KADC 2026: The openFuyao Forum
The Kunpeng Ascend Developer Conference (KADC) 2026 takes place May 22-23 at the Zhongguancun International Innovation Center in Beijing. The openFuyao sub-forum on May 23rd focuses on a single mission: releasing the potential of diversified AI computing power through openFuyao practices.
As someone who has been tracking the openFuyao project since its early days, this forum represents a significant milestone — the ecosystem has matured from a Huawei-led initiative into a multi-vendor collaboration spanning China’s largest cloud providers, academic institutions, and open-source communities.
What is openFuyao?
openFuyao is a cloud-native AI platform that provides unified scheduling, resource management, and inference serving across heterogeneous AI accelerators — including Huawei Ascend, NVIDIA GPUs, and domestic Chinese AI chips. Think of it as the Kubernetes-native layer that abstracts away hardware diversity for large-scale AI workloads.
Forum Agenda Highlights
Opening: Community Progress and Future Direction
Hu Hao (openFuyao Technical Committee Chair) kicks off with the state of the community — technical progress, ecosystem growth, and the roadmap ahead.
Aether: Elastic Scheduling for Large-Scale AI Workloads
Ji Wanqiang (JD.com Retail Architect) presents Aether — a highly available elastic scheduling framework designed for large-scale AI-native workloads. In production at JD.com’s retail infrastructure, Aether addresses the challenge of dynamically scaling training and inference jobs across thousands of accelerators while maintaining fault tolerance.
China Mobile Cloud: Ultra-Large-Scale Cluster Practice
Luo Gangyi (China Mobile Cloud Capability Center, Product Director) shares operational insights from running AI clusters at massive scale. China Mobile operates one of the world’s largest telecommunications infrastructures, and their cloud AI platform serves billions of inference requests daily.
Distributed Inference in Smart Compute Containers
Li Tao (China Telecom Cloud, R&D Director) covers distributed inference scenarios using smart compute containers. This talk bridges the gap between container orchestration and the unique requirements of inference workloads that span multiple nodes — a challenge familiar to anyone running large language models in production.
LingqueCloud: Overseas Deployment of openFuyao
Du Dongming (LingqueCloud Chief Architect) discusses taking openFuyao international — deploying the platform outside China. This is significant for the project’s global ambitions and addresses challenges around data sovereignty, latency optimization, and compliance with regional regulations.
Open-Source AI Infrastructure in the AI-Native Era
Li Haoyang (Director of CNCF Asia) provides the Cloud Native Computing Foundation perspective on open-source AI infrastructure. With CNCF projects like Kubernetes, KServe, and Kueue forming the backbone of AI platforms worldwide, this talk bridges the Chinese and global ecosystems.
AI-Native Reference Architecture and Core Capabilities
Zhu Haopeng (Huawei Fellow) presents the reference architecture for AI-native infrastructure — the theoretical foundation and core capabilities that openFuyao implements. As a Huawei Fellow, his perspective spans hardware (Ascend NPUs), software (MindSpore), and platform (openFuyao) layers.
Panel: Key Technical Challenges in AI-Native Infrastructure
The forum closes with a panel moderated by Deng Hui (Huawei AI-native technology expert) featuring:
- Zhu Haopeng — Huawei Fellow
- Hu Hao — openFuyao Technical Committee Chair
- Wo Tianyu — Beijing University of Aeronautics and Astronautics, Software School Professor
- Lei Chao — vLLM-ascend Community Maintainer
- Yang Ke — Mooncake Core Developer, Frontier Technology Expert
Key discussion topics include heterogeneous accelerator scheduling, the convergence of training and inference infrastructure, and building vendor-neutral AI platforms that scale across hardware generations.
Why This Matters
The openFuyao ecosystem represents China’s answer to the AI infrastructure challenge — and it is evolving rapidly:
- Multi-vendor convergence — JD.com, China Mobile, China Telecom, and LingqueCloud all building on the same platform
- Global ambitions — overseas deployment signals openFuyao is not just a domestic project
- CNCF alignment — CNCF Asia’s involvement ensures compatibility with the global cloud-native ecosystem
- vLLM integration — the vLLM-ascend maintainer on the panel confirms LLM serving is a first-class citizen
- Mooncake participation — Microsoft Research’s distributed KV-cache project collaborating with the Chinese ecosystem
Connection to the Global AI Infrastructure Landscape
For platform engineers and AI infrastructure architects, openFuyao addresses the same challenges as Western equivalents:
- Scheduling: Like Kueue and Volcano, but optimized for Ascend hardware
- Inference serving: Comparable to NVIDIA NIM but hardware-agnostic
- Distributed inference: Similar goals to llm-d for KV-cache-aware routing
The convergence of these approaches — East and West — suggests the industry is arriving at common architectural patterns for AI-native infrastructure regardless of the underlying hardware.
Key Takeaways
- openFuyao has grown from a Huawei project into a multi-vendor ecosystem
- China’s largest cloud providers (Mobile, Telecom) are standardizing on the platform
- International deployment is underway via LingqueCloud
- CNCF alignment ensures interoperability with global cloud-native standards
- The panel features vLLM-ascend and Mooncake contributors — bridging Chinese and international open-source AI communities
Related Content
- openFuyao: Cloud-Native AI Platform
- openEuler: The Linux Distribution Powering China’s AI Infrastructure
- Huawei Atlas 950 SuperPod
- NVIDIA NIM Multi-Node Kubernetes Deployment
Building AI platforms that scale across heterogeneous hardware? Let’s discuss your architecture.
