Nvidia open-sources cuFile, launches Storage-Next and SCADA to fix AI's GPU starvation problem

At the Future of Memory and Storage (FMS) 2026 conference in Santa Clara, Nvidia made three interconnected announcements: open-sourcing the cuFile APIs, formally launching the Storage-Next industry initiative, and productizing the SCADA (Scaled, Accelerated Data Access) framework . Together they target the core AI infrastructure triad of GPU starvation, storage bottlenecks, and storage-security gaps.

What Nvidia announced

cuFile open-source APIs — Nvidia is open-sourcing the cuFile APIs and the full vertical storage software stack underneath them, which allow GPUs (not just CPUs) to read from and write to storage directly . cuFile is a core component of GPUDirect Storage (GDS) . The code is hosted on GitHub under the new xio-sig (Accelerated IO Special Interest Group) organization, with Google, Intel, Nvidia, and Meta as inaugural maintainers . This makes cuFile an open, community-driven standard rather than a proprietary Nvidia-only interface .

A skeptical analysis published shortly after the announcement noted that as of August 4, 2026, there was no public code, license, or migration path yet available for existing GPUDirect Storage deployments, meaning storage architects should treat the move as an announced direction rather than a shipped product .

Storage-Next industry initiative — A formal consortium of over 40 storage and flash vendors including DDN, KIOXIA, and Micron, plus controller makers, cooling specialists, and standards bodies . The goal is to align on how GPU-driven storage should behave and turn those advancements into interoperable, open industry standards .

SCADA framework (Scaled, Accelerated Data Access) — The productized form of Nvidia's earlier BaM (Big Accelerator Memory) research from ASPLOS 2023 . SCADA is a runtime that lets massively parallel GPUs initiate and manage their own storage I/O operations — each of the hundreds of thousands of GPU threads can independently request data from storage, with GPU-side caching and direct NVMe or remote storage access .

How these address the three challenges

GPU starvation (idle GPUs waiting on data)
Traditional host-mediated I/O forces GPUs to wait while the CPU sets up every storage transfer — each operation carries microseconds of kernel-launch and CPU-GPU synchronization overhead . cuFile eliminates the CPU from the data path entirely (DMA directly into GPU memory) . SCADA goes further by removing the CPU from the control path too — threads on the GPU initiate their own reads, batch small scattered requests, and serve them from a GPU-side cache . In identical hardware, the underlying BaM research ran data-analytics workloads 5.3x faster than CPU-initiated access and graph workloads with up to 21.7x lower hardware cost versus storing data in host memory .

AI storage bottlenecks
GPUs can now process data faster than most storage systems can deliver it — that gap is the core bottleneck for AI training and inference . cuFile enables microsecond-level secure data access from storage using hundreds of thousands of GPU threads and high-bandwidth memory . This provides a low-latency "superhighway" between deep storage and GPU memory, critical for retrieval-augmented generation (RAG), mixture-of-experts models, and agentic AI workflows that need massive, rapid parallel data pulls . The Storage-Next initiative ensures this isn't a one-vendor solution: 40+ vendors collaborate on interoperable standards so the entire storage ecosystem can keep pace with GPU speeds .

Cybersecurity challenges
Direct GPU-to-storage access, if unprotected, could let one application read or corrupt another's data . Nvidia's SCADA design separates privilege levels: performance-critical application code runs unprivileged outside the trusted compute base, while a privileged component configures protected access between each application and only its approved storage during setup, using standard Linux security mechanisms . Nvidia frames cuFile as foundational for AI-powered cybersecurity: "fast, secure access to data and storage is a foundational element to powering preventive and detective cybersecurity measures. Making cuFile openly available will help make security context, data and storage accessible at the speed AI-powered defenses need" . This also supports the new Open Secure AI Alliance . The full hardware stack (Nvidia Vera BlueField-4 STX processors) uses the unified NVIDIA DOCA security stack for continuous policy enforcement in the AI data path .

What this means for AI infrastructure

These three announcements from FMS 2026 represent Nvidia's most comprehensive push yet to solve the fundamental architectural mismatch between GPU compute speed and storage delivery. By open-sourcing cuFile, Nvidia is betting that community-driven development will accelerate adoption and interoperability . The Storage-Next consortium provides the industry coordination needed to ensure that flash storage, controllers, and networking can all keep pace with GPU-driven I/O patterns . And SCADA, building on proven BaM research, offers a concrete path to eliminating the CPU as a bottleneck in both the data path and the control path .