BareStax is for institutions that operate their own GPU clusters
BareStax fits organizations that own NVIDIA GPU capacity in their data centers. Those organizations remain responsible for the operating system, fabric, scheduler, access, and recovery. BareStax gives their platform teams one control model from installed server to job queue.
Enterprise AI teams, research centers, universities, public institutions, and regulated sectors use it on connected or fully air-gapped networks.
Four roles divide cluster ownership and access
-
Cluster operators control state instead of script order
Operators declare node profiles and inspect observed status; BareStax handles inventory, stack installation, BareStax Core configuration, drift handling, health gates, staged recovery, canary updates, and node removal.
The console and
barestaxctluse the same API, policy, and audit path. -
GPU ownership
Hardware, data, identity, storage, software repositories, and license operation remain inside the institution.
Different GPU classes share one pool through separate queues; BareStax Core assigns full GPUs directly to jobs without MIG, vGPU, Kubernetes, or a virtualization layer.
-
Security and audit
mTLS, cluster PKI, role separation, built-in TOTP, four-eyes approval, and signed images apply to web and CLI operations. Tenants remain isolated from one another. Audit data can be exported or sent to a SIEM.
Air-gapped releases carry signatures, integrity data, compatibility checks, scan reports, pinned repository snapshots, and key fingerprints across the approved boundary.
-
Researcher access
Researchers submit through the command line,
barestax run, or a portal form and see job state, waiting reason, quota, logs, and cancellation controls.Interactive Jupyter, VS Code, RStudio, and remote desktop sessions run as BareStax Core jobs. The browser portal also provides files and a terminal under the researcher's own Linux identity.
Four recurring cluster shapes use the same join, queue, and release paths
-
Enterprise training clusters need repeatable node state
Tested profiles cover the kernel, driver, CUDA, fabric software, Apptainer, and system settings; drift handling determines whether correction can run in service.
-
Research environments need tenant and quota visibility
Queues, tenants, quotas, fairshare, accounting, and waiting reasons share one scheduler state without exposing another tenant's jobs, data, or queues.
-
Isolated institutions need a complete offline release path
The connected-side bundler creates a signed release bundle. Approved media or a data diode transfers it for verification, independent approval, canary deployment, and budgeted rollout inside the air-gap.
-
Mixed GPU generations need one operating pool
BareStax inventories each node and assigns its matching profile, while each GPU class keeps its own BareStax Core queue under the same control plane.
The operating approach determines what remains manual
| Approach | Primary strength | Work retained by the institution |
|---|---|---|
| Cluster manager and scripts | Mature installation tooling | Upgrade, recovery, and air-gap procedures remain manual; a script ordering error can leave a silent partial installation. |
| Kubernetes with a GPU operator | Strong service workload model | Batch training adds scheduling and RDMA layers; operating system, driver, fabric, and BMC control remain outside. |
| Virtualization with PCIe passthrough | Familiar virtual machine operations | A GPU stays assigned to a virtual machine while idle, without scheduler-based sharing across jobs. |
| BareStax | One product from installed bare-metal host to job queue, with declarative reconciliation and air-gap release control | The institution racks the hardware and installs Ubuntu. |
BareStax keeps deliberate boundaries
BareStax is not a Kubernetes distribution and does not place Kubernetes above the GPU cluster. It is not a virtualization platform; BareStax Core assigns the GPU to a job rather than fixing it to a virtual machine.
It is not a cloud service: it runs in your data center, sends no data outside, and connects to no remote server.
BareStax does not install the operating system, provide automatic scaling, or provide a load balancer. Your team installs Ubuntu, and the institution's load balancer exposes long-running service jobs.
BareStax starts above the institution's Ubuntu installation
Your team racks the hardware and installs Ubuntu; BareStax takes over at cluster control.