Huawei Cloud used KubeCon EU 2026 to present a systematic “AI-native” infrastructure stack focused on AI scheduling, multi-cluster orchestration, sidecarless service mesh, and edge intelligence. It addresses three core pain points in the Agentic AI era: low resource utilization, complex cross-cluster governance, and high inference latency. Keywords: Agentic AI, cloud native, Kubernetes.
Technical specification snapshot
| Parameter | Details |
|---|---|
| Focus areas | Agentic AI infrastructure, cloud native, open source collaboration |
| Key projects | Volcano, Karmada, Kmesh, KubeEdge |
| Primary languages | Go, Rust, and eBPF-related kernel technologies |
| Core protocols/platforms | Kubernetes, CNCF, multi-cluster scheduling, service mesh |
| GitHub stars | Not provided in the source |
| Core dependencies | Kubernetes, vLLM, Argo CD, Argo Rollouts, eBPF |
| Typical scenarios | LLM training and inference, Agent orchestration, hybrid cloud delivery, edge AI |
This appearance showcased a complete infrastructure stack for Agentic AI
KubeCon + CloudNativeCon Europe 2026 sent a clear signal: cloud native is no longer just a container orchestration layer. It is evolving into the runtime foundation for AI. Under the theme “Powering the Agentic Future,” Huawei Cloud showcased a systematic blueprint for AI-native infrastructure.
AI Visual Insight: This image shows the conference’s main visual identity and on-site atmosphere, highlighting KubeCon Europe’s role as a global cloud-native bellwether and emphasizing the accelerating convergence of AI and cloud-native infrastructure.
This stack is not a loose collection of point products. It forms a closed loop around training, inference, Agents, service communication, and edge collaboration. Its core value lies in moving AI workloads from isolated execution toward unified scheduling and global governance.
A minimal view of AI infrastructure collaboration
components:
volcano: ai-scheduler # Handles training, inference, and Agent scheduling
karmada: multi-cluster # Handles cross-cluster orchestration and unified control
kmesh: service-mesh # Handles low-latency service governance
kubeedge: edge-ai # Handles edge-cloud collaboration and edge runtimeThis configuration summarizes how the four projects divide responsibilities in an AI-native architecture.
Volcano is evolving from a job scheduler into an AI lifecycle foundation
Volcano is no longer focused only on batch scheduling. It is expanding into large-model training, inference, and Agent workloads. The core problem it targets is resource fragmentation and throughput loss caused by running AI jobs across multiple isolated systems.
The source article notes that Volcano highlighted Kthena and AgentCube in 2025. Kthena targets LLM inference services and supports frameworks such as vLLM. AgentCube targets AI Agent workloads and emphasizes high-performance orchestration. This means Volcano is extending beyond the resource scheduling layer into the AI application runtime layer.
AI Visual Insight: This image corresponds to the topic of splitting and governing oversized LLM jobs across clusters. Its central message is that a single large workload can be decomposed and executed across multiple clusters to break through single-cluster capacity and scheduling limits.
AI Visual Insight: This image shows a technical session focused on Volcano, Kthena, and AgentCube, illustrating how Volcano has evolved from a single project capability into a coordinated AI scheduling system with multiple subprojects.
Volcano’s capability path can be abstracted into three layers
Training scheduling -> Inference serving -> Agent orchestration
Single cluster -> Multi-cluster -> Heterogeneous compute pooling
Resource allocation -> Lifecycle management -> Unified scheduling control planeThese three lines of evolution show that Volcano’s value has risen to the abstraction layer of AI infrastructure itself.
Karmada is becoming the unified control plane for multi-cluster AI orchestration
Karmada’s upgrade focus is shifting from a “multi-cluster management tool” to a “multi-cluster AI orchestration foundation.” This change is critical because production deployments for Agentic AI naturally span regions, clouds, and compute pools.
The source highlights several enhancements: application-priority scheduling, federated resource quotas, failover for stateful applications, and collaboration with Volcano Global. Together, they serve one goal: moving optimization for massive LLM and Agent workloads from the single-cluster level to the global level.
AI Visual Insight: This image shows a Karmada community meeting and presentation session, reflecting a mature open source collaboration model built around multi-cluster scalability, workload distribution, and roadmap alignment.
At the delivery layer, Karmada’s integration with Argo CD and Argo Rollouts is even more practical. It extends mature progressive delivery capabilities from a single cluster to multiple clusters, making “define once, deliver to any cluster” a realistic operating model.
AI Visual Insight: This image focuses on the unified progressive delivery model built with Karmada and Argo. Technically, it represents coordinated canary releases, policy propagation, and consistency control across multiple clusters through a global orchestration layer.
A simplified multi-cluster delivery example
apiVersion: policy.karmada.io/v1alpha1
kind: PropagationPolicy
metadata:
name: ai-inference-policy
spec:
resourceSelectors:
- apiVersion: apps/v1
kind: Deployment
name: llm-gateway
placement:
clusterAffinity:
clusterNames:
- cluster-eu-1 # Deploy the inference gateway to the Europe cluster
- cluster-ap-1 # Deploy it to the Asia-Pacific cluster as wellThis example shows how Karmada propagates the same workload to multiple target clusters.
Kmesh rebuilds the service mesh data path with Rust and eBPF
In AI inference scenarios, service mesh latency and resource overhead are amplified even further. Kmesh follows a sidecarless design: it pushes governance capabilities down toward the kernel layer to minimize the forwarding overhead introduced by traditional sidecar proxies.
The source explains that Kmesh introduced Orion, rebuilt in Rust for L7 traffic processing, and paired it with eBPF-based L4 processing to create a unified high-performance data plane. This design targets both throughput and long-term memory safety and runtime stability.
Kmesh matters not only because it is “faster,” but because it is better suited to high-concurrency, low-latency, continuously running AI inference paths. For large-scale service-based model invocation, this architecture directly affects both cost and tail latency.
Kmesh’s core design logic can be summarized as follows
fn traffic_path() {
// L4 traffic goes through eBPF first to accelerate the forwarding path
let l4_fast_path = "ebpf";
// L7 traffic is handled by the Rust proxy to improve memory safety
let l7_proxy = "orion";
println!("{} + {}", l4_fast_path, l7_proxy);
}This pseudocode expresses Kmesh’s core idea: coordinated L4/L7 governance through an eBPF + Rust architecture.
KubeEdge continues to push cloud-native capabilities into edge AI scenarios
As a CNCF graduated project, KubeEdge’s recent growth has centered on edge AI and industry workload management. It solves more than simple node onboarding. It addresses the broader system problem of edge-cloud collaboration, model execution, and distributed device management.
As demand rises for inference at the edge in smart manufacturing, transportation, energy, and endpoint devices, the edge is no longer just an extension of the cloud. It is becoming the frontline execution environment for real-time decision-making. KubeEdge’s value lies in extending Kubernetes control capabilities into unstable networks and heterogeneous hardware environments.
AI Visual Insight: This image shows an edge application case featured in a KubeEdge keynote, emphasizing that KubeEdge has moved beyond project-level validation toward scaled deployment for real physical devices and industry scenarios.
Huawei Cloud’s booth presented a full-stack closed loop from infrastructure to developer platform
Beyond the open source projects themselves, Huawei Cloud also showcased its CCE intelligent computing cluster, full-stack Agent platform, and container product capabilities. This indicates that its strategy is not limited to participating in open source communities. It is trying to connect community innovation, commercial platforms, and real-world industry practice.
AI Visual Insight: This image captures interaction at Huawei’s exhibition booth and conveys a complete technical showcase path for developers, spanning AI compute clusters, Agent platforms, and open source infrastructure projects.
Overall, the real focus of this appearance was not an upgrade to any single project, but a broader trend: cloud native is becoming the default operating system for Agentic AI. The vendors and communities that can integrate scheduling, orchestration, networking, and edge capabilities into a unified platform will be closest to the core of next-generation AI infrastructure.
FAQ
1. What is the core value-add of Volcano compared with the default Kubernetes scheduler?
Volcano adds AI-oriented capabilities such as queues, batch scheduling, heterogeneous resource management, and cross-cluster task decomposition. That makes it suitable for mixed training, inference, and Agent scenarios rather than only general-purpose Pod scheduling.
2. Why is Karmada especially important for Agentic AI?
Because Agentic AI often runs across regions, clouds, and compute pools. Karmada provides unified orchestration and policy propagation, turning multiple clusters from isolated islands into a global control plane.
3. What scenarios fit Kmesh’s sidecarless architecture?
It is especially well suited to high-throughput, low-latency, long-connection-intensive microservices and AI inference workloads. By combining eBPF with a Rust proxy, it reduces sidecar overhead and improves long-running stability.
AI Readability Summary
This article reconstructs Huawei Cloud’s KubeCon EU 2026 technical showcase and focuses on four major projects: Volcano, Karmada, Kmesh, and KubeEdge. It explains how they support Agentic AI across training, inference, multi-cluster orchestration, service governance, and edge-cloud collaboration.