NVIDIA open-sources cuFile APIs, launches Storage-Next initiative with 40+ storage vendors

NVIDIA is open-sourcing the cuFile APIs that let GPUs read from and write to storage directly, and it has launched Storage-Next, an industry-wide initiative with more than 40 storage and flash vendors to build the next generation of AI-focused storage technology. Both moves were announced at the FMS (Future of Memory and Storage) conference, alongside a new AI-native storage tier called NVIDIA CMX Context Memory Storage.

What's new

The headline move is open access to GPU-direct storage. "At FMS, NVIDIA announced it is open sourcing its cuFile application programming interfaces (APIs) — and the vertical storage software stack underneath them — which let GPUs, not just CPUs, read from and write to storage directly," NVIDIA says. Historically, moving data between storage and GPU memory has routed through the CPU, adding latency and consuming CPU cycles that could otherwise go to other work; cuFile bypasses that path so GPUs can pull data straight from storage.

Alongside the open-sourcing, NVIDIA is organizing the storage industry around a shared roadmap. "Storage-Next includes over 40 leading storage and flash vendors — including DDN, KIOXIA and Micron — each contributing to the next generation of AI storage technologies with NVIDIA," the company says. The scale of that vendor list signals an attempt to set a de facto standard for how storage hardware talks to GPUs, rather than leaving each vendor to build its own proprietary integration.

NVIDIA also introduced a new storage tier purpose-built for AI inference workloads. "NVIDIA CMX Context Memory Storage provides an AI-native context tier for long-context, multi-turn, agentic AI inference, built on NVIDIA STX," the company says — aimed at the growing class of agentic applications that need to retain and retrieve large amounts of conversational or task context across many turns, more than fits comfortably in GPU memory alone.

Context

As models handle longer context windows and multi-turn agentic workloads, the datasets and intermediate state involved routinely exceed what fits in a system's memory, pushing more of the AI pipeline's performance bottleneck onto storage. NVIDIA has been building toward this shift with its Vera CPU and BlueField platforms, which are designed to handle compression, encryption, and data-path work with far higher throughput than general-purpose x86 CPUs, freeing storage to act as an active, low-latency part of the pipeline rather than a passive archive tier. Open-sourcing cuFile removes a barrier that previously limited GPU-direct storage access to NVIDIA's own stack and a narrower set of certified partners; making it open, combined with Storage-Next's 40-plus-vendor coalition, extends that architecture across the broader storage industry.

Why it matters

Storage has historically lagged compute and networking in AI infrastructure conversations, but as context windows grow and agentic systems keep more state alive across long-running tasks, storage throughput becomes a direct constraint on model performance and cost. By open-sourcing cuFile, NVIDIA lowers the barrier for storage vendors to build GPU-direct products without negotiating individual technical partnerships, which should accelerate hardware availability and competition in GPU-attached storage. The Storage-Next coalition's breadth — spanning both established players like Micron and specialized flash vendors like KIOXIA and DDN — suggests the storage industry sees GPU-direct access becoming a baseline requirement rather than a premium feature, and CMX gives NVIDIA a concrete answer to the question of where agentic AI systems keep the context they can't hold in memory.