AIBRIEF
The cuFile API moves data directly from NVMe drives to GPU memory; 40 storage and flash vendors are signing on to standardize the path through a new Storage Next consortium.
Nvidia open-sourced the cuFile API at the Future of Memory and Storage conference on Tuesday and pulled 40 storage and flash vendors into a new consortium called Storage-Next, formalizing a direct NVMe-to-GPU data path that had previously lived inside its proprietary GPUDirect Storage software.
cuFile lets a GPU pull data directly from NVMe drives using direct memory access (DMA), bypassing the CPU and system main memory. The interface first reached general availability in July 2021 as part of CUDA Toolkit 11.4. The code is now on GitHub, and the GPUDirect Storage documentation remains the canonical reference for how the pieces fit together.
The Storage-Next consortium, anchored by DataDirect Networks and 39 other storage and flash memory vendors, gives that fast lane a shared industry target. Nvidia framed the move as a fix for "GPU starvation," the pattern where expensive accelerators sit idle waiting for data. That bottleneck matters more as retrieval-augmented generation and agentic AI workloads push larger context windows and longer-running inference jobs through the same hardware.
What is still unknown: the full Storage-Next roster beyond DataDirect Networks, and whether independent benchmarks confirm the millisecond access times Nvidia describes. No third-party vendor reaction was visible at announcement time.