Request a demo

Built for GPU clusters that cannot depend on anyone else

BareStax is made for institutions whose clusters must run without an outside connection: no internet, no license server, no call home. Where a connection is allowed, it works exactly the same way.

Four things that set BareStax apart

  • Nothing needs the internet

    No part of BareStax needs internet access to install, operate, update, recover, or license the cluster. Updates arrive as signed release bundles on approved media or through a data diode.

  • Nothing depends on us

    The license is a signed file installed offline. There is no license server, no outbound connection, and no telemetry. If you do not renew, the cluster keeps running.

  • Every layer, one product

    Kernel, NVIDIA drivers, CUDA, NVLink and InfiniBand fabric, BMC, and BareStax Core are managed as one declared state, from racked server to running job.

  • A second person on every critical action

    Removing a node, promoting an image, or restarting hardware needs approval from someone other than the requester, who reads the impact first and types the target name.

One product instead of a stack of tools

BareStax applies the declarative, self-healing model to the layers that usually sit outside any one tool: operating system settings, kernel, drivers, the NVLink, NVSwitch, and InfiniBand fabric, BMC, and the job scheduler.

There is no Kubernetes or virtualization layer between the job and the GPU. Each job gets the whole GPU, and every server enters production through the same command, the same checks, and the same approval.

Built for the team on call

  • The impact comes first

    An alarm opens straight into the jobs, tenants, and services it touches. Verified and possible impact are kept apart, and every view shows how old its data is.

  • No false green

    If health data is stale, the state is shown as unknown. Old data is never presented as live.

  • Recovery with a brake

    Restarts escalate step by step under policy. When many nodes fail at once, the recovery chain stops so the cluster does not harm itself.

  • Answers without a call

    Every error carries a code, its impact, what BareStax already did, the next step, and an offline runbook.

How it compares to the usual approaches

  • Instead of a cluster manager and scripts

    Installation tools are mature, but upgrades, recovery, and air-gap procedures stay manual, and a script ordering error can leave a silent partial install. BareStax covers all of it in one product.

  • Instead of Kubernetes with a GPU operator

    That model is strong for service workloads, but batch training needs extra scheduling and RDMA layers, while the operating system, drivers, fabric, and BMC stay outside. BareStax manages those layers directly.

  • Instead of virtualization with PCIe passthrough

    VM operations are familiar, but each GPU stays pinned to a VM and sits idle when the VM is idle. With BareStax, GPUs are shared across jobs through queues and quotas.

Questions about independence

Is BareStax tied to one server vendor?

No. BareStax supports HGX systems of the Hopper and Blackwell generations and PCIe GPU servers, in one pool with a queue per GPU class. Validated hardware classes are listed in each release note.

Does any data leave our data center?

No. BareStax runs in your data center, connects to no remote server, and sends no telemetry.

Can the head node run on our virtualization platform?

Yes. The head node can be a server without GPUs or a virtual machine on your own virtualization platform.

What happens if we stop working with you?

The cluster keeps running on the last imported release and your licensed capacity. Jobs, draining, and recovery continue.

What can we review before we buy?

The technical overview, the pilot acceptance checklist, and the air-gap update checklist are open to download. Resources

See the difference on your own hardware

Start with a guided remote demo, then prove it in a paid pilot with written acceptance drills.