PyTorch Conference China 2026 brought the PyTorch community together in Shanghai on September 8–9 alongside KubeCon + CloudNativeCon and OpenInfra Summit, following sponsor-hosted co-located events on September 7. Technical discussions spanned models, frameworks, distributed training, inference, hardware, cloud native infrastructure, and agents.
“Open Source for the AI Era” framed work across those layers. Across co-located sessions, keynotes, technical demonstrations, a PyTorch Foundation press conference, community meetings, and conversations at the PyTorch booth, the program covered hardware adaptation, training and serving, open infrastructure, accelerator integration, and collaboration across open source communities.
PyTorch Foundation welcomed Alibaba Cloud, Ant Group, and Cambricon as new members, joining Huawei and other existing Foundation members.
Building Frontier Intelligence in the Open
Speaker: Mark Collier, Executive Director, PyTorch Foundation
Mark’s September 8 keynote focused on an open ecosystem where researchers, developers, model builders, and hardware companies can work together across an expanding landscape of models and compute. The session connected PyTorch with the path from experimentation to production, new accelerator architectures, and the infrastructure that ultimately runs AI workloads.
Additional Keynotes Across the Open Source AI Stack
The keynote program brought together speakers working across models, PyTorch, cloud native infrastructure, accelerators, serving, and agents.
Day 1
From Open Source Adoption to Open Source Innovation: China’s Next Chapter
Speaker: Professor Lu Shouqun
Organization: China OSS
Open Source for the AI Era
Speakers: Jonathan Bryce and Horace Li
Organization: The Linux Foundation
Beyond Multimodal: Building Full-Modal AGI via Open-Weight
Speaker: Ryan Lee
Organization: MiniMax
Operating Frontier Intelligence at Scale
Speakers: Chris Aniszczyk and Xiao Zhang
Organizations: The Linux Foundation and Dynamia.ai
Tidal Auto-scaling for Training and Inference Based on Kubernetes + KEDA
Speakers: Jun Zheng and Tan Pei Xiang
Organization: China Merchants Bank
Ascend & PyTorch: Pioneer New AI Open Ecosystem
Speaker: Liang Zhang
Organization: Huawei
Towards Device-agnostic PyTorch: Building Unified Infrastructure for a Multi-Backend Ecosystem
Speaker: Wei Li
Organization: Cambricon
Inside vLLM: Production Best Practices, Model Integration and Road Map
Speaker: Kaichao You
Organization: Inferact Inc.
PD Disaggregation vLLM Deployment on Alternative AI Accelerators Using llm-d
Speakers: 纪飞 王 and Mengxuan Li
Organizations: Dynamia and Dynamia.ai
Building an Agent Runtime with Open Infrastructure
Speakers: Yaya Xia and Xu Wang
Organization: Ant Group
What AI Agents Need from Open Infrastructure
Speaker: Yaya Xia
Organization: Ant Group
Day 2
Welcome Back + Opening Remarks
Speakers: Jonathan Bryce and Horace Li
Organization: The Linux Foundation
What Powers Frontier Intelligence
Speaker: Thierry Carrez
Organization: OpenInfra Foundation
Road from Kata Containers to Confidential Containers + GPUs: From First Commit to CNCF Incubation
Speaker: Zvonko Kaiser
Organization: NVIDIA
HyperParallel: A SuperPoD-Aware Distributed Acceleration Library
Speakers: Teng Su and Shendi Wang
Organization: Huawei
Build a Unified Heterogeneous AI Computing Ecosystem for PyTorch
Speaker: Zesheng Zong
Organization: Huawei
Serving Qwen at Scale: Multi-Cluster AI Infrastructure on Karmada
Speaker: Jionghang Cai
Organization: Alibaba Cloud
Meet the Community Behind the Open Source AI Stack
Speakers: Jane Lyu; Fupan Li; Zesheng Zong; Hiu Yeung, Sunny Chan
Organizations: The Linux Foundation; Ant Group; Huawei; TCC Consulting Limited
Building Frontier AI Infra: SGLang and Miles
Speaker: Ke Bao
Organization: RadixArk
A Cloud Native Stack from Bare Metal to Tokens for Large-Scale AI Inference
Speaker: Trong Vinh Nguyen
Organization: Viettel
Alibaba Cloud, Ant Group, and Cambricon Join PyTorch Foundation
PyTorch Foundation announced three new members in Shanghai on September 8.
Alibaba Cloud joined as a Platinum Member, with work spanning cloud infrastructure and the Qwen model family.
