Your own GPU cluster.
Bare metal. Root access.
Choose NVIDIA GB300 NVL72, NVIDIA B300, AMD MI355X, or AMD MI350P hardware. No hypervisor, full performance, your stack. Want us to run it? Add managed Slurm, Kubernetes, or 24/7 operations.
Single-tenant or multi-tenant, your choice, with no hypervisor on bare metal. Deploy in as little as 4 weeks into Tier II+ and Tier III facilities across the US, Canada, Mexico, Europe, and Asia.
Dedicated nodes on your own fabric, or a multi-tenant option to start smaller at lower cost.
Run your own stack, or add managed Slurm, Kubernetes, or 24/7 operations on the same hardware.
Clusters built to order and handed over in as little as 4 weeks, with your deployment window in writing.
Start with the hardware. Add what you want us to run.
Bare metal is the product: dedicated GPUs with root access. Managed Slurm, managed Kubernetes, and operations are optional layers on the same cluster, for teams that would rather not carry the pager.
Bare Metal GPU
The core product. GPU nodes with root access, single-tenant or multi-tenant, and no virtualization tax. Non-blocking InfiniBand or RoCEv2, NVMe local scratch, and shared high-throughput storage.
- Tenancy
- Single-tenant or multi-tenant
- Platforms
- GB300 NVL72 / B300 / MI355X / MI350P
- Fabric
- InfiniBand or RoCEv2, non-blocking
Managed Slurm
Optional add-on. A research-grade scheduler stood up, tuned, and operated for you: partitions, QoS, fair-share accounting, container support, and job telemetry.
- Scheduler
- Slurm, tuned per workload
- Containers
- Pyxis / Enroot / Apptainer
- Included
- Accounting, telemetry, upgrades
Managed Kubernetes
Optional add-on. Production Kubernetes with the NVIDIA GPU operator, autoscaling node pools, ingress, and full observability. Cluster lifecycle and upgrades are ours to carry.
- Distribution
- Upstream-conformant K8s
- GPU
- GPU operator, MIG or time-slicing
- Included
- Observability, ingress, lifecycle
Operations & Maintenance
Optional add-on. 24/7 NOC coverage, hardware sparing and RMA, firmware and driver lifecycle, fabric health, and thermal monitoring, on our fleet or on hardware you already own.
- Coverage
- 24/7 NOC, SLA-backed
- Hardware
- On-site sparing and RMA handling
- Scope
- Our fleet or your own
Configured to the workload.
Three shapes cover most of what we deploy. Every one can be delivered single-tenant or multi-tenant, burned in, and acceptance tested before handover.
Training block
For pre-training and large-scale training runs that need tightly-coupled, high-density compute.
- Hardware
- GB300 NVL72 rack-scale or B300 HGX
- Fabric
- InfiniBand or RoCEv2 non-blocking
- Density
- High density, liquid cooled
- Managed layer
- None, or Slurm if wanted
Inference fleet
For serving models in production, where scale and uptime matter more than peak interconnect.
- Hardware
- B300 HGX, AMD MI355X, or AMD MI350P
- Cooling
- Air or liquid
- Scale
- 64-1,024 GPUs
- Managed layer
- None, or Kubernetes if wanted
Fine-tuning cluster
For teams adapting existing models on a modest, air-cooled footprint with a shorter commitment.
- Hardware
- B300 HGX or AMD MI350P, 8-128 GPUs
- Cooling
- Air-cooled friendly
- Managed layer
- Optional
256 B300 GPUs on InfiniBand, bare metal with root access, single-tenant hall.
GB300 NVL72 rack-scale training block, liquid cooled, with optional O&M added.
512-GPU MI355X inference fleet, bare metal with managed Kubernetes added.
Illustrative configurations to gauge fit. Customer identities and pricing are never disclosed publicly.
Nothing is handed over untested.
A cluster that passes a smoke test is not a cluster that survives a three-week training run. Four stages, each with an artifact you receive.
DESIGNConfigured against the workload
Node count, GPU platform, fabric topology, storage throughput, power density, and the managed layer, sized to the jobs you actually intend to run. Written down before anything is ordered.
CONFIGURATION SHEET
DEPLOYMENT WINDOW
BUILDRacked, cabled, burned in
Hardware racked and cabled in a qualified facility matched to your density and geography, then burned in under sustained load until the weak components fail here rather than in your first run.
BURN-IN REPORT
CABLING AND TOPOLOGY MAP
VALIDATEAcceptance tested end to end
NCCL all-reduce across the full cluster, per-link bandwidth and error counters, storage throughput, thermal soak, and scheduler smoke tests. You see the numbers before you accept the cluster.
NCCL / FABRIC RESULTS
ACCEPTANCE CRITERIA SIGN-OFF
OPERATERun, monitored, kept current
24/7 monitoring of GPUs, fabric, power, and thermals. Sparing and RMA handled on site, firmware and drivers kept on a tested track, and a named engineer who knows your cluster.
24/7 NOC AND SLA
MONTHLY HEALTH REPORT
This is what prevents silent throughput loss, mystery job failures, and the slow drift that turns a fast cluster into an average one.
Your cluster picks the building.
We deploy into independent Tier II+ and Tier III facilities qualified for your specific cluster: your density, your geography, your term. Every facility clears the same bar before it carries a deployment.
Tier II+ / III
N+1
Density matched
Carrier-neutral
- Tier certification / design dossier
- REVIEWED
- Power headroom / N+1 topology
- REVIEWED
- Density and cooling limits
- DOCUMENTED
- Fiber providers and latency
- DOCUMENTED
- Remote hands and escalation path
- ON FILE
- Site walk
- BEFORE DEPLOYMENT
Facility detail shared under NDA / metros disclosed there
Live in as little as 4 weeks.
Every cluster is built to order for your workload. From a signed configuration, handover can come in as little as 4 weeks, and your quote puts the deployment window in writing.
Requirements and configuration
One working session: workload, scale, density, interconnect, storage, term length, and the managed layer you want. You leave with a written configuration.
Quote and cluster spec
A configuration sheet with the numbers in it: node counts, fabric, storage, the managed layer, the SLA, and the deployment window. Every number is in writing before you commit.
Build, validate, handover
Racked, burned in, and acceptance tested in a qualified facility. Bare-metal access, your scheduler or ours, and a direct line to the engineers who stood it up. Handover can come in as little as 4 weeks, on the date written into your quote.
Choose single-tenant when your workloads must never share hardware, or multi-tenant to start smaller. Each facility's security attestations, such as SOC 2 or ISO 27001 where held, are reviewed during qualification and shared under NDA. Private networking and dedicated storage.
Already own the hardware?
Keep your hardware and hand us the operations. For fleets you already own, we run deployment, scheduling, monitoring, sparing, firmware, and the on-call rotation under an SLA, in your facility or one we qualify for you.
Questions.
What does Crystal Cloud actually provide?
Dedicated bare-metal GPU clusters on four platforms: NVIDIA GB300 NVL72, NVIDIA B300, AMD MI355X, and AMD MI350P. You get root access to your own hardware on a high-speed fabric. If you want us to run it, add managed Slurm, managed Kubernetes, or 24/7 operations and maintenance.
Is the hardware shared with other customers?
Only if you choose to. Pick single-tenant bare metal, where the nodes, the fabric, and the storage are yours alone for the term, or a multi-tenant option when you want a smaller footprint and lower cost.
How fast can a cluster be deployed?
Deployments start in as little as 4 weeks from a signed configuration. Every cluster is built to order for your workload, and your quote states the deployment window in writing, so you can plan around a real date.
Why is there no pricing on the site?
Every cluster is configured to the workload, the density, and the term, so the numbers live in a written quote rather than on a webpage. Tell us the shape of the cluster and an engineer replies with a configuration and a deployment window.
Start with the cluster you need.
I need a cluster
CONFIGURATION AND WINDOW IN WRITING
I have hardware to run
24/7 NOC / SPARING / SLA
I have capacity
WE DEPLOY OUR CUSTOMERS DIRECTLY
contact@crystalcloud.ai / Los Angeles, CA