NVIDIA's CUDA platform received a strong positive signal from the first AgentPerf benchmark from Artificial Analysis,...

NVIDIA's CUDA platform received a strong positive signal from the first AgentPerf benchmark from Artificial Analysis, published on the NVIDIA blog (https://blogs.nvidia.com/blog/nvidia-blackwell-agentperf-artificial-analysis/). The benchmark measures agentic AI performance, a fundamentally different workload from conversational AI. The NVIDIA Blackwell Ultra NVL72 platform achieved leading performance, running up to 20x more agents per megawatt than the H200 system. The performance is attributed to CUDA kernels and TensorRT LLM, highlighting CUDA's software advantage in this emerging workload. This could drive demand for NVIDIA GPUs and reinforce CUDA's position.

In parallel, the HuggingFace Transformers v5.11.0 release (https://github.com/huggingface/transformers/releases/tag/v5.11.0) added support for new models and fine-grained fp8/fp4 Triton kernels, further integrating with the GPU ecosystem. While Triton is an open-source kernel language, it compiles to CUDA on NVIDIA hardware, indicating continued reliance on the platform.

A Manifold Markets prediction (https://manifold.markets/_deleted_/will-cuda-remain-a-monopoly-for-gpu) shows a 56.95% probability that CUDA remains a monopoly through 2027, reflecting cautious optimism. Together, these events suggest CUDA is strengthening its performance and ecosystem moat in AI workloads.