GPU cluster software that never needs the internet
BareStax installs, operates, updates, and recovers your on-premises NVIDIA GPU cluster, from racked server to running job. It works the same way connected or fully air-gapped, and your cluster never depends on a connection to us.
- Cluster summary
- Cluster
- Node profiles
- Nodes
- Onboarding2
- Fabric
- Storage
- Driver stacks
- Tenants
- Queues
- Jobs
- Usage
- Operations
- Approvals1
- Releases & images
- Policies
- Events
- Alarms
- Audit log
- Runbooks
- E-NODE-008gpu-02Node unreachable10 mindrained, takes no new work
- E-PROV-007gpu-03Configuration drift that requires downtime6 minmaintenance window
- train-h20092%6 pending
- train-b30071%2 pending
- inference-l40s48%0 pending
- E-NODE-008gpu-02Node unreachable
- E-PROV-007gpu-03Configuration drift that requires downtime
RB-NODE-008
Why institutions choose BareStax
-
Every server enters production the same way
One command and one approval. BareStax inventories the hardware, installs a tested driver stack, runs a two-node fabric check, and hands the node to the job queue. No hand-built driver stacks. How a node joins
-
Operations stay predictable
Drift is fixed in place or held for a maintenance window. Faulty nodes stop taking work while running jobs finish. Critical actions need a second person who reads the impact first. Health and recovery
-
Control stays inside your institution
Full operation without internet access. No outbound connections, no telemetry, an offline license file, and cluster keys and audit records that stay in your data center. Security
You rack and install. BareStax does the rest.
-
Rack the servers
Connect GPU nodes, the head node, storage, and the cluster network.
-
Install Ubuntu
Keep your own Ubuntu 24.04 installation. BareStax never erases the disk.
-
Run one command, give one approval
The node joins, validates, and starts taking jobs.
No Kubernetes. No hypervisor. Just the GPU.
No layers. Just compute.
BareStax brings declarative, self-healing control to the layers Kubernetes never manages: kernel, NVIDIA drivers, CUDA, NVLink, InfiniBand, BMC, and the job scheduler.
Every job gets the whole GPU. Nothing sits in between.
One product for the whole cluster lifecycle
-
Day 0
Install the control plane, connect identity, and test the alert path.
barestax-server init -
Day 1
Servers join with one command and one approval.
barestaxd joinbarestaxctl node approve -
Day 2
Drift is corrected; faulty nodes stop taking new work.
barestaxctl impact node gpu-03 -
Growth
New servers follow the same path; each GPU class gets its own queue.
barestaxd join -
Updates
Signed releases start on one canary node, then roll out within your disruption budget.
barestax-server upgrade --offline -
Removal
Drain a node without losing work, then remove it.
barestaxctl node drainbarestaxctl node remove
Everything a GPU cluster needs to run
-
Declarative stack management
Kernel, drivers, CUDA, and fabric software ship as tested sets.
Details -
BareStax Core and job execution
Queues, quotas, and accounting are configured for you.
Details -
Health and recovery
Faults stop new work and show the impact before anything restarts.
Details -
Air-gapped releases
Signed bundles cross the boundary by approved media or data diode.
Details -
Security and audit
mTLS, six roles, four-eyes approval, and a hash-chained audit record.
Details -
Console and portal
One console for operators, one portal for researchers.
Details
See what a fault affects before you act
When a node fails, BareStax shows the jobs, tenants, and services it touches. Verified and possible impact are kept apart, and every view shows how old its data is.
| JOBID | TENANT | USER | STATE | GPU | IMPACT |
|---|---|---|---|---|---|
| 4521 | research-a | researcher-1 | RUNNING | 4 | Verified |
| 4533 | research-a | researcher-2 | RUNNING | 4 | Verified |
- E-PROV-007 Configuration drift that requires downtime Show impact
Air-gapped by design, not by exception
Nothing in BareStax needs the internet. Updates arrive as signed bundles, are verified and approved inside your network, and roll out one canary at a time. How releases work
Your team fixes faults without calling us
Every error carries a code, its impact, what BareStax already did, and the next step, with an offline runbook. Standard support is still part of every subscription. The product interface is in English.
Check every claim before you sign
-
The full specification, in writing
Every capability on this site is specified in the technical overview: architecture, lifecycle, security, deployment, and licensing.
Download the PDF -
A pilot with written acceptance drills
Run BareStax on your own hardware for 90 days. Acceptance is proven by drills your team witnesses, from node join to a control plane restore.
Acceptance checklist -
A license that cannot strand you
If you do not renew, the cluster keeps running on the last imported release and your licensed capacity. The license is a signed offline file.
Licensing -
Security your team can test
Signed release bundles, an exportable hash-chained audit record, and a published contact for vulnerability reports.
Security
See it on hardware like yours
Book a guided remote demo. Then run a paid pilot on your own servers, with written acceptance tests.