Four Logos, One Framework: Alibaba Cloud, Cambricon, and Ant Group Take Seats in the PyTorch Foundation

SHANGHAI — The ballroom of the Shanghai International Convention Center hosted a quiet reshuffling of power in open-source AI this week. At the first-ever KubeCon + CloudNativeCon + OpenInfra Summit + PyTorch Conference China 2026, running September 7–9, the PyTorch Foundation announced that Alibaba Cloud and Cambricon have joined as Platinum members, each earning a seat on the Governing Board and a seat on the Technical Advisory Council, while Ant Group joins as a Gold member. The three take the keynote stage alongside existing member Huawei — and their arrival puts Chinese cloud providers, chip designers, and fintech infrastructure builders directly inside the governance of the framework that trains most of the world’s open models.

It is, on its face, a membership announcement. In practice, it is the clearest signal yet of where the center of gravity in open AI infrastructure is moving.

What was actually announced

The structure of the deal matters more than the headlines. Platinum membership in the PyTorch Foundation — the vendor-neutral home for PyTorch under the Linux Foundation — comes with governance rights: one seat on the Foundation’s Governing Board and one on the Technical Advisory Council for each of Alibaba Cloud and Cambricon. Gold membership, Ant Group’s tier, carries participation without the same board weight.

The announcement lands at a conference that is itself a first: KubeCon + CloudNativeCon + OpenInfra Summit + PyTorch Conference China 2026, a four-way co-location organized by the CNCF, OpenInfra Foundation, and PyTorch Foundation, explicitly framed around “scaling AI” and the open AI stack — the chips and accelerators, open-weight models, frameworks, and infrastructure that carry AI into production.

The scale of Chinese participation is the subtext throughout the press release: more than 250 organizations across China contribute to PyTorch Foundation projects, including PyTorch itself alongside the Foundation’s growing project portfolio — vLLM, DeepSpeed, Ray, Helion, and Safetensors. That portfolio now spans model serving, distributed training, cluster orchestration, kernel authoring, and model serialization: effectively the full lifecycle of open-weight AI, from training run to inference endpoint.

The keynotes map the stack, from silicon to agents

The most revealing part of the announcement is not the membership tiers but the keynote program, which reads as a guided tour of the open AI stack — deliberately sequenced from hardware to application.

Alibaba Cloud opens at the platform level. Jionghang Cai, Senior Technical Expert on AI Infrastructure, delivers “Serving Qwen at Scale: Multi-Cluster AI Infrastructure on Karmada” — a walkthrough of how Alibaba serves the Qwen model family, one of the world’s most widely adopted open-weight series since its 2023 debut, across multiple clusters using Karmada, the CNCF-hosted multi-cluster Kubernetes orchestrator. This is hyperscale operations knowledge, the kind that used to live only inside private engineering blogs.

Huawei takes the hardware-software co-design theme. Liang Zhang, Vice President of the Ascend Computing Product Line, presents “Ascend & PyTorch: Pioneer New AI Open Ecosystem” — exploring the interoperability needed between Chinese and global AI technology stacks, a question that grows more foundational with every new accelerator that enters production.

Cambricon picks up the chip-level thread. Wei Li, Principal Software Engineer, presents “Towards Device-agnostic PyTorch: Building Unified Infrastructure for a Multi-Backend Ecosystem,” detailing how the company — China’s leading listed AI-chip designer, whose MLU (Machine Learning Unit) accelerators support large-scale training and inference — is hardening PyTorch’s device-agnostic foundation so any backend can offer broader reach and richer native capabilities.

Ant Group closes at the application layer. Yaya Xia, Open Source Analyst, and Xu Wang, Vice Chair of Ant’s Open Source Technology Committee, deliver two linked talks — “What AI Agents Need from Open Infrastructure” and “Building an Agent Runtime with Open Infrastructure” — showing how existing cloud-native building blocks like Kubernetes Agent Sandbox and Kata Containers can be assembled into a secure, on-demand runtime for AI agents.

From silicon to software to deployment: the program is an argument that open source is now built into every layer of AI in China, and the PyTorch Foundation is where those layers meet.

What the new members are saying

The quotes sketch each company’s strategic interest with unusual candor.

Dr. Feifei Li, CTO of Alibaba Cloud, framed the join as a natural extension of existing practice: “We believe the future of AI is built on open, production-proven infrastructure — and PyTorch sits at the heart of that future. Joining the PyTorch Foundation is a natural step given our years of running PyTorch at scale across heterogeneous hardware on Alibaba Cloud. We look forward to working alongside the PyTorch Foundation to raise the bar for AI infrastructure and help developers build the next generation of models with confidence.”

Zhengyu He, CTO of Ant Group, emphasized adoption economics: “AI proves its value through real tasks and real users. We want more people to be able to build with AI and afford to use it. Joining the PyTorch Foundation is another step in Ant Group’s long-term participation in global open-source collaboration.”

Elton Gong, Vice President of Software Engineering at Cambricon, made the co-optimization case: “We believe that realizing the full potential of AI computing requires both software-hardware co-optimization and a thriving open ecosystem… We want to work closely with the community to help PyTorch deliver a native, out-of-the-box developer experience across a broader range of backends.”

And Fred Li, who leads Huawei’s Computing Open Source Development Team, supplied the historical arc — and a welcome that doubles as a strategic statement: “When Huawei joined the PyTorch Foundation in 2023, we came with a goal we still hold today: to help make diverse computing power ubiquitous, and to do that work upstream first. That is what led us to help start the Accelerator Integration Working Group, and what carried Ascend NPU through to becoming the first additional platform in PyTorch… I’m glad to welcome Alibaba Cloud, Ant Group and Cambricon to the Foundation. We will keep working toward native support for Ascend — native development, native release, native distribution — so that developers get the same PyTorch experience whichever accelerator they reach for.”

Mark Collier, Executive Director of the PyTorch Foundation, cast the welcome in ecosystem terms: “Open source has become the default way the world builds AI, and the PyTorch Foundation is a leading hub for this innovation… Welcoming Alibaba Cloud, Cambricon, and Ant Group as members alongside Huawei further strengthens the global open source AI ecosystem.”

Context: this has been building all year

Today’s announcement completes a sequence. Alibaba Cloud actually joined as a Platinum member in late May 2026 — a move the company paired with pledges to contribute AI compiler optimization, multi-chip compatibility, and large-scale stability practices, built on production experience with projects like SGLang, vLLM, PAI-TurboX, and TorchEasyRec. The Shanghai announcement formalizes the China bloc: Cambricon’s arrival as the second new Platinum, Ant Group as Gold, and the four companies appearing together on one keynote stage.

The timing is not accidental. Chinese AI has shifted decisively toward the deployment phase: Qwen’s open-weight releases, DeepSeek’s training and serving stack, and hundreds of downstream deployments all run on PyTorch-family infrastructure. Meanwhile, US export controls have cut Chinese labs off from Nvidia’s best hardware, pushing Huawei Ascend and Cambricon MLU accelerators into the gap — accelerators that are only useful to the broader ecosystem if the software layer supports them natively. A chip without a framework is a very expensive paperweight.

That is the real transaction here. The Chinese members get upstream influence over the framework their entire AI economy depends on — and a venue to harden support for domestic silicon inside the same codebase the rest of the world uses. The Foundation gets the contributors and production-scale engineering that China’s 250+ participating organizations already represent. And the “device-agnostic” work — Huawei’s Ascend as “the first additional platform in PyTorch,” Cambricon’s push for “a native, out-of-the-box developer experience across a broader range of backends” — becomes the mechanism by which non-Nvidia accelerators achieve parity without forking the ecosystem.

Why it matters beyond Shanghai

Three implications stand out.

First, governance follows gravity. Open-weight AI is increasingly a China-origin phenomenon — by some measures, Chinese models account for 30–46% of enterprise usage in the US alone. When the companies behind that gravity take board seats in the framework’s foundation, the governance of open AI infrastructure starts to reflect where its center of mass actually is. Expect the debate over what “vendor-neutral” means to get more interesting.

Second, the multi-backend bet is now institutional. Huawei’s Accelerator Integration Working Group, Ascend’s status as PyTorch’s first additional platform, Cambricon’s device-agnostic keynote — the Foundation is consciously building a PyTorch that treats accelerator diversity as a first-class concern rather than an afterthought. For a hardware ecosystem locked out of CUDA, that is the difference between second-class citizenship and native support.

Third, agents are the new workload shaping infrastructure. Ant Group’s keynote pairing — turning Kubernetes Agent Sandbox and Kata Containers into an on-demand agent runtime — signals that the next infrastructure battle is not model training but agent hosting: secure, isolated, elastic environments where autonomous systems do real work. The cloud-native stack and the AI stack are converging, and this conference is where the merger is being drafted.

None of this comes without tension. The same week these companies joined the Foundation, Huawei’s US racketeering trial opened in Brooklyn, Belgium charged a former semiconductor executive with espionage, and Washington continued weighing further chip-export restrictions. Open-source governance makes strange bedfellows — and the PyTorch Foundation has just invited four of the most strategically significant ones to sit at the same table.

The next milestone is already scheduled: PyTorch Conference North America, San Jose, October 20–21. Watch the multi-backend sessions there. What was announced in Shanghai is a governance change; what follows will be the code.