The portability guarantee that AI workloads never quite had
Kubernetes built its reputation as the de facto container orchestration standard on a specific promise: workloads behave roughly the same regardless of which cloud they run on. AI/ML workloads have been the exception. GPU scheduling implementations vary by vendor, inference-serving configurations differ across platforms, and it's common for a setup that works on one cloud to simply not run on another.
CNCF's Certified Kubernetes AI Conformance Program is the industry's answer to that fragmentation. In 2026, the list of certified platforms nearly doubled, adding OVHcloud, SpectroCloud, JD Cloud, and China Unicom Cloud, among others. This post covers what the program actually standardizes, and how webhani factors it into AI infrastructure design for clients.
What "conformance" actually means here
At its core, the program defines a shared technical baseline for running AI workloads on Kubernetes. The 2026 update leans heavily on alignment with Kubernetes v1.35.
Key v1.35 alignment points
├─ Stable In-Place Pod Resizing
│ adjust inference model resources without a restart
├─ Workload-Aware Scheduling
│ avoids resource deadlocks during distributed training
└─ Expanded agentic workload validation
applies the same trusted sandbox model used for
containers to agent execution, constraining agents
to their assigned limitsThe addition of agentic workloads to the validation scope is worth calling out specifically. As autonomous AI agents become a normal part of production systems, this is the industry trying to answer a concrete question — how do you constrain what an agent can do, at the Kubernetes level — with a shared standard instead of every vendor inventing their own answer.
Why In-Place Pod Resizing matters in practice
Inference workloads see resource demand swing sharply with traffic. Historically, changing a pod's resource allocation meant recreating the pod, which introduced downtime or latency spikes right when serving capacity mattered most.
# Conceptual example: a pod configured for in-place resizing
apiVersion: v1
kind: Pod
metadata:
name: inference-server
spec:
containers:
- name: model-server
resources:
requests:
cpu: "2"
memory: "8Gi"
limits:
cpu: "4"
memory: "16Gi"
resizePolicy:
- resourceName: cpu
restartPolicy: NotRequired
- resourceName: memory
restartPolicy: NotRequiredrestartPolicy: NotRequired lets you scale a pod's resources up during a traffic spike without recreating it. For availability-sensitive workloads like inference serving, that's a meaningful practical difference, not a marginal one.
What webhani tells clients evaluating AI infrastructure
1. Select on conformance, not just "it runs"
Confirming that an AI workload runs on a given cloud or distribution isn't the same as confirming it's portable. Adding CNCF certification as a selection criterion reduces future multi-cloud migration risk and vendor lock-in exposure.
2. Design for sandboxed agent execution from the start
If autonomous agents are on your production roadmap, verify during initial platform selection whether the infrastructure can run agents under the same isolation model used for containers. Retrofitting isolation later costs considerably more than building it in from the beginning.
3. Don't assume GPU scheduling maturity is a solved problem
The fact that the program is adding tougher requirements for v1.35 is itself a signal — conformance validation in this space is still maturing. Workload-Aware Scheduling only recently became a certification requirement, so it's worth actually verifying whether an existing setup meets this bar rather than assuming it does.
Takeaway
CNCF's Kubernetes AI Conformance Program brings concrete standards to a part of AI infrastructure — portability — that's long been vague. Developments like in-place pod resizing and sandboxed agentic workload execution give organizations running AI across multiple clouds a clearer basis for platform selection. webhani recommends factoring certification status into AI/ML infrastructure design decisions rather than treating it as a footnote.
Sources: CNCF Nearly Doubles Certified Kubernetes AI Platforms (CNCF), CNCF Launches Certified Kubernetes AI Conformance Program (HPCwire)