Modular

Modular is the company building MAX, an AI inference platform designed by Chris Lattner (creator of LLVM, Clang and Swift) to collapse the fragmented stack between model authoring and production serving. The MAX engine compiles Hugging Face and PyTorch models down to highly optimised kernels that run across NVIDIA, AMD, Apple and CPU backends with a single API and dramatically better performance per dollar than vendor-specific runtimes. Modular also develops Mojo, a Python-superset language that gives kernel authors and model researchers C++/CUDA-level performance without leaving the Python ecosystem; Mojo is increasingly the language of choice for custom GPU kernels in 2026. On top of that is MAX Cloud, a managed inference service for hosting open-weight models with autoscaling, observability and OpenAI-compatible endpoints, and MAX Builds, a registry of pre-packaged optimised models. Modular's pitch — kernel-to-cloud, AMD-friendly, vendor-neutral — has been particularly resonant for teams trying to escape CUDA lock-in or run cost-efficient open models like Llama, Qwen and DeepSeek at production scale. The platform is used by infrastructure teams, AI labs and inference providers who need to squeeze every dollar out of their GPU fleet.