NVIDIA ships native Rust GPU kernel programming in two tracks, and one already runs on stable Rust
NVIDIA Developer Blog·high signal
NVIDIA announced CUDA Rust, compiling Rust directly to PTX rather than wrapping C++ or Python. The SIMT track (cuda-oxide) is early alpha and needs a pinned nightly toolchain plus custom LLVM; the tile track (cutile-rs) runs on stable Rust 1.89+, is already on crates.io, and is in use by Hugging Face's Grout and mistral.rs. Both require CUDA 13.3 or newer and compute capability 8.0 or later, and both enforce memory safety at compile time to catch the aliasing bugs that cause GPU race conditions.