KubeCon China 2026: CNCF, OpenInfra, and PyTorch Share Stage for First Time, Rewriting AI Infra Boundaries

The most obvious change at this year's KubeCon + CloudNativeCon + OpenInfra Summit + PyTorch Conference China 2026 is that CNCF, OpenInfra, and PyTorch—three open-source communities that used to be at different technology layers—shared the same stage at the same industry conference for the first time. At the media briefing after the keynote, OpenInfra Foundation General Manager Thierry Carrez, CNCF Executive Director Jonathan Bryce, PyTorch Foundation Executive Director Mark Collier, and CNCF Chief Technology Officer Chris Aniszczyk repeatedly emphasized the same thing: after AI truly enters production environments, the original infrastructure capabilities are far from enough.

This looks more like catching up with how enterprises actually deploy. Jonathan Bryce said that no one runs only PyTorch, and no one runs only Kubernetes; real enterprise environments are inherently about combining multiple open-source projects to build systems. Chris Aniszczyk put it more directly: a single project cannot do it, and a single foundation cannot do it either.

What needs to be distinguished is that this is currently not a new unified organization, joint standard, or fixed cooperation mechanism. According to Jonathan, future collaboration will happen simultaneously at several levels: community interoperability, standards alignment, and joint project building—for example, cross-community collaboration such as MCP and A2A, and project connections like vLLM and llm-d that sit at the intersection of different communities. The three communities are gradually moving from "developing separately and being stitched together by users in production" toward discussing each other's interfaces and coordination earlier.

Thierry Carrez's judgment is "no infrastructure, no AI." Models, inference services, and Agents ultimately all have to run on compute, accelerators, networking, and storage, but the requirements AI brings are clearly different from traditional cloud workloads: GPUs and various accelerators are becoming increasingly diverse, and enterprises need to organize heterogeneous resources into unified resource pools; GPUs are expensive, so utilization cannot be "good enough" but must be squeezed as close to full load as possible.

Jonathan Bryce mentioned that the core direction of cloud computing over the past ten or twenty years has been to standardize environments as much as possible and then dynamically access resources through APIs, but AI is bringing heterogeneity back to infrastructure. To unlock performance, enterprises will use custom hardware and custom kernels, with tighter connections from chips all the way up to Agents, and no single open-source project can cover the entire technology chain.

This change has already penetrated inside a single AI request. A workload may need to assign different tasks to different types of hardware: KV Cache can be placed in memory or offloaded to high-speed storage; the Prefill and Decode phases also have different resource characteristics. Infrastructure is no longer just facing "scheduling a Pod onto a GPU," but needs to understand the resource requirements of different phases inside a request and then decide which type of compute, storage, and network resource the task lands on.

Jonathan Bryce mentioned that large-scale model training has in fact long been running heavily on cloud-native infrastructure such as Kubernetes; the reason inference has been pushed to the forefront separately is that it is becoming the fastest-growing AI workload and at the same time requires a large amount of specialized engineering capability that had not been fully addressed in the past.

The boundary between training and inference is blurring. Mark Collier said that hardware architectures, model architectures, and inference software are changing almost every day, and if software evolves separately for each chip, it will quickly fragment. The PyTorch Foundation hopes to see a shared software layer: the upper layer maintains a unified interface, while the lower layer adapts to different accelerators and backends. The long-term goal is that any model can be trained and served on any chip, any cloud, and connected to any Agent.

Two numbers Mark gave illustrate the optimization space in the software stack: over the past year, the number of vLLM contributors grew by more than 180%; after MiniMax's open model was released, a ComfyUI integration appeared within 48 hours, and AMD, combined with vLLM, improved performance by 11x in less than 19 days on a new hardware architecture. He summarized this type of problem as a coordination problem—chips, model architectures, and inference engines are all changing, and the key is whether these layers can complete information transfer and adaptation faster.

Agents operate differently from traditional services: traditional services receive requests and return relatively predictable results, and infrastructure cares more about availability and error rates; an Agent receives a goal and will break down tasks itself, call models and tools, and access external systems. Jonathan Bryce believes that Agents are generating workloads with higher parallelism and exponentially growing resource demands, which will directly affect routing, caching, capacity planning, and even security policies designed around human access patterns. He cited a Wikipedia case saying that Agent-driven page visits accounted for about 35%, but resource-intensive usage exceeded 60%.

Isolation is therefore pushed to a more central position. Thierry Carrez mentioned that Agent workloads require stricter isolation and constraints, and enterprises must prevent Agents from escaping predetermined boundaries to perform unauthorized operations; Ant Group uses Kata Containers' lightweight virtual machine capabilities to provide lower-level isolation for Agent workloads. Isolation capability also affects whether enterprises are willing to open more internal data to AI. Over the past year, the Kata Containers community has seen a large number of teams building Agent Sandbox or similar services, and Kata Containers 4.0 also further pushed upstream the Rust VMM architecture previously used internally at companies such as Alibaba and Ant.

The problems with Agents do not only occur at runtime. A talk at the keynote about Agent Infra mentioned that an Agent's execution process may be very short, but the permissions it obtains, the state it modifies, and the impact it has on other systems may persist for a long time, summarized as "short-lived processes with long-term consequences." Therefore infrastructure cannot only manage when a process starts and when it exits; it must also continuously track what it has done: every action should be traceable to a clear authorization subject, access should be controlled by policy, and the system should not only observe latency and errors but also track the decision process, tool calls, and execution results, and send them back into the evaluation loop. The ecosystem observations shown at the keynote said that user attention is still concentrated on upper-layer applications such as Coding Agents and personal assistants, but one of the fastest-growing areas has shifted to infrastructure layers such as Runtime, sandbox execution, observability, and evaluation.

Jonathan Bryce said that platform engineering teams that used to be responsible for general-purpose computing and application systems are now increasingly taking on AI, including both enterprises' own model training and the deployment and operation of open-weight models. In some enterprises he has encountered, AI teams are well-resourced and move quickly, but after entering production environments they quickly encounter availability, reliability, scalability, and security problems, so some enterprises have begun to have AI teams collaborate more closely with platform and operations teams.

In his view, when a workload first appears it can be very new and exciting, but for an enterprise to ultimately depend on it, the premise is that it must become "stable and boring." This is also the change that AI infrastructure is undergoing: in the past few years the industry cared more about "can this model run," and next it needs to solve a set of less sexy but directly decisive problems that determine whether enterprises dare to rely on it long term—utilization, latency, cost, scaling, security, isolation, observability, and reliability.

Models are still changing rapidly, hardware is far from converging, and Agents are at a very early stage, but enterprises will not lower their requirements for production systems just because a workload is new enough. AI competition is continuing to spread from single-model capability to the system layer: when AI is still just a demo, the model determines almost everything; when AI truly enters production environments, the model is only the beginning of the system.