Alibaba Cloud Launches Next-Gen CPFS Storage for AI Training Clusters

On September 22, Alibaba Cloud unveiled its next-generation full-stack self-developed high-performance storage CPFS at the 2026 Yunqi Conference. Targeting next-generation AI training clusters, the product is built to accelerate model training at million-GPU scale and support large-scale training and multimodal data processing.

According to Alibaba Cloud, in actual training workloads the new CPFS reduces average model startup time by 50%, improves peak compute utilization by 30%, and cuts AI storage costs by 69%, while doubling business capacity. Compared with the previous generation, the file system capacity has expanded fivefold, metadata performance improved 10 times, and the number of supported files increased 100 times, allowing a single file system to scale to 100 PiB and host trillions of files.

At the architecture level, enterprises can manage training corpora, model parameters, checkpoints and experiment results in a unified file system without repeatedly splitting or migrating data for different training phases, and can scale capacity and throughput together with GPU clusters. As large-model training enters the scale-up stage, datasets are expanding from text to images, video and audio, while model parameter sizes have grown about tenfold, making concurrent reads of massive small files and burst writes of checkpoints new bottlenecks.