PyTorch as a Portability Layer: Why Cambricon and Ant Group Joined

It is easy to read this story primarily as geopolitics: Alibaba Cloud, Ant Group, Cambricon, and Huawei appeared together at PyTorch Conference China in Shanghai. But a different question is more interesting here. Why would a vendor of its own AI accelerators pay $350,000 per year for Platinum membership in the PyTorch Foundation when PyTorch's code is already open?

First, the news framing needs a correction. The four companies did not all “just join PyTorch.” Cambricon was announced as a Platinum member on September 7, 2026, and Ant Group as a Gold member on September 8. Alibaba Cloud has been a Foundation member since May 26, 2026, and Huawei since October 17, 2023. The new part is Cambricon and Ant. The joint Shanghai announcement brought the Chinese AI stack onto one stage: Qwen models, cloud infrastructure, MLU and Ascend chips, and an agent runtime.

My interpretation is that the Chinese companies are not buying control over PyTorch. They are buying a seat at the table where the cost of moving away from CUDA can be reduced.

For $350,000, a Platinum member receives one voting seat on both the Governing Board and the Technical Advisory Council. Gold membership costs $150,000, but does not guarantee Ant an individual seat: Gold members collectively elect one representative for every three companies. This creates influence over budgets, working groups, CI, and ecosystem priorities. It does not buy the right to press Merge: PyTorch's technical governance assigns authority to individual maintainers based on contribution, not to companies based on membership tier.

This is where the important part begins. The problem for an alternative AI chip is not only FLOPS. It needs operators, a compiler, distributed collectives, profiling tools, library integrations, and compatibility with every new PyTorch release. That long software tail is what turned CUDA into a moat that cannot be closed with a single impressive performance table.

PyTorch is gradually building a common accelerator integration layer through PrivateUse1, the OpenReg reference backend, and the Accelerator Integration Working Group. The idea is for the model and most application code to see a common PyTorch API while CUDA, ROCm, CANN, or Neuware remain underneath. Vendors then compete on backend quality: operator coverage, torch.compile support, distributed-training stability, and release velocity.

Hardware interchangeability is still a long way off. Cambricon currently requires a separate torch_mlu package and vendor components including CNToolkit, CNNL, and CNCL. Huawei is further along: Ascend is already included in the working group's public CI, but torch_npu is still installed separately and depends on CANN. In other words, a common facade is emerging while different staircases, keys, and evacuation instructions remain behind it :)

Ant Group is working on an adjacent layer. The company presented an AI-agent runtime based on Kubernetes Agent Sandbox and Kata Containers. This is neither a PyTorch core contribution nor an accelerator. Instead, it suggests that a broader open-source AI stack is beginning to gather around the Foundation, from the tensor runtime to isolated agent execution. The four companies have not announced a shared technical architecture.

So I would watch three boring engineering signals rather than the number of Chinese logos in the Foundation:

1️⃣ whether Cambricon MLU appears in the common public CI next to Ascend; 2️⃣ whether torch_mlu and torch_npu can support the current stable PyTorch release without special builds or long delays; 3️⃣ whether upstream changes land in Inductor, the distributed stack, and domain libraries instead of growth remaining confined to external vendor repositories.

If that happens, PyTorch will become a real portability layer for AI hardware and modestly reduce NVIDIA's software lock-in. It will not replace CUDA, its mature kernels, NCCL, or its tooling—it will reduce the cost of taking the first step toward another accelerator. If it does not happen, membership will remain an expensive logo on a website.

#AI #OpenSource #PlatformEngineering #Infrastructure #Architecture #Bigtech