At the Open Source Summit in Denver, Matt White, Executive Director of the PyTorch Foundation, laid out a vision that goes well beyond maintaining a machine learning framework. The foundation has now become an umbrella organization supporting a growing ecosystem of projects, standards, and community efforts — all focused on making AI infrastructure more open, more accessible, and more robust.

From Framework to Foundation

“PyTorch is the de facto framework of choice for deep learning,” White explained. It’s used by nearly every major lab and enterprise building large language models (LLMs) today. But its roots run deeper — from fraud detection to computer vision and classical NLP, PyTorch has quietly powered much of modern AI.

To meet the demands of today’s evolving ecosystem, the PyTorch Foundation has transitioned into an umbrella organization. It now hosts projects like vLLM — a high-performance inference framework — and DeepSpeed, an orchestration framework for scalable training. The goal is to enable tight collaboration across projects and provide a stable, neutral home for AI innovation.

Defining the Modern AI Stack

White broke down the current AI value chain into key stages: data ingestion, preprocessing, training, fine-tuning, inference, and deployment. “We’re transitioning from very model-centric to very system-centric now,” he said. As a result, the ecosystem is expanding rapidly — from low-level infrastructure like CUDA and silicon vendors to orchestration tools like LangChain and agentic communication protocols like A2A and MCP.

The PyTorch Foundation aims to be the connective tissue in this layered environment, supporting projects across the entire AI lifecycle — whether they emerge from big tech, academia, or community Discord servers.

Training, Certification, and Community Growth

With PyTorch now serving as a cornerstone in many production AI systems, the foundation is investing in education and community. Its ambassador program launched recently and has already received strong interest. Training and certification programs are also in development, with a focus on building talent pipelines that connect academia and industry.

“We’re running PyTorch Days around the world,” said White, noting recent events in China and France, along with workshops at conferences like NeurIPS and MLSys. “Our intent is to build a vibrant, trusted community — just like CNCF did for cloud-native.”

Solving AI’s Licensing Problem 

One of the most urgent challenges in open AI is licensing. Traditional open source licenses like MIT and Apache 2.0 were never designed for the complexity of AI models, which include not just code, but data, weights, and other artifacts.

To address this, the Linux Foundation quietly released the OpenMDW license — short for Model, Data, and Weights. The license allows model developers to specify how each component is licensed, while maintaining compatibility with permissive standards.

“There’s definitely an appetite in the industry,” White noted. “We just need that first adopter.” With interest already growing from AI research labs, OpenMDW could help standardize how models are shared, studied, and built upon — without the ambiguity of legacy licenses or the restrictions of community use licenses.

Embracing Open Source AI

For White, the push for openness goes beyond models. He defines open source AI as any platform, tool, or system released under OSI-approved licenses — including frameworks, datasets, and training infrastructure. “Open source AI has been thriving for quite a few years,” he said, pointing to early tools like scikit-learn and Caffe.

The key today is collaboration. “There’s a lot of cross collaboration between organizations and academia,” White said. “It’s a very exciting time. And the PyTorch Foundation is here to support that — wherever the innovation comes from.”