China's largest tech companies are now shaping the future of open source artificial intelligence infrastructure, with Alibaba Cloud, Ant Group, and Cambricon joining the PyTorch Foundation as major members. This expansion reflects a broader shift in how AI computing power is being developed and deployed worldwide, moving beyond proprietary systems toward collaborative, open ecosystems that span hardware, software, and cloud platforms.

Why Are Chinese Tech Giants Joining PyTorch Now?

The PyTorch Foundation announced three new members during its conference in Shanghai on September 8 and 9, 2026. Alibaba Cloud joined as a Platinum Member, bringing expertise in cloud infrastructure and the Qwen model family. Ant Group joined as a Gold Member, contributing experience running AI systems at financial services scale. Cambricon, a hardware accelerator developer, also joined as a Platinum Member.

This timing matters. The open source AI stack has become critical infrastructure for enterprises worldwide. Rather than relying on a single vendor's proprietary tools, companies increasingly want flexibility to run AI models on different hardware, across multiple cloud providers, and with customizable software layers. Chinese companies have built significant expertise in this area and are now formalizing their role in shaping global standards.

"No single organization builds the whole system. Together, our members span it," the PyTorch Foundation stated in its press conference summary.

PyTorch Foundation

This philosophy underpins the entire initiative. The Foundation outlined an open source AI stack spanning five layers: applications and agents at the top, building and delivering intelligence, running and scaling AI workloads, infrastructure and isolation, and heterogeneous compute at the foundation.

What Does This Mean for AI Infrastructure Development?

The expansion signals that open source AI infrastructure is no longer a Western-dominated effort. More than 250 organizations across China now contribute to PyTorch Foundation projects, including DeepSpeed, Helion, PyTorch, Ray, Safetensors, and vLLM. Over 100 China-based developers contribute to PyTorch specifically across 40 or more affiliated organizations.

The practical impact is visible in model development. One case study presented at the conference showed how DeepSeek-R1, an open source model, was optimized through kernels, routing, parallelism, and serving improvements. The optimized system delivered 2.77 times higher throughput and 60 percent lower token cost compared to earlier configurations.

Another metric underscores the shift: the share of OpenRouter token traffic served by open-weight models developed in China increased from 2 percent in late 2024 to 45 percent in April 2026. This dramatic rise shows that open source models from China are now handling nearly half of all inference traffic on a major routing platform.

How to Participate in Open Source AI Infrastructure Development

• Join the Accelerator Integration Working Group: Hardware vendors and developers can participate in PyTorch's TAC Accelerator Integration Working Group, co-chaired by Huawei and Intel, which delivers standardized hardware onboarding guidelines and cross-repository testing mechanisms for new accelerators.
• Contribute to Core Projects: Developers can contribute to PyTorch Foundation projects including vLLM for model serving, Ray for distributed computing, and Safetensors for model storage, all of which are actively maintained by global communities.
• Adopt Multi-Backend Infrastructure: Organizations can build systems using device-agnostic APIs and unified infrastructure that work across NPUs, CPUs, GPUs, and other accelerators, reducing vendor lock-in and enabling flexibility.

Cambricon's approach exemplifies this collaborative model. The company develops MLU accelerators and follows an "Upstream First" approach to its PyTorch contributions, meaning it contributes improvements directly to the open source project rather than keeping them proprietary. Its contributions span torch.compile, Eager Operators, Device Runtime, Distributed Computing, Automatic Mixed Precision (AMP), Dataloader, and Profiler.

What Hardware and Software Layers Are Being Standardized?

The PyTorch Foundation's vision centers on four core principles: Any Model, Any Chip, Any Cloud, and Any Agent. This means AI systems should work regardless of which model you choose, which hardware accelerator you use, which cloud platform you deploy on, or which production service you build.

To achieve this, the Foundation identified key software components across the stack. For building and delivering intelligence, PyTorch, vLLM, and Ray represent the core layer. For running and scaling AI workloads, Kubernetes, KServe, Kueue, OpenTelemetry, and llm-d provide orchestration and observability. For infrastructure and isolation, OpenStack and Kata Containers handle virtualization and security.

The Accelerator Integration Working Group addresses a fundamental challenge: AI computing hardware faces significant heterogeneity, meaning different accelerators require different software adaptations. This creates high costs and fragmentation. The working group delivers standardized hardware onboarding guidelines, cross-repository continuous integration testing, generalized device-aware test suites, and platform incubation workflows to reduce these barriers.

Huawei and Intel co-chair this effort, reflecting the global nature of the initiative. During the conference, Zesheng Zong from Huawei shared achievements including refined device-agnostic APIs and an expanded multi-backend test matrix, signaling progress toward truly interoperable AI infrastructure.

The PyTorch Foundation also welcomed Beijing Academy of Artificial Intelligence (BAAI) as an Associate Member, further strengthening the Chinese research community's role in shaping open source AI standards. With Alibaba Cloud, Cambricon, and Huawei as Platinum Members and Ant Group as a Gold Member, China now has substantial influence over how the global open source AI stack evolves.