From the NVIDIA edition of October 13, 2026
NVIDIA makes the full Run:ai GPU orchestration platform open source as part of DSX AI Factory
NVIDIA · Technology vendor · Global
- Publisher
- NVIDIA Technical Blog
- Original publication
- 2026-10-13
- Source retrieved
- 2026-10-14
- Event date
- 2026-10-13
What happened
On Oct 13 NVIDIA released the complete Run:ai platform as open source, plus four separately adoptable components: KAI Scheduler (Apache 2.0, CNCF Sandbox), GPU Fractioning Operator (CUDA 13.4-native fractions; the proprietary implementation retired), KAI Resource Management (projects, departments, quotas, node pools) and Workload Map (a Go library). Eight DSX partners are integrating it: Crusoe, Mirantis, Nebius, OpenNebula Systems, Rafay, Saturn Cloud, Spectro Cloud and vCluster. The NGC Community Product requires accepting NVIDIA license terms and excludes enterprise support; paid editions continue through Mission Control and AI Enterprise.
Why it matters
High for research universities and state shared AI computing programs that pool GPUs across labs or agencies on Kubernetes.
Evidence and measured results
Vendor announcement; Model Streamer benchmark figures cited are vendor and partner measurements; no utilization results for the open release itself.
Limitations and uncertainty
Vendor claims; utilization and throughput benefits asserted without data for this release; community edition unsupported.
Put this evidence to work
Lighthouse Advisory interpretation, grounded in this source. Enriched 2026-10-14; this does not change the original publication date. Each role follows the same checklist so takeaways compare across sources.
Sales
Role takeaway
- Customer problem
- Shared campus and state GPU clusters suffer from idle reserved GPUs and queue conflicts between labs, and commercial orchestration licenses have been a barrier to fixing it.
- Stakeholders
- University research computing / HPC director
- Campus CIO and IT finance
- Principal investigators sharing GPU clusters
- State or system-wide shared AI computing program leads
- Discovery questions
- Is the GPU estate already on Kubernetes, or still Slurm on bare metal?
- Who would own upgrades and security patching without a vendor support contract?
- Can counsel accept the NVIDIA license terms attached to the NGC Community Product?
- Value hypothesis
- Interpretation: a scoped pilot could raise GPU utilization for notebook and small-model workloads through fractions and quotas; the source reports no utilization figures for this release.
- Potential engagement
- Bounded pilot of open-source Run:ai or KAI Scheduler on one shared Kubernetes GPU cluster, with utilization measured before and after.
- Unsupported claims
- Any specific utilization or cost savings
- That the free edition is production-supported
- That it replaces Slurm for tightly coupled HPC jobs
Pre-sales engineering
Role takeaway
- Fit
- High for research universities and state shared AI computing that run Kubernetes GPU clusters; low where workloads are single-team or Slurm-only.
- Architecture and integration
- Control plane plus per-cluster services on Kubernetes; KAI Scheduler, GPU Fractioning Operator (CUDA 13.4-native, scheduler-agnostic), KAI Resource Management for projects, departments and node pools; Workload Map as a Go library.
- Prerequisites
- Kubernetes GPU cluster with NVIDIA GPU Operator
- CUDA 13.4 drivers for native fractions
- Identity integration for role-based access
- Constraints
- Free edition excludes NVIDIA enterprise support
- NGC download requires NVIDIA license acceptance in addition to open-source licenses
- Commercial editions' terms vary by offering
- Security
- Shared multi-tenant clusters need tenant isolation review; GPU fractions isolate memory but share compute, so sensitive-data workloads need separate policy.
- Proof of value
- Measure utilization on a pilot cluster for four weeks before and after
- Test fraction memory isolation with representative notebooks
- Confirm a security-patch path for the community edition
Delivery
Role takeaway
- Work
- Deploy and configure open-source Run:ai or individual components on an existing Kubernetes GPU cluster, mapping labs to projects and departments.
- Dependencies
- Kubernetes platform team
- CUDA 13.4 driver rollout
- Campus identity provider
- Ownership
- Research computing platform team, with the CIO's office accountable for the support model.
- Skills and adoption
- Platform engineers need Kubernetes scheduler and GPU operator skills; researchers need training on requesting fractions and quotas.
- Governance checkpoints
- Support model decision before production use
- Legal review of NGC Community Product terms
- Fair-share policy approved by research governance
- Acceptance criteria
- Proposed: measured GPU utilization improvement on the pilot cluster versus baseline
- Proposed: documented patch and upgrade procedure tested once
- Risks
- Unsupported software in production
- Kubernetes skills gap
- Policy disputes between labs over quotas
Implementation considerations
Lighthouse Advisory interpretation across the operating dimensions a public-sector buyer must settle before this evidence becomes a design. Each note answers the question under its heading for this specific source.
Architecture and integration
What must connect, and where does the AI sit in the workflow?
Kubernetes-native GPU scheduling, fractions and quotas; complements but does not replace Slurm-based HPC.
Governance
Who approves, reviews and stays accountable for outcomes?
Departmental quota hierarchies can encode research governance decisions about who gets guaranteed GPU capacity.
Security and privacy
What data, permissions and controls need testing?
Not addressed in depth; multi-tenant isolation depends on configuration and partner tooling such as per-tenant clusters.
Accessibility and workforce
Who is affected, and what skills or accommodations follow?
Requires Kubernetes and GPU platform skills many campus teams lack.
Procurement
What should contracts, pricing and exit terms secure?
Removes license cost for self-supported use; support must be bought from NVIDIA or partners or staffed internally.
Operating model
Which teams own the service once it runs?
Shifts from vendor-supported subscription to self-operated or partner-managed service.
Publication history
- 2026-10-13NVIDIA · Issue 387 resources
Stable resource ID: nvidia-runai-open-source-dsx-20261013