01CSTM-C
AI as an internal service
One pool of GPU capacity for many departments, each with its own quotas and access rules.
CSTM-C · REV AcstmAI™ · Hardware
A GPU cluster for the largest models and the most users.
cstmAI™ Cluster joins two or more GPU servers with a high-bandwidth fabric, shared storage and a scheduler, for organizations that need the biggest open-weight models, thousands of users, or regular training runs on their own hardware.
Fig. 01 · Materials test laboratory, NASA Lewis
A cluster is a different project from a server. The network fabric, storage throughput, power density and cooling decide whether the GPUs are busy or waiting, and each one is a design decision. We design all of it, and plan the facility work with your team or your colocation provider before anything ships.
On the software side, the cstmAI Platform schedules inference and training across nodes, so capacity is shared instead of stranded on one box. You grow by adding nodes of the same design.
| Form factor | One or more full racks (42U–48U) |
|---|---|
| Nodes | 2 or more GPU servers, typically 8 GPUs each |
| GPU memory | 80–192 GB per GPU |
| Fabric | 200–400 Gb/s per GPU; InfiniBand or RoCE Ethernet |
| Storage | Shared high-throughput storage, tens to hundreds of TB |
| Typical models | The largest open-weight models at full precision; many models at once |
| Concurrent users | Hundreds to thousands |
| Power & cooling | Often 20–40 kW or more per rack; liquid cooling at the high end |
| Setting | Data center or colocation |
Typical ranges. The quote names every part.No list prices
Front elevation
The drawing before the build.
A representative elevation. The chassis we specify varies by vendor and generation; the callouts don't.
Fig. 1 CSTM-C · REV A · Front elevationNot to scale
01
Fabric switches
02
GPU nodes, 8 GPUs each
03
Rear power distribution (hidden)
04
Shared storage
Built for
Capacity you share, schedule and grow.
02CSTM-C
Large-model serving
The biggest open-weight models at full precision, for the tasks smaller models can't handle.
03CSTM-C
Training at scale
Full fine-tunes and continued training on large private datasets, without renting GPUs by the hour.
04CSTM-C
Batch work on archives
Processing millions of documents, records or images in days instead of months.
Reel 02 · Instrument assembly in a clean room, NASA JPL2021
How it's delivered
Four steps, each one signed off.
01
Capacity plan
Users, models, training cadence and growth over three years, turned into node counts.
02
Facility & fabric design
Power, cooling, rack layout, network topology and storage, agreed with your facilities team.
03
Staged build & burn-in
Nodes and fabric tested together, not just one at a time.
04
Install & acceptance
Installed, then accepted against a test plan you sign off on.
FAQ
Questions we hear first.
How long does a cluster take?
Longer than a single server. GPU and networking lead times plus facility work set the pace, so we give a dated plan after the site survey. If you need production capacity sooner, start with a Rack designed to join the cluster later.
InfiniBand or Ethernet?
Both work. InfiniBand is the established choice for large training jobs; RoCE Ethernet suits teams that want one network technology across the data center. We recommend one based on workload, your staff's skills and supply.
Can we start with one server and grow?
Yes, and it is often the right path. We design the first Rack with the cluster's network and power in mind, so it joins later as a node instead of becoming a stranded box.
Get a quote
Spec your system.
Tell us the models you want to run, how many people will use them and where the hardware should live. An engineer replies with a first configuration and the questions that decide the quote.