Nvidia Launches CUDA Rust: Now Write GPU Kernels in Rust

Nvidia has launched CUDA Rust — two paths for writing GPU kernels (code that runs directly on the GPU): SIMT via cuda-oxide on nightly Rust, and Tile via cutile-rs on stable Rust 1.89+. Tile kernels compile once for any GPU architecture and are already in use in Hugging Face's Grout inference engine and mistral.rs.

Rust in GPU programming isn't a trend — data races in CUDA code are notoriously hard to catch at debug time, and Rust eliminates them at compile time. I don't write GPU kernels in my day-to-day work with agents, but solid inference infrastructure directly affects the speed and reliability of the APIs I use every day.