From empty white space
to a cluster that runs.
Full-stack engineering for teams building their own AI clusters. We design power, liquid cooling, rack layout, the GPU fabric, storage, and the orchestration layer, with no hardware to sell you and no vendor to steer you toward.
Retainer or phased engagement / scope in writing
- Scope
- Facility floor to workload
- Vendors
- Strictly agnostic
- Platforms
- GB300, B300, MI355X, MI350P
- Cooling
- Air or direct-to-chip liquid
- Engagement
- Retainer or phased
What you get.
Power and rack architecture
Rack elevations, power distribution, PDU and busway redundancy, and floor loading checks for heavy liquid-cooled racks, designed inside the facility's real limits.
Liquid cooling design
CDU sizing, manifold layout, secondary loop design, and thermal envelopes matched to the platform, so the facility and the hardware agree before anything ships.
Fabric and network topology
Rail-optimized InfiniBand or RoCEv2, non-blocking spine-leaf, optics and cabling bills of materials, and separate management and storage networks.
Storage architecture
Parallel file system and object storage sized for checkpoint bursts and steady training reads, with a clear path to grow.
Scheduler and platform design
Slurm or Kubernetes architecture, identity, observability, and image management designed for how your teams actually run jobs.
Scope and deliverables.
Scoped per engagement. Your proposal names the phases, deliverables, and the engineers on them.
- Power
- Distribution, redundancy, busway
- Cooling
- CDU, manifolds, secondary loop
- Structure
- Floor loading and weight
- Tiers
- Tier II+ and Tier III
- Compute
- InfiniBand or RoCEv2
- Topology
- Rail-optimized, non-blocking
- Optics
- Counts, reach, power budget
- Out-of-band
- Management network
- Parallel FS
- Sized to throughput
- Checkpoints
- Burst ingest planned
- Object
- Datasets and artifacts
- Growth
- Expansion path
- Design
- Architecture document
- Drawings
- Rack elevations, single-lines
- BOM
- Vendor-neutral parts list
- Runbook
- Bring-up and acceptance plan
Where this fits.
Teams buying hardware for the first time at scale who want it designed right before purchase orders go out.
Infrastructure teams that know data centers but not 130 kW liquid-cooled GPU racks.
Providers standing up new GPU halls who need a second set of senior eyes on the design.
How the work runs.
Phased, with a written deliverable at every gate. Stop after any phase, or keep us on retainer.
- Discovery: workloads, growth plan, facility constraints, and budget.
- Architecture: power, cooling, fabric, storage, and platform design with drawings and a parts list.
- Review: design walkthrough with your team and the facility, revised until signed off.
- Build support: answers for integrators and facility teams through delivery.
- Handover: runbooks to your team, or continue with Crystal Cloud operations.
Tell us what you are building.
An engineer replies with a proposed scope and engagement plan. No pricing on the website, no sales sequence.
contact@crystalcloud.ai / Los Angeles, CA