A reference guide to the keynote by Chris Latner, co-founder and CEO of Modular, on September 30, 2026.
Main Theme
The keynote, “Open Source, Open Cloud, Open Software,” addresses the critical challenge of programming a rapidly diversifying and complex hardware landscape for AI. The main message is that the current software paradigm, heavily reliant on proprietary systems like CUDA and developer-hostile languages like C++, is a bottleneck. The solution is a new, unified, open-source stack built on Mojo and Max. This new way of developing software aims to “crack open” the existing silos, democratize AI compute, and give developers, researchers, and hardware vendors the choice and extensibility needed to unlock the full performance of all hardware, from the data center to the edge.
Subthemes and Chapters
Chris structured the talk by first identifying the core problem and then introducing the two fundamental, open-source components of the solution: Mojo and Max.
1. The Problem: A New Way to Develop Software is Needed
The explosion of new, specialized AI hardware (GPUs, TPUs, etc.) has created a “developer hostile” environment. While powerful, these chips are often locked behind proprietary, complex programming models like CUDA and C++. This approach is slow, difficult to use, and prevents developers from accessing the full performance of a wide range of hardware, ultimately making AI more expensive and less accessible. We need a system that is both easy for developers to use and powerful enough to program any chip.
2. Mojo: The Programming Language for All AI Hardware
To solve the programmability problem, Modular created Mojo. It’s not just another language; it’s a new foundation designed for the modern AI era.
- It Feels Like Python, Runs Like C++: Mojo is a member of the Python family, making it familiar and easy to write for the millions of developers who love Python. However, it compiles to high-performance machine code, delivering the speed of C++ or Rust without the complexity or long compile times.
- Unlocks All Hardware: Built on modern compiler technology (MLIR), Mojo is designed from the ground up to target the specialized features of diverse accelerators (NVIDIA, AMD, Google TPUs, etc.). This makes code written in Mojo inherently portable across different chips. As Chris mentioned, “one stack with six architectures is pretty good.”
- A Fully Extensible, General-Purpose Language: While its focus is on unlocking accelerators, Mojo is a complete language you can use for anything from web servers to games. Its open nature allows the community to extend it, and AI agents can even be used to automatically port CUDA code to Mojo, making it portable.
3. Max: The Modern AI Framework for Performance and Portability
Moving up the stack from the language, Max is an AI framework built entirely on Mojo. It is designed to overcome the limitations of older frameworks like PyTorch, which were built for training and struggle with inference and performance on new hardware.
- Designed for GenAI and Deployment: As a new framework, Max is built for modern workloads like GenAI. It excels at inference and is designed to get the full performance out of the hardware, which is critical for managing the total cost of ownership (TCO).
- Co-designed with Mojo: Because Max and Mojo are built together, extending the framework is seamless. Unlike patching PyTorch with foreign kernels from Triton or CUDA, you can write custom, high-performance operations directly in Mojo. This creates a cohesive, “best of both worlds” system.
- A Full-Stack, Open Platform: Max provides a complete, open-source stack, including serving components and modeling libraries. Chris highlighted a key difference: “we don’t lock up things in proprietary kernel libraries.” By open-sourcing everything, including performance-critical kernels, Max empowers researchers and developers to understand, modify, and improve every part of the system.
4. The Alliance: Building an Open Ecosystem for AI
Modular recognizes that one company cannot solve this problem alone. To accelerate the vision of a truly open AI software ecosystem, Chris announced the formation of a new alliance program. The goal is to create “LLVM, but for AI”—an open, collaborative platform where hardware vendors, startups, and researchers can work together to build on a common technology base. This initiative aims to break the mold of proprietary software and ensure that hardware innovation is no longer limited by software fragmentation.
Inspiring Quotes
- “Shouldn’t we have something that writes like Python but runs on C++? That’s what Mojo is.”
- “We’re going to open source everything. You have all the kernels. They run on full performance… And if our stuff isn’t good enough, please help us fix it. We accept patches.”
- “Let’s crack open this big problem. Let’s break the molds. Let’s get people together.”
- “If you’re a researcher, you want an open box. You don’t want arbitrary closed silos that prevent you from actually changing or understanding things.”
- “I want this to end up feeling like LLVM, but for AI, an open platform everybody can use, everybody can build into.”
Key Action Points
- Install and Experiment with the Modular Stack: The easiest way to start is to run
pip install modular. This gives you the full Mojo and Max stack. Explore the documentation and tutorials on modular-lang.org to see how you can achieve C++-level performance with Python-like syntax. - Contribute to the Open-Source Community: The entire stack is being opened up. Check out the projects on GitHub (github.com/modular). Whether it’s fixing a bug, adding a new kernel, or porting a model, your contributions can help build this new ecosystem.
- Explore the Alliance Program: If your organization is a hardware vendor, startup, or research institution that believes in an open, differentiated software platform for AI, consider joining the new alliance. Email alliance@modular.com to get involved in shaping the future of AI infrastructure.