Ant Group joined as a Gold Member, bringing experience with production AI at financial-services scale.
Cambricon joined as a Platinum Member. Cambricon develops MLU accelerators and follows an “Upstream First” approach to its PyTorch contributions. Its contributions span torch.compile, Eager Operators, Device Runtime, Distributed Computing, AMP, Dataloader, and Profiler.
Representatives from Alibaba Cloud, Ant Group, Cambricon, and Huawei took the keynote stage to speak about the open AI stack across hardware, models, and infrastructure.
ANY MODEL · ANY CHIP · ANY CLOUD · ANY AGENT
Read the full membership announcement.
Read more about Cambricon joining PyTorch Foundation as a Platinum Member.
One Open Source AI Stack
At the September 8 press conference, PyTorch Foundation outlined an open source AI stack spanning applications and agents, building and delivering intelligence, running and scaling AI workloads, infrastructure and isolation, and heterogeneous compute.
PyTorch, vLLM, and Ray represented the layer focused on building and delivering intelligence. Kubernetes, KServe, Kueue, OpenTelemetry, and llm-d represented the layer focused on running and scaling AI workloads. OpenStack and Kata Containers represented infrastructure and isolation. Those software layers run across NPUs, CPUs, GPUs, and other accelerators.
PyTorch Foundation members work across four areas:
- Any Model: AI labs and builders
- Any Chip: silicon and systems
- Any Cloud: clouds and platforms
- Any Agent: production and services
Many members work across more than one layer. The Foundation summarized the model directly: “No single organization builds the whole system. Together, our members span it.”
China in the Stack
100+ China-based developers contribute to PyTorch across 40+ affiliated organizations.
PyTorch Foundation members represented in the China-focused overview included Alibaba Cloud, Cambricon, and Huawei as Platinum Members, Ant Group as a Gold Member, and Beijing Academy of Artificial Intelligence (BAAI) as an Associate Member.
The September 8 membership announcement also noted that more than 250 organizations across China contribute to PyTorch Foundation projects, including DeepSpeed, Helion, PyTorch, Ray, Safetensors, and vLLM.
The conference also highlighted open model development and optimization through the open software layer.
For DeepSeek-R1, one case study covered optimization through kernels, routing, parallelism, and serving on the same GB300 hardware six months later. In the configuration presented, the optimized system delivered 2.77x throughput and 60% lower token cost. The source cited in the conference material was NVIDIA, 2026.
A second example showed the share of OpenRouter token traffic served by open-weight models developed in China increasing from 2% in late 2024 to 45% in April 2026. The source cited in the conference material was Mozilla.
Join the PyTorch TAC Accelerator Integration Working Group
AI computing hardware faces heterogeneity challenges with high adaptation costs and a lack of unified standards.
The PyTorch TAC Accelerator Integration Working Group, co-chaired by Huawei and Intel, delivers standardized hardware onboarding guidelines, a cross-repo CI testing mechanism, generalized device-aware test suites, and platform incubation workflows.
Zesheng Zong shared the Accelerator Integration Working Group’s achievements and future roadmap during PyTorch Conference China, including refined device-agnostic APIs and an expanded multi-backend test matrix.
Hardware vendors and developers interested in co-building an open, efficient heterogeneous computing ecosystem for PyTorch can participate in the Accelerator Integration Working Group.
Join a PyTorch Foundation Working Group.
Community in Shanghai
PyTorch Foundation Technical Advisory Council members met in Shanghai during Day 1 of PyTorch Conference China 2026.
At the PyTorch booth, developers, contributors, hardware vendors, and attendees discussed work across the open source AI ecosystem.
September 7: Starting with Diverse Hardware
Two September 7 co-located events highlighted the challenges created by increasingly heterogeneous AI hardware and the software work required to support it.
Open Computing: Building Open Software for Diverse Hardware
Speaker: Mark Collier, Executive Director, PyTorch Foundation
Host: FlagOS
Mark Collier joined experts from Beijing Academy of Artificial Intelligence (BAAI), Shanghai AI Laboratory, vLLM, SGLang, NVIDIA, and others in a technical discussion about open system software for diverse hardware.
AI chip diversification creates repeated integration work across operators, compilers, training and inference frameworks, communications, and other parts of the software stack. Talks focused on technical roadmaps and collaboration models for a multi-chip AI system software stack, including hardware-software co-optimization across emerging chip architectures.
Hardware-Aware AI: Building PyTorch Workloads Across Accelerators
Opening remarks: Mark Collier, Executive Director, PyTorch Foundation
Host: Huawei
Huawei’s co-located session focused on PyTorch workloads across GPUs, NPUs, and XPUs, including framework adaptation, operators, compiler co-optimization, workload migration, and performance tuning.
Register for PyTorch Conference North America 2026
If you couldn’t join us in Shanghai, PyTorch Conference North America 2026 is the next chance to hear directly from maintainers, researchers, developers, and AI engineers working across the open source AI stack.
PyTorch conferences are the open source AI community’s town square, where what’s next gets decided.
PyTorch Conference North America comes to San Jose October 20–21, with maintainers from PyTorch, vLLM, DeepSpeed, and Ray among the speakers.
Facts Only
* PyTorch Conference China 2026 occurred September 8–9 in Shanghai.
* Co-located events took place on September 7 and alongside KubeCon + CloudNativeCon and OpenInfra Summit.
* Alibaba Cloud and Cambricon joined the PyTorch Foundation as Platinum Members.
* Ant Group joined the PyTorch Foundation as a Gold Member.
* Huawei is an existing Foundation member.
* Mark Collier serves as Executive Director of the PyTorch Foundation.
* The PyTorch TAC Accelerator Integration Working Group is co-chaired by Huawei and Intel.
* Over 100 China-based developers from 40+ organizations contribute to PyTorch.
* More than 250 organizations in China contribute to projects including DeepSpeed, Helion, PyTorch, Ray, Safetensors, and vLLM.
* A cited NVIDIA 2026 case study on DeepSeek-R1 showed 2.77x throughput and 60% lower token cost after optimization.
* Mozilla data indicated OpenRouter token traffic from Chinese open-weight models rose from 2% in late 2024 to 45% in April 2026.
* PyTorch Conference North America 2026 is scheduled for October 20–21 in San Jose.
Executive Summary
PyTorch Conference China 2026 took place in Shanghai on September 8–9, co-located with KubeCon + CloudNativeCon and the OpenInfra Summit. The event centered on "Open Source for the AI Era," focusing on the integration of models, frameworks, and hardware to create a device-agnostic AI stack. Key themes included the transition from AI experimentation to production and the necessity of standardized infrastructure to support a diverse array of accelerators, including GPUs, NPUs, and XPUs.
The PyTorch Foundation expanded its membership with the addition of Alibaba Cloud (Platinum), Cambricon (Platinum), and Ant Group (Gold), joining existing members like Huawei. Technical discussions highlighted the role of open-weight models and the efficiency gains achieved through software optimization, as seen in DeepSeek-R1 case studies. While the event emphasized a collaborative "Any Model, Any Chip, Any Cloud" philosophy, the practical application relies on ongoing work by the TAC Accelerator Integration Working Group to reduce the high costs of hardware adaptation.
Full Take
The strongest version of this narrative is one of global technical democratization. By building a "device-agnostic" stack, the community aims to prevent vendor lock-in and lower the barrier to entry for AI development, ensuring that frontier intelligence isn't gated by the ownership of a specific chip architecture.
However, the narrative operates through a distinct "ecosystem" framing. By emphasizing a "Unified Heterogeneous AI Computing Ecosystem," the focus shifts from individual corporate competition to a collective infrastructure play. This is a strategic alignment: when hardware becomes commoditized through standardized software layers (like the TAC Accelerator Integration Working Group), the value shifts from the chip itself to the orchestration and the models. The mention of Chinese open-weight models capturing 45% of OpenRouter traffic serves as a signal of shifting geopolitical influence in the AI supply chain, framed here as a technical achievement of open source.
The underlying paradigm is "Co-opetition"—competitors like Alibaba, Huawei, and Ant Group collaborating on the foundation layer to ensure their specific hardware or cloud services remain compatible with the dominant framework. This echoes historical patterns in the early internet and Linux development, where industry rivals standardized the plumbing to compete on the services.
Patterns detected: none
The root cause is the urgent need to decouple AI software from proprietary hardware silos. If the "Any Chip" vision succeeds, human agency increases as developers gain the freedom to migrate workloads based on cost and efficiency rather than contractual obligation.
Bridge Questions:
1. If a "unified" standard is co-chaired by a few dominant players, does it actually prevent lock-in or simply move the lock-in to the steering committee level?
2. How does the rise of open-weight models from specific regions affect the global governance of AI safety and alignment?
Counterstrike Scan: A coordinated campaign to project "technological independence" would emphasize the number of local contributors and the success of local hardware in a global framework to build confidence in regional autonomy. The content aligns with this pattern but remains within the bounds of a standard corporate/community announcement.