NVIDIA's CUDA toolkit has long provided the foundation for GPU-accelerated computing. GPUs — originally designed for graphics — excel at parallel processing, powering everything from deep learning to scientific simulation. Until now, developers had to write CUDA code in C or C++ (or rely on higher-level bindings) to unleash a GPU. With its recent announcement, NVIDIA is breaking that barrier: Python developers can now tap directly into CUDA's performance without leaving Python.
Python has surged into the world's most popular language for AI and data science, but its interpreted nature long made it seem “too slow” for high-performance computing. Tools like Numba and CuPy bridged the gap — but they were still workarounds, tapping GPU power indirectly. NVIDIA's native support eliminates that gap, letting Python call CUDA APIs directly.
At GTC 2025, NVIDIA declared 2025 the “Year of Python” and made Python a first-class citizen in the CUDA ecosystem: developers can launch GPU kernels, manage device memory, and invoke CUDA libraries with no C++ in between. It's not just a new wrapper — it's a full integration of Python into CUDA's toolchain.