When experts say that cloud native is now AI-native, they mean that the same principles that made cloud native successful for microservices are now being used to reimagine how we build and operate AI in production.

According to leaders from AWS, Google Cloud, Microsoft, and solo.io, moving AI into enterprise production isn’t just about adding GPUs. It requires three core elements:

• Vendor-neutral, mature platforms
  Organizations need a foundational, vendor-neutral infrastructure that supports AI training and serving at scale. A key signal of maturity is alignment with the Kubernetes AI Conformance program, which defines the essential primitives for AI workloads and helps guarantee interoperability across environments.
• Security by design for AI and agents
  AI introduces new risks, especially with agentic flows that can act autonomously. Security now has to cover not just containers, but also the model supply chain and the behavior of non-deterministic models.
• Active community contribution
  The CNCF community expects organizations to move beyond just consuming open source. Contributing to CNCF Special Interest Groups (SIGs) helps shape the next wave of AI-native capabilities and ensures your needs are reflected in emerging standards.

In practice, becoming AI-native means:

• Building on open, interoperable, vendor-neutral standards rather than proprietary stacks.
• Designing platforms that support both research workflows (e.g., Python-heavy experimentation) and production-grade operations.
• Embedding security, evaluation, and governance into the AI lifecycle from day one.

AI workloads behave very differently from traditional microservices. Instead of many small, loosely coupled services, AI often looks like a large monolith that needs to initialize multidimensional matrices in memory across many nodes. Standard Kubernetes wasn’t built for this kind of tight coupling and high-performance compute.

To address this, engineers across the cloud native ecosystem are collaborating on several key initiatives that reshape Kubernetes for AI without locking users into rigid architectures:

• Pod Groups (Workload API)
  Pod Groups treat a set of pods as a single failure domain. This helps ensure the proximity and reliability needed for large-scale AI matrix initialization, where many pods must start and operate together for training or inference jobs.
• Dynamic Resource Allocation (DRA)
  DRA integrates specialized chips and GPUs directly into the Kubernetes scheduler. This allows the platform to understand hardware nuances and allocate resources efficiently, which is critical for AI training and high-intensity serving.
• Inference Gateways
  Using Gateway API standards, the community is building AI-specific gateways that handle prompt management and high-intensity generative model responses. These inference gateways help standardize how traffic is routed to models and how responses are managed at scale.

Together, these efforts aim to:

• Make Kubernetes a first-class platform for AI training and inference.
• Preserve openness and interoperability so organizations are not tied to a single vendor.
• Provide a path from experimentation to production on the same cloud native foundation.

AI is not just a new workload type; it is reshaping how teams work and how they think about security.

Changes in engineering workflows

• Prototyping before PRDs
  Instead of starting with a traditional Product Requirements Document (PRD), product managers increasingly begin with AI-generated prototypes to test ideas quickly. Documentation often follows once a concept has been validated.
• Code review bottlenecks
  AI can generate large volumes of code, which creates a review bottleneck because humans still need to validate quality, security, and maintainability. This shifts the challenge from writing code to reviewing and governing it at scale.
• Toward agentic SRE
  The panelists see a future of agentic SRE, where AI agents assist with root-cause analysis and remediation. These agents help triage incidents and propose fixes, while humans remain in control of mission-critical decisions.

Evolving security practices for AI

• Beyond container scanning
  Security now extends to the model supply chain and the risks of non-deterministic outputs. Teams must understand where models come from, how they are trained, and how they behave under different prompts.
• Consistent evaluation and guardrails
  The community is investing in consistent evaluation frameworks (Evals) and guardrails that are applied before models reach production. This helps organizations systematically test for safety, reliability, and policy compliance.
• Open standards for citation and prompt safety
  To reduce risks like remote code execution via prompt injection, the ecosystem is adopting open standards such as llms.txt and standardized schema markups. These standards help ensure that AI models crawling the web cite and recommend only authoritative, trusted open source sources.

Overall, AI is pushing organizations to:

• Rethink how they prototype, review, and ship software.
• Integrate AI-specific security and evaluation into their existing DevSecOps practices.
• Rely on open, vendor-neutral standards so that when someone asks, “How do I scale this?”, the answer is grounded in interoperable cloud native approaches.