Anyscale has announced an integration of Ray Data with NVIDIA cuDF for GPU-native multimodal data processing, among upcoming capabilities in Ray and its platform, the company said on 16 March 2026.
Anyscale, founded by the creators of Ray, says the integration enables 80% lower cost for multimodal data processing on NVIDIA RTX PRO 4500 Blackwell Server Edition, which it says will be available soon on AWS EC2.
The target workload is turning images, video and documents into AI-ready datasets, a bottleneck for teams fine-tuning vision-language-action models in robotics or scaling document processing for retrieval-augmented generation and search.
Anyscale is also introducing rack-aware scheduling in Ray for NVIDIA GB300 NVL72 clusters, placing distributed workloads such as large-scale reinforcement learning to make use of NVLink interconnects.
The cost figure is the company’s own and depends on the hardware becoming available on AWS.