At the Future of Memory and Storage (FMS) conference this week, NVIDIA unveiled one of its most significant storage software initiatives to date, announcing that it is open-sourcing its cuFile APIs and the underlying storage software stack that powers GPUDirect Storage.
The move is designed to remove one of the biggest bottlenecks facing AI infrastructure: getting data to GPUs fast enough to keep them properly fed with data and fully utilized.
In a traditional setup, data is loaded from a storage array into the CPU and then into the memory, and then it is transferred to the GPU. That’s a lot of steps and it adds latency. NVIDIA has a technology called GPUDirect, which allows the GPU to communicate directly with the storage device without requiring the CPU or system memory as an intermediary.
That architecture has become increasingly inefficient as AI models grow larger and inference workloads require GPUs to access massive datasets in real time. NVIDIA’s cuFile technology bypasses the CPU, allowing GPUs to read data directly from storage with dramatically lower latency. And now it is being released as open source for the entire industry.
By open sourcing cuFile, NVIDIA hopes to transform what had been a proprietary technology into an industry standard supported by storage vendors, hyperscalers, and systems manufacturers. According to the company, Google, Intel, Meta, and NVIDIA will serve as founding maintainers of the project.
“This upends the old economics of determining when data belongs in memory (where applications can fetch it faster) versus on a storage drive (where it can be held in cheap and plentiful space),” the company said in a blog post announcing the plan. “The tradeoff was first framed 40 years ago, when the answer was measured in accessing that data in minutes. On today’s GPUs, paired with AI storage solutions from NVIDIA and partners, the same tradeoff now plays out in microseconds. Closing the gap between AI’s needs and memory shortage depends on extreme codesign across the whole ecosystem, from memory and storage manufacturers to the software built on them.”
The announcement reflects a broader shift in AI infrastructure design. Rather than treating storage as a passive repository for data, NVIDIA is positioning it as an active component of the AI processing pipeline.
The cuFile announcement was accompanied by the formal launch of Storage-Next, an industry initiative bringing together more than 40 storage, flash memory, controller, cooling, and standards organizations to define open specifications for GPU-native storage.
The effort seeks to establish interoperable standards for storage devices that communicate directly with GPUs rather than relying on host CPUs. Participants include storage vendors such as DDN along with flash manufacturers including Kioxia and Micron.
NVIDIA also emphasized security as part of the initiative. The company said open-sourcing cuFile will make it easier for vendors to integrate secure, high-speed storage into AI environments while enabling interoperability with Linux-based security practices.
The company acknowledges that speed at the storage layer comes with a catch: letting an application talk straight to a drive is quick, but done carelessly, it can scribble over other processes’ memory. That’s a security hole, not a feature.
So NVIDIA invented SCADA, or SCaled Accelerated Data Access, a new storage server architecture designed specifically for AI infrastructure. Its goal is to address AI bottlenecks while maintaining security and data integrity.
It’s all designed to offer fast, massively parallel, efficient, secure AI storage infrastructure that can feed better data to applications and AI factories so they can produce more useful, accurate intelligence at scale.
