NVIDIA at FMS pushes AI storage upgrades, open-sources cuFile to speed GPU access

NVIDIA used this week’s Future of Memory and Storage (FMS) conference in Santa Clara, California to argue that the next performance gains in AI will be constrained less by raw GPU power and more by the storage systems feeding “AI factories.” In a technical post focused on memory and storage pressures, the company said AI’s growing reliance on massive datasets and long context windows is pushing workloads beyond the practical limits of system memory, even when advanced memory is available.

The company’s central point is that increasing storage capacity alone does not solve the problem. Instead, storage has to be designed to handle surges in parallel requests, while continuously performing security and data-processing functions. NVIDIA said AI agents now consume very large amounts of data and, importantly, GPUs can directly initiate storage requests. That combination can generate thousands of concurrent operations hitting the same storage infrastructure.

NVIDIA described several functions that storage systems must perform at scale, including continuous encryption, compression, verification, and reconstruction of data. It warned that these “critical data services” can become bottlenecks when many agents access storage simultaneously. In other words, the storage path must sustain high throughput without falling behind the pace of accelerated compute, or the system’s end-to-end latency and data access efficiency suffer.

To make that case with numbers, the blog highlighted benchmark results involving NVIDIA’s Vera CPU, which is part of the NVIDIA Vera BlueField-4 STX. In a two-stage compression and encryption pipeline, NVIDIA said the Vera CPU can deliver up to 3.21 times higher throughput than an x86 CPU. The company tied that result to the idea that storage platforms can absorb high-volume AI data traffic more efficiently, potentially reducing the amount of compute infrastructure needed for the same storage-side processing load.

Storage integration also sits at the heart of NVIDIA’s software changes at FMS. NVIDIA said it is open sourcing its cuFile application programming interfaces (APIs) and the vertical storage software stack underneath them. cuFile is described as an open source component of NVIDIA GPUDirect Storage, a technology that lets GPUs, not just CPUs, read and write data directly to storage. NVIDIA said the approach uses hundreds of thousands of GPU threads along with other methods to achieve secure data access measured in microseconds.

NVIDIA also framed the cuFile open sourcing as a security-oriented interoperability effort. The company said it is helping unify a security-first storage stack based on Linux best practices, enabling interoperability between GPUs and data. It added that fast, secure access to data and storage can support preventive and detective cybersecurity measures, and that making cuFile openly available can help make security context, data, and storage accessible at the speed AI-driven defenses need. NVIDIA referenced the Open Secure AI Alliance as an initiative that aligns with that objective, and said the open APIs site is intended for contributions and cross-platform optimization with Google, Intel, NVIDIA, and Meta as inaugural maintainers.

Beyond software, NVIDIA announced an initiative called Storage-Next, aimed at aligning storage makers, controller vendors, thermal and cooling contributors, and orchestration operators on how GPU-driven storage should behave. NVIDIA said Storage-Next includes more than 40 leading storage and flash vendors, naming DDN, KIOXIA, and Micron among the participants. The initiative’s stated goal is to translate those behavioral alignments into interoperable, open industry standards, in part to support accelerated data access for large AI datasets.

Central to that effort is a framework NVIDIA called SCADA, short for scaled, accelerated data access. NVIDIA said SCADA allows massively parallel GPUs to pull only the data necessary for an application directly from storage into their own high-speed memory, rather than bringing broader data sets into memory in bulk. The post gave one integration example: DDN is said to be integrating SCADA with Infinia, DDN’s software-defined, AI-native data intelligence platform designed to eliminate storage bottlenecks at scale. DDN Chief Technology Officer Sven Oehme said the collaboration helps create a “more direct, efficient connection” between GPUs and data, keeping accelerated computing resources productive and speeding “time to insight,” a framing that ties infrastructure improvements to customer returns.

NVIDIA positioned these developments as part of its broader AI storage infrastructure work, including NVIDIA Vera BlueField-4 STX. It also described NVIDIA STX as a modular, rack-scale foundation powered by the NVIDIA Vera Rubin platform, NVIDIA Vera BlueField-4 storage processors, and NVIDIA Spectrum-X Ethernet networking. For enterprise security controls in the data path, the company said STX uses a unified NVIDIA DOCA security stack to enable continuous policy enforcement. It additionally mentioned NVIDIA CMