NVIDIA has open sourced its cuFile application programming interfaces and the vertical storage software stack beneath them. Google, Intel, Meta and NVIDIA will serve as the inaugural maintainers in the project's new home.
The software lets GPUs read from and write to storage directly instead of routing those requests through CPUs. NVIDIA describes cuFile as an open source component of GPUDirect Storage, which moves data between storage and GPU memory with lower latency.
NVIDIA also outlined a broader push into AI storage through an initiative called Storage-Next, which brings together more than 40 storage and flash suppliers. Participants include DDN, Kioxia and Micron, along with controller vendors, cooling groups, orchestration operators and standards bodies.
At the centre of the effort is a change in how AI systems access data. NVIDIA said AI agents are generating large volumes of storage traffic and that GPUs can now initiate storage requests directly, creating thousands of concurrent operations that require storage systems to encrypt, compress, verify and reconstruct data at the same time.
That load is putting more pressure on infrastructure as AI models use larger datasets and longer context windows that exceed system memory. NVIDIA argues this shifts storage from a passive repository to an active part of the data path for AI workloads.
Open stack
By releasing cuFile and the software stack beneath it, NVIDIA aims to broaden interoperability across hardware and software platforms. The project is open to external contributions and is intended to align GPU access to storage with Linux-based security practices.
Fast and secure storage access also matters for cyber defence, where AI systems need rapid access to security context and data. NVIDIA linked the open source move to broader industry efforts around secure AI infrastructure.
cuFile uses GPU threads and high-bandwidth memory to access data stored outside system memory. According to NVIDIA, that approach can cut storage access times to microseconds.
Storage-Next
Storage-Next is intended to define how GPU-driven storage should operate and translate those approaches into open industry standards. The initiative focuses on moving large AI datasets and reducing bottlenecks between compute and storage layers.
As part of that work, NVIDIA is promoting a framework called SCADA, short for scaled, accelerated data access. It is designed so parallel GPUs pull only the data an application needs directly from storage into GPU memory.
NVIDIA said SCADA splits direct data access into two parts to address security risks that can arise when applications communicate closely with storage devices. Under that design, user-level application components remain outside the trusted computing base, while a privileged component establishes protected access between the application and approved storage using standard Linux security protocols.
NVIDIA tied those storage efforts to its broader infrastructure products, including Vera BlueField-4 STX, which combines Vera Rubin, Vera BlueField-4 storage processors and Spectrum-X Ethernet networking. According to the company, STX uses the DOCA security stack for continuous policy enforcement in the AI data path, while CMX Context Memory Storage provides a context tier for long-context and multi-turn AI inference.
Performance claims
NVIDIA also pointed to benchmark data for the Vera CPU, which forms part of Vera BlueField-4 STX. According to figures highlighted by the company, the processor delivered up to 3.21 times higher throughput than an x86 CPU in a two-stage compression and encryption pipeline.
The benchmark is meant to show that storage functions such as encryption, compression, integrity checking and recovery can be handled with less strain on conventional host compute. That matters because those functions can become a bottleneck when many AI agents access storage at once.
For storage suppliers, the changes reflect a broader architectural shift rather than a simple increase in storage capacity. The challenge is increasingly about moving, securing and processing data at a rate that matches modern GPU systems, especially as the historical gap between memory and drive access times narrows to microseconds.
One early implementation comes from DDN, which is integrating SCADA with its Infinia software-defined data platform as part of the broader Storage-Next effort.
"AI success will be defined not by how much infrastructure organizations own, but by how productively they use it," said Sven Oehme, Chief Technology Officer, DDN. "Our collaboration with NVIDIA is helping create a more direct, efficient connection between GPUs and data - keeping accelerated computing resources productive, speeding time to insight and enabling customers to achieve stronger business and financial returns from their AI investments."