Chris Lattner is the co-founder and chief executive of Modular, the company developing the Mojo programming language and MAX inference platform. The original creator of LLVM and Clang, originator of Swift, and a principal architect of MLIR, he builds the compiler and programming infrastructure that makes complex hardware more accessible to developers.
Lattner studied computer science at the University of Portland and earned his master’s degree and doctorate at the University of Illinois at Urbana-Champaign. His work there with Vikram Adve established LLVM, which grew into foundational infrastructure for compilers, debuggers, and programming languages.
At Apple, where he spent more than a decade beginning in 2005, Lattner developed compiler infrastructure and led developer-tools teams responsible for products including Xcode, Instruments, and Swift Playgrounds. He began developing the Swift programming language in 2010, creating a modern alternative to Objective-C for Apple-platform developers.
After briefly leading Tesla’s Autopilot software organization in 2017, he joined Google, working on TensorFlow, accelerator support, and machine-learning infrastructure. There he helped create MLIR compiler infrastructure, which allows specialized compilers to share reusable representations and optimization machinery across different hardware and programming abstractions. The MLIR research paper, which he co-authored, addresses the mounting complexity of maintaining separate compiler stacks for increasingly diverse accelerators.
From 2020 to early 2022, Lattner led engineering and product teams at SiFive, extending his work into the RISC-V processor ecosystem. In January 2022, he founded Modular with Tim Davis, a former Google colleague, to simplify the software stack connecting AI models with the hardware that runs them.
- Production AI demands infrastructure control. Hosted model APIs accelerate prototyping, but production systems often require control over private data, security, model customization, deployment costs, and hardware selection. Lattner argues that proliferating model-specific frameworks make those requirements unnecessarily difficult to satisfy.
- Hardware portability must extend below the API. Switching a model between CPU and GPU through a single line of code can conceal entirely different libraries, kernels, and runtime behavior. Lattner advocates hardware-portable AI infrastructure with shared compiler, execution, and kernel foundations, while preserving the ability to write specialized accelerator code.
- Python familiarity can coexist with systems-level performance. The Mojo programming language combines Python-oriented syntax with hardware-aware compilation and lower-level control. Its purpose is to let developers implement tokenizers, ordinary loops, branching logic, and custom kernels without routinely moving into separate C++, Rust, or CUDA codebases. Lattner traces Mojo’s design to the practical demands of accelerator-oriented code generation.
- Inference should function as a coherent stack. The MAX inference platform integrates model execution, optimized kernels, compilers, and deployment infrastructure while supporting familiar Python and PyTorch workflows. Modular’s public repository includes Mojo components, accelerator libraries, model pipelines, and inference-serving infrastructure.
- Accelerator leadership depends on software ecosystems. In his analysis of CUDA’s success, Lattner describes NVIDIA’s advantage as a combination of hardware, libraries, tooling, research practices, and developer familiarity. His response is to make alternative hardware practical through better programming models and reusable compiler infrastructure.
