Cambricon, Alibaba Join PyTorch Foundation: Chinese Chips Move Up

The PyTorch Foundation moved deep into Chinese silicon this week. At the inaugural PyTorch Conference China in Shanghai (September 7-9, held alongside KubeCon, CloudNativeCon and the OpenInfra Summit), the Linux Foundation-hosted group announced that Alibaba Cloud and Cambricon (寒武纪) had joined as Platinum members and that Ant Group had joined as a Gold member. The additions seat three more of China's largest AI-infrastructure names inside the body that steers the world's most-used open-source deep-learning framework — and, stacking on Huawei's earlier membership, give Beijing's accelerator builders a direct voice where the software that powers global model training and inference is standardized.

A seat at the standards table

Platinum membership buys one seat on the Governing Board and one on the Technical Advisory Council. Cambricon is sending Jin Wang, its senior director of AI frameworks and infrastructure, to the board and Jing Zhu, a lead maintainer on its PyTorch team, to the TAC. Alibaba Cloud CTO Feifei Li and Ant Group CTO Zhengyu He both framed the decision as the natural continuation of running PyTorch at scale; Li pointed to years spent operating PyTorch across heterogeneous hardware on Alibaba Cloud, while He described it as another step in Ant's long-running participation in global open-source collaboration. The foundation says more than 250 organizations across China already contribute to the PyTorch Foundation project portfolio, which spans vLLM, DeepSpeed, Ray and Safetensors.

The exact calendar matters less than the sequence. Huawei joined in 2023, when its Ascend unit helped start the foundation's Accelerator Integration Working Group and drove the Ascend NPU to become the first additional platform carried natively in PyTorch — the first time a Chinese accelerator stepped outside CUDA's orbit to ship as a first-class backend in the upstream project rather than as a CUDA-compatible imitation. Fred Li, head of Huawei's computing open-source development team, told the conference the standing goal is native Ascend support — native development, native release, native distribution — so that developers reach the same PyTorch experience whichever accelerator they pick.

Upstream first, not CUDA-compatible

Cambricon makes the same "native" pitch for its MLU line. A long-standing PyTorch contributor that describes itself as "Upstream First," it has pushed work in torch.compile, eager operators, device runtime, distributed computing and mixed-precision training back into the mainline, and collaborated with the vLLM community so Cambricon hardware wins day-zero support for leading open-weight models such as DeepSeek-V4 and GLM-5. Its Torch-MLU extension rides PyTorch's PrivateUse1 backend mechanism to deliver operator coverage, Inductor integration, CNCL-based distributed training, graph-mode acceleration and a profiling toolchain. The stated aim, in Elton Gong's words, is "not simply to integrate Cambricon products with PyTorch" but to help PyTorch broaden support for diverse backends — a more generalized device abstraction that lowers forking and migration cost for the whole ecosystem. Cambricon's conference keynote carried the theme literally in its title: "Towards Device-agnostic PyTorch: Building Unified Infrastructure for a Multi-Backend Ecosystem."

What joining really means

Read as strategy, the week was about escaping CUDA more than about Chinese chips per se. PyTorch's device-agnostic backend layer — where accelerators register as first-class citizens instead of being patched in as CUDA's poorer relation — is the genuine battleground. Denied current Nvidia parts by US export controls, China's chipmakers have poured resources into making their own silicon work natively in the one framework global developers already know, betting that an open, multi-backend base lowers the moat a proprietary stack like CUDA depends on. Membership is a low-cost, high-signal move in that contest: it lets a vendor shape the standard rather than fork and fragment it, keeps the developer pool and toolchain compatible, and reduces migration cost for any team moving off Nvidia hardware.

What remains genuinely open is software maturity. A "device-agnostic" abstraction and a curated training stack are different things from two decades of CUDA tooling, libraries and tuned kernels; day-zero support for flagship open models is meaningful progress but not yet a substitute for that accumulated ecosystem. Neither the foundation nor Cambricon has released benchmark or throughput figures this week, and such numbers remain unverified. The significance of these announcements is organizational and directional rather than quantitative: five of China's largest AI-infrastructure players — Huawei, Alibaba Cloud, Cambricon and Ant Group among the members, with hundreds more contributing — now sit inside the foundation, co-defining the software interface that determines how the next generation of AI accelerators is programmed. That, more than any single chip, is what moved upstream this week and why the contest over who shapes the open AI stack is being decided here.