Opening the code enables broader storage-stack collaboration, but it does not guarantee compatibility with every GPU and storage configuration.
At FMS 2026 on August 4, NVIDIA announced that it was opening the source code for the cuFile API and its supporting software layer. cuFile is part of GPUDirect Storage and is designed to move data between storage and GPU memory without the conventional route through CPU memory.
NVIDIA, Google, Intel, and Meta were named as initial maintainers. NVIDIA is also working with more than 40 memory, storage, and infrastructure vendors through the Storage-Next initiative.
The company also described SCADA, a mechanism intended to let a GPU request selected pieces of data instead of reading an entire dataset. That matters for AI workloads where data preparation and delivery can leave accelerators waiting.
An open API does not remove every bottleneck by itself: compatible GPUs, drivers, file systems, networks, and storage devices are still required. Vendor performance results describe specific tests and should not be treated as guaranteed gains for every server configuration.