<h2>What happened</h2>
<p>At KubeCon Europe 2026 in Amsterdam, NVIDIA announced it’s donating its Dynamic Resource Allocation (DRA) Driver for GPUs to the Cloud Native Computing Foundation, transferring ownership from NVIDIA to the broader Kubernetes community. AWS, Google Cloud, Red Hat, and several other industry partners are backing the move.</p>
<p>The announcement came alongside a few other notable items: GPU support for Kata Containers (for confidential computing), the KAI Scheduler joining CNCF as a Sandbox project, and a handful of new open source projects including Grove, NVSentinel, AI Cluster Runtime, NemoClaw, and OpenShell.</p>
<h2>Why the DRA driver matters</h2>
<p>The legacy Kubernetes device plugin model treats GPUs as simple integer counts. You ask for one GPU, you get one GPU. That works fine until you’re trying to do anything interesting - like requesting a specific MIG slice, targeting GPUs connected via NVLink, or sharing a GPU across multiple workloads without stepping on each other.</p>
<p>DRA replaces that model with a ResourceClaims-based framework driven by CEL expressions. Instead of asking for “1 GPU,” a pod can request a GPU by attributes: model, memory size, MIG profile, compute capability, NVLink topology, or multi-node memory domain. It’s a meaningful shift from counting resources to describing them.</p>
<p>Three capabilities are worth calling out specifically:</p>
<ul>
<li><strong>Attribute-based allocation</strong> - workloads can target specific GPU types or MIG slice configurations rather than taking whatever’s available</li>
<li><strong>Prioritized fallback</strong> - you can specify device preferences in order, and the scheduler falls back automatically if the first choice isn’t available</li>
<li><strong>Namespace-controlled admin access</strong> - monitoring tools can access in-use devices without disrupting running workloads</li>
</ul>
<p>DRA reached General Availability in Red Hat OpenShift 4.21, built on Kubernetes 1.34, with full YAML manifests and live NVIDIA A100 cluster output demonstrating each feature.</p>
<h2>The broader GPU stack</h2>
<p>DRA doesn’t exist in isolation. The GPU Operator handles driver and runtime lifecycle management automatically. The KAI Scheduler - now a CNCF Sandbox project - adds fractional GPU allocation, gang scheduling, and hierarchical queuing on top of that. Together these pieces form a more complete picture of how GPU infrastructure is being managed in Kubernetes at scale.</p>
<p>The multi-node NVLink support is probably the most significant capability for anyone training large models. It lets Kubernetes-managed workloads span NVLink-connected GPU clusters across nodes, which previously required stepping outside the standard Kubernetes resource model entirely.</p>
<h2>What the CNCF donation actually changes</h2>
<p>Donating the driver to CNCF means governance moves from NVIDIA to the community. In practice, that means other vendors and cloud providers can contribute to the driver’s direction without going through NVIDIA’s internal processes. Whether that translates to faster iteration or just more committee meetings depends on how the project is run - but the structural change is real.</p>
<p>For teams running Canonical Kubernetes, OpenShift, or managed Kubernetes on AWS and GCP, the practical implication is that GPU scheduling capabilities will evolve through the same open governance process as the rest of the Kubernetes ecosystem rather than on NVIDIA’s product roadmap alone.</p>


Collections

NVIDIA announced at KubeCon Europe 2025 that it is donating its Dynamic Resource Allocation (DRA) Driver for GPUs to the CNCF, transferring governance to the broader Kubernetes community with backing from AWS, Google Cloud, and Red Hat. The DRA driver replaces the legacy device plugin model — which treated GPUs as simple integer counts — with a ResourceClaims-based framework using CEL expressions. This enables attribute-based GPU allocation (by model, MIG profile, NVLink topology, etc.), prioritized fallback scheduling, and namespace-controlled admin access. DRA reached GA in Red Hat OpenShift 4.21 on Kubernetes 1.34. The donation means GPU scheduling capabilities will evolve through open community governance rather than NVIDIA's internal roadmap. Additional announcements include GPU support for Kata Containers, the KAI Scheduler joining CNCF Sandbox with fractional GPU and gang scheduling support, and several new open source projects.

NVIDIA donates GPU Dynamic Resource Allocation driver to CNCF at KubeCon Europe