The HuggingFace Transformers library released v5.11.0 on June 10, 2026, which includes 'Triton finegrained fp8/fp4'...

The HuggingFace Transformers library released v5.11.0 on June 10, 2026, which includes 'Triton finegrained fp8/fp4' kernel support (source 2). Triton is an open-source GPU programming language developed by OpenAI that directly competes with NVIDIA's CUDA. By integrating Triton kernels as a first-class option, HuggingFace is enabling developers to run models without being tied to CUDA-specific implementations. This move, while incremental, signals that the ML ecosystem is diversifying its GPU programming stack. Meanwhile, NVIDIA's CUDA Toolkit 13.2 (as of June 14, 2026, source 1) continues to enhance its own features, including Tile C++ and Python support, but the presence of Triton in a major library like Transformers could accelerate developer adoption of alternatives. A prediction market on Manifold (source 3) shows only 56.95% belief that CUDA will remain a monopoly through 2027, indicating growing doubt. This event alone is not decisive, but combined with other signals (like OpenAI's Triton gaining traction), it points to potential erosion of CUDA's dominant position. The direction is therefore negative for CUDA's long-term developer lock-in. Confidence is low given this is a single data point from one library, but the pattern is notable.