NVIDIA CUDA 13.5 Brings NVIDIA Vera Rubin GA Support | NVIDIA Technical Blog - NVIDIA Developer

CUDA 13.5 is not a routine release—it is NVIDIA's pre-emptive ecosystem lock-in, shipped roughly 12 months ahead of Rubin mass production. Classifying Rubin as a "General Accelerator" rather than a GPU is an explicit encroachment into the CPU+GPU heterogeneous territory Intel and AMD have defended for over a decade. The compiler backend, memory scheduler, and operator libraries in 13.5 are being rewritten around Rubin's HBM4 architecture. Blackwell-optimized code will not port cleanly. Developers face another forced migration cycle, structurally raising switching costs. On compliance: Rubin as a next-gen flagship will almost certainly trip tighter export-control thresholds. Enterprises hardwiring core AI workloads to CUDA 13.5 are effectively betting on a product line vulnerable to geopolitical severance—software-layer dependency amplifies single-point supply-chain risk multiplicatively. Competitively, AMD's ROCm and Intel's oneAPI have an 18-month window to achieve operator parity. Miss it, and the Rubin cycle cements NVIDIA's 80%+ share of AI training for another full generation. 12–24 month outlook: CUDA 13.5 becomes the de facto entry ticket for the Rubin era. The window for open-source alternatives—Triton, OpenXLA—is narrowing. NVIDIA's true moat has shifted from silicon to compiler infrastructure itself.