NVIDIA’s $4 Million CNCF Pledge Tests the GPU Layer Around Kubernetes

A common interface does not settle GPU control

NVIDIA frames the effort as an upstream contribution under community governance. It says it joined the CNCF Governing Board this year, and that its KAI Scheduler entered the CNCF Sandbox. KAI is an open-source Kubernetes scheduler for GPU clusters; NVIDIA’s KubeCon program describes it as supporting large-scale training and inference, alongside features including gang scheduling and hierarchical queues. NVIDIA says it has run the scheduler on clusters with more than 10,000 GPUs.

Those are claims about NVIDIA’s software and its own experience, not a finding that one scheduler or driver design is best for every operator. The alternative approaches displayed at the same event make that plain. SUSE Virtualization’s Harvester uses native NVIDIA MIG vGPU support to create hardware-level partitions. VMware Cloud Foundation, by Broadcom, presents MIG slices through its hypervisor to virtual machines and can subdivide them for tenants, according to an analysis of the show. The analysis says that approach offers a different degree of granularity from bare-metal Kubernetes device plugins, especially for regulated workloads.

So the value of conformance and CI is compatibility, not a verdict on deployment architecture. An operator may still prefer hardware partitions, hypervisor-mediated isolation or a different scheduling policy depending on its tenancy, security and workload requirements. NVIDIA’s funding can reduce the cost of discovering failures among those layers; it cannot remove the trade-offs.

NVIDIA company-reported that the Kubernetes AI Conformance Program grew from 18 to 31 certified platforms in its first few months. Source: CNCF.

Production usage is the unresolved measure

NVIDIA cites CNCF’s 2025 survey to argue that the software gap is operational: 82% of container users run Kubernetes in production, and 66% of organizations hosting generative AI use it for some or all inference workloads. In the same company account, only 7% of organizations deploy models daily and 47% do so intermittently. The survey figures describe the respondents cited by NVIDIA; they do not establish that testing capacity is the cause of the deployment gap.

The company also says the Kubernetes AI Conformance Program grew from 18 to 31 certified platforms within a few months, with version 1.35 requirements covering agentic-workflow support and in-place pod resizing for inference serving. That is a count of certifications, not a measure of how many workloads run in production or how well they perform.

What the next evidence must show

The pledge becomes consequential if it yields public, reproducible evidence that different drivers and platforms work reliably on real GPUs. The practical test is whether CNCF and participants disclose the hardware covered, the software versions and scenarios tested, failures found and fixed, and the mix of providers taking part.

Without that evidence, the commitment is a strategically sensible investment by a GPU supplier in an ecosystem that needs its hardware, but not a demonstrated answer to the operational problems it identifies. With it, operators would have a more useful basis for judging whether a certified stack fits their own isolation and scheduling choices.