NVIDIA CUDA Rust: Two Tracks, One Already in Production

NVIDIA shipped Rust support for GPU kernel programming at RustConf 2026 on September 8, introducing two tracks: cutile-rs for tile-based ML workloads and cuda-oxide for low-level SIMT-model kernels. cutile-rs runs on stable Rust 1.89+ with CUDA 13.3 and is already in production at HuggingFace's Grout inference engine and mistral.rs, while cuda-oxide requires a pinned nightly Rust toolchain (2026-04-03), LLVM 21+, clang 21+, and CUDA 12.x or newer and is labeled early alpha. NVIDIA also joined the Rust Foundation as a Platinum Member the same day.

NVIDIA shipped Rust support for GPU kernel programming at RustConf 2026 on September 8, and the developer community has had opinions about it ever since — 886 upvotes and 354 comments on Hacker News worth of opinions, to be precise. The announcement introduces two distinct tracks: cuda-oxide for low-level SIMT-model kernels, and cutile-rs for tile-based ML workloads. They are not equally ready, and that distinction matters more than the headline. One Track Is Production-Ready. The Other Isn’t. cutile-rs is the one to pay attention to now. It works on stable Rust 1.89+, requires CUDA 13.3, and installs with a single cargo add cutile . HuggingFace’s Grout inference engine and mistral.rs are already running it in production, with throughput that NVIDIA Research describes as “comparable to vendor-optimized cuBLAS.” That’s not a preview claim — it’s a deployed result. cuda-oxide is a different story. It requires a pinned nightly Rust toolchain 2026-04-03 , LLVM 21+, clang 21+ with libclang headers, a custom LLVM build, and CUDA 12.x or newer. NVIDIA labels it “early alpha” and explicitly warns to expect “bugs, incomplete features, and API breakage.” The GitHub repository has 3,500 stars and genuine engineering ambition, but this is not something to build production infrastructure on today. It’s a project to watch. NVIDIA’s own blog states both projects are early-stage. What it doesn’t fully acknowledge is that cutile-rs has already escaped that framing — real production use at major organizations tells you more than a version label does. The Actual Case for Rust in GPU Kernels The reason this isn’t just “Rust is popular, let’s add Rust” is that GPU kernel aliasing bugs are a specific, nasty problem. They rarely reproduce in development but fail silently in production — exactly the class of bug Rust’s ownership system was built to eliminate. CUDA C++ offers no compile-time aliasing enforcement. Both CUDA Rust tracks do. cuda-oxide uses DisjointSlice<T types to guarantee each thread exclusive access to its element; passing the same buffer as both input and mutable output produces a compile error. cutile-rs goes further: it operates on data blocks rather than individual threads, uses tensor partitioning to grant exclusive tile ownership, and requires zero unsafe blocks for typical ML kernel work. Both eliminate data races through Rust’s &mut exclusivity rules extended to device-side code. This is a legitimate improvement over “write CUDA C++ and hope the sanitizers catch it before the customer does.” Is This Just a Triton Counter-Move? The elephant in the room is OpenAI’s Triton, which has offered a similar abstraction — tile-based GPU programming with high-level safety — since 2022, in Python. Triton-based kernels hit 90–105% of hand-tuned CUDA performance. A kernel a senior CUDA engineer takes three days to write takes four to eight hours in Triton. Nicholas Wilt, a founding member of the original CUDA team, wasn’t subtle about it: “It’s hard not to suspect that cuTile was developed directly to counter Triton.” The community largely agrees — “cuTile feels like NVIDIA’s response to Triton, Mojo, and ThunderKittens” is a representative sentiment. The vendor lock-in debate on Hacker News cut predictably: CUDA critics pushed for Triton, OpenCL, or Metal; pragmatists pointed out CUDA has the talent pool and tooling that alternatives don’t. The honest read: if you’re already a Rust shop running NVIDIA hardware for ML inference, cutile-rs lets you keep your kernel code in the same language as everything else, with real production proof points. If you’re starting fresh and don’t have a Rust constraint, Triton in Python https://openai.com/index/triton/ is still the lower-friction path to custom GPU operators. NVIDIA’s Institutional Bet on Rust The same day as the RustConf announcement, NVIDIA joined the Rust Foundation as a Platinum Member. That’s not a coincidence and not a casual commitment. The company explicitly named Rust’s growing adoption in AI infrastructure systems as the driver, and its roadmap calls for CUDA Rust to mature “into 2027 and beyond” with inter-language interoperability across CUDA C++, CUDA Python, and CUDA Rust. The goal is that choosing a Rust frontend won’t lock you out of the rest of the ecosystem. Combined with HuggingFace Candle, mistral.rs, and Burn’s CubeCL already in active use, the pattern is clear: Rust is becoming a first-class language in AI infrastructure, and NVIDIA is making sure CUDA doesn’t sit out that transition. The Bottom Line If you write Rust for ML inference and run NVIDIA GPUs, cutile-rs is worth evaluating today — the production proof points are real and the installation story is sane. cuda-oxide is genuinely interesting engineering with strong compile-time safety guarantees, but the setup requirements put it in “follow the GitHub” territory rather than “build something with it” territory for most teams. If you’re not in a Rust shop already, nothing about this announcement changes your current setup. Full details are on NVIDIA’s technical blog https://developer.nvidia.com/blog/introducing-cuda-rust-two-tracks-for-writing-gpu-kernels/ . The cuda-oxide repository on GitHub https://github.com/NVlabs/cuda-oxide is the place to follow cuda-oxide’s progress. And the Hacker News discussion https://news.ycombinator.com/item?id=49724881 — 886 points and counting — has the vendor lock-in debate in full if you want to read the community’s unfiltered take.