedicated VLANs, private interconnect between your nodes, controlled egress, and optional direct connectivity back to your own network or cloud accounts.
Burn-in and acceptance testing
Sustained thermal and load soak, per-link error counters, storage throughput, and GPU health screening before handover. Weak components fail here, not in your first long run.
The technical sheet.
Configured per deployment. These are the ranges we build within; your quote names the exact parts.
- NVIDIA
- GB300 NVL72 / HGX B300
- AMD
- Instinct MI355X / MI350P
- Node form
- 8-GPU HGX / OAM or rack-scale NVL72
- CPU / RAM
- Dual socket, 1-2 TB typical
- Compute network
- InfiniBand NDR/HDR or RoCEv2
- Topology
- Rail-optimized, non-blocking
- Per-GPU bandwidth
- Up to 400 Gb/s
- In-node
- NVLink / NVSwitch
- Management
- Separate out-of-band network
- Local scratch
- NVMe, per node
- Shared
- Parallel filesystem, sized per dataset
- Object
- S3-compatible, optional
- Checkpointing
- Throughput sized to run cadence
- Standard
- Tier II+ or Tier III, certified or equivalent
- Density
- Matched to the cluster, air or liquid
- Cooling
- Air or direct-to-chip liquid
- Redundancy
- N+1 power and cooling
- Geography
- US / Canada / Mexico / Europe / Asia
Where this fits.
Long, tightly-coupled training runs where a single degraded link costs days of wall-clock time.
Teams with security, residency, or audit requirements who want hardware that is theirs alone.
Groups that want the metal and the fabric, and intend to run their own scheduler and tooling on top.
Who carries the pager.
Bare metal includes hardware monitoring and replacement. Want more? Add managed Slurm, managed Kubernetes, or full operations on the same cluster.
- 24/7 monitoring of GPU health, fabric errors, power draw, and thermals.
- On-site sparing and RMA handling, so a failed node is swapped rather than ticketed.
- Firmware, BMC, and driver lifecycle managed on a tested version track.
- Fabric health checks and link-flap remediation before they show up as job failures.
- A named engineer who knows your cluster, reachable directly, not through a queue.
Tell us the shape of the cluster.
An engineer replies with a configuration and a deployment window. No pricing on the website, no sales sequence.
contact@crystalcloud.ai / Los Angeles, CA