In this briefing
- 01OpenShell places Agent operating boundaries under open governance
- 02KubeRay brings training, jobs and online inference into one lifecycle
- 03NVIDIA places cluster-management and recovery components in preview and open source
- 04Shipping server processors provide a computing foundation for an open control plane
- →What to watch next
- ↗Sources and verification
- The Cloud Native Computing Foundation has accepted OpenShell and KubeRay as Sandbox projects. The two projects respectively cover Agent operating boundaries and Ray workload orchestration, but Sandbox status only marks the starting point of open governance; Sandbox status is not proof of security certification or production maturity.
- NVIDIA Kubernetes Engine and NVIDIA Kubernetes Recipes have entered public preview, while NVIDIA Snapshot start-up figures come from tests by NVIDIA and its design partners. Enterprises need to verify open-source code, preview status and specific test results separately.
- 6th Gen AMD EPYC SP7 server processors are generally available, and related server platforms have begun shipping. The core, bandwidth and Agent-capacity figures published by AMD still depend on the specific processor, memory, software, orchestration and complete-system configuration.
OpenShell places Agent operating boundaries under open governance
On 10 November, the Cloud Native Computing Foundation accepted OpenShell as a Sandbox project. NVIDIA contributed the open-source runtime for securing and governing autonomous Agent systems to the foundation, and OpenShell will move from NVIDIA's GitHub organisation to the neutral ai-openshell organisation. The project has adopted a public request-for-comments process and an explicit commitment to vendor neutrality, and the project uses Cedar as its policy language.
OpenShell places policy enforcement in the infrastructure surrounding an Agent system. The project provides extension points for computing, inference providers, credential storage, isolation, approval workflows and observability, but the foundation material still lists conformance testing, real approval workflows, cross-platform isolation, security reviews, software bills of materials, signing, continuous integration, multi-tenancy and identity integration as work that needs to continue.
For enterprise adoption, open governance can allow platform teams to inspect policy-enforcement code, participate in interface development and reduce the risk that one supplier controls a core operating boundary. Enterprises still need to validate denial rules for files, networks, credentials and model endpoints in their own environments, and enterprises still need to validate audit records, human approval, fail-closed behaviour and upgrade rollback. Sandbox status cannot replace a security assessment, compatibility testing or a formal support contract.
KubeRay brings training, jobs and online inference into one lifecycle
On 10 November, the Cloud Native Computing Foundation announced that KubeRay had joined the foundation. KubeRay uses RayCluster to manage Ray cluster nodes, health checks, control-service fault tolerance and dynamic scaling. KubeRay uses RayJob to create temporary clusters for batch processing, distributed training and reinforcement learning, and RayJob releases those clusters after a job finishes. KubeRay uses RayService to manage long-running online inference, rolling upgrades, automatic failover and traffic routing.
The official KubeRay material also lists integrations with OpenTelemetry, Prometheus, Fluent Bit, Kueue, Volcano and KAI. The integrations cover metrics, logs, batch scheduling, multi-tenant quotas, priority pre-emption and topology scheduling, and KubeRay can combine automatic mutual Transport Layer Security with Kubernetes network policies to protect Ray clusters. KubeRay is entering the foundation's Sandbox stage, not a higher maturity stage.
For enterprise adoption, one public set of lifecycle objects can reduce duplicate work that would otherwise create separate orchestration interfaces for training, batch processing and online inference, and common objects can help platform teams connect resource quotas, logs and release records. Foundation governance does not automatically assume responsibility for cluster upgrades, model correctness, data residency, failure recovery or service levels. Enterprises need to use their own workloads to validate the version combination across Ray, Kubernetes, the scheduler, storage and inference services.
NVIDIA places cluster-management and recovery components in preview and open source
On 10 November, NVIDIA announced NVIDIA Kubernetes Engine, NVIDIA Kubernetes Recipes and NVIDIA Snapshot. NVIDIA Kubernetes Engine and NVIDIA Kubernetes Recipes are currently in public preview, and both projects use open-source artefacts to help teams manage cluster lifecycles and use signed, version-pinned images. NVIDIA Snapshot reduces model reloading and warm-up work by restoring a previously saved copy.
NVIDIA and its design partners published two sets of Snapshot tests. Crusoe tested GLM-5.3 on an NVIDIA HGX B300 system, and Crusoe reduced workload-ready time from about 8 minutes to 31 seconds. Saturn Cloud reduced vLLM start-up time during scaling from zero from 2 to 8.5 minutes to 6 to 34 seconds. The figures come from NVIDIA and its design partners, and the figures cover only the hardware, models and start-up conditions listed in the public material.
For enterprise adoption, version-pinned images and inspectable recovery components can place cluster changes and model-copy state in the same release record. Public preview components can still change their interfaces, support scope and upgrade methods, and design-partner tests cannot replace the target enterprise's acceptance tests for model bytes, cache state, storage failures, network congestion, recovery consistency and complete cost. Enterprises need to manage preview features separately from formal production dependencies.
Shipping server processors provide a computing foundation for an open control plane
On 10 November, AMD announced that 6th Gen AMD EPYC SP7 server processors were generally available and that server manufacturers had begun shipping platforms that use the processors. AMD says that related cloud instances from Microsoft Azure and Oracle Cloud Infrastructure will become available later. The AMD product page positions the processors for Agent sandboxes, graphics-processing-unit host nodes and general services, and the AMD product page lists up to 256 cores, 512 threads, up to 5.3 GHz, up to 1.6 TB/s of memory bandwidth per socket and PCIe Gen 6.
AMD's performance and Agent-capacity figures come from supplier tests or estimates. The AMD product page states explicitly that thread count is only a proxy for estimated Agent capacity, and that actual capacity depends on the workload, model, memory, software, orchestration and system configuration. The public material does not provide model, quantisation, context, batch, time to first token, complete-system power, complete-system price or common after-sales conditions that apply to every server.
For enterprise adoption, shipped processors and servers represent a purchasable state, but that state cannot prove that a particular Agent workload has reached the required concurrency, latency or cost target. Enterprises need to check processor power, memory capacity, networking, storage, graphics processing units, virtualisation, supply region, price and support contract for each server model, and enterprises need to replay capacity tests with real Agent systems, tool calls, retrieval and business applications. Planned cloud instances cannot be counted in advance as currently available capacity.
What to watch next
- Can OpenShell and KubeRay progress from the Sandbox stage to public conformance testing, security reviews, long-term support and multi-supplier production cases?
- Can NVIDIA publish stable interfaces, upgrade commitments, a complete compatibility matrix and independent recovery tests for the three new components?
- Can AMD and its server partners publish power, memory, concurrency, time to first token, complete price and supply regions for Agent workloads on specific complete systems?
Sources and verification
Golden Data has edited this briefing from the public materials listed above. The original sources govern facts and figures. The enterprise relevance sections are Golden Data editorial analysis and do not constitute an endorsement of any third-party product.
← Back to AI Daily Briefing