My observations from KubeCon India, 2025.

Welcome Board at KubeCon India, 2025

[CNCF provided me a Media + Analyst Pass, thank you CNCF]

Namaste, KubeCon!

KubeCon + CloudNativeCon Hyderabad 2025 felt like a redux of cloud-native victories—yet something in the air hinted at change. Jonathan Bryce keynoted at the event (his first KubeCon keynote as the Executive Director at CNCF), and in my hallway conversations with him and CNCF CTO Chris Aniszczyk, both emphasized the same point: the bridge to AI-native workloads must be built by first strengthening cloud-native fundamentals.

Bryce’s keynote slide laid out three pillars shaping the future:

  1. Observability – Standardized, vendor-neutral, and driven by OpenTelemetry.
  2. Platform engineering – Essential to ship faster and safer.
  3. Cloud-native as the pathway to AI production – Leveraging community expertise to meet AI’s operational demands.

That framing made the question sharper: not if Kubernetes can run AI, but how it can remain relevant as AI-native architectures take hold.

Kubernetes: From Superstar to Supporting Role?

Kubernetes earned its status as the de facto operating system of the cloud by orchestrating containerized workloads across heterogeneous environments with unmatched reliability. It was the star of the microservices era—managing deployments, scaling, and resilience at a global scale.

But AI-native workloads (post GenAI era) are a different beast:

  • Centralized models are trained and hosted inside hyperscaler environments, rarely leaving their tightly integrated infrastructure.
  • Inference is increasingly API-driven, rather than being deployed as containerized services (at the application level).
  • Training and scheduling demand GPU-aware, high-throughput orchestration that often leans on specialized frameworks.

The AI-native stack is layered differently:

  1. Agents & AI applications
  2. Context pipelines & data layers
  3. Models & inference APIs
  4. Accelerated compute

In this architecture, Kubernetes risks slipping into “invisible plumbing”—important, but no longer where the visible innovation happens.

Where Kubernetes Still Matters

It would be premature to write Kubernetes out of the AI-native script. There are key scenarios where it remains essential:

  • Hybrid AI deployments – For organizations keeping sensitive data or models on-prem, Kubernetes still serves as the universal control plane.
  • Unified operations – AI workloads need integration with CI/CD, observability, and security pipelines—domains where Kubernetes is already entrenched.
  • Mixed workloads – AI services that must run alongside traditional microservices benefit from Kubernetes’ ability to manage both seamlessly.

Challenges Kubernetes Must Address

If Kubernetes wants to remain relevant, it must evolve to address AI’s specific demands:

  1. Advanced GPU scheduling – Matching or integrating with the capabilities of frameworks like Run:ai.
  2. Deep AI framework integration – Making orchestration for Ray, PyTorch, and other distributed AI stacks seamless.
  3. Optimized data pipeline handling – Enabling low-latency, high-throughput access to massive datasets.
  4. Inference scaling – Treating model APIs as first-class deployment targets.

Failing here risks reducing Kubernetes to background infrastructure—useful but strategically sidelined.

Conclusion: From Orchestrator to Hybrid Compute Fabric

Kubernetes isn’t going away. It powers AI workloads across hyperscalers and enterprises, but it must adapt to stay central in the AI-native era.

With extensions like Ray (via KubeRay) and Run:ai, Kubernetes can be the unifying operational layer for both microservices and AI-native systems. The opportunity is not to fight for the top of the AI stack, but to own the fabric underneath it—ensuring AI runs reliably, securely, and at scale wherever it lives.

The real question isn’t whether Kubernetes can run AI—it clearly can. The question is whether Kubernetes will evolve fast enough to remain part of the AI-native conversation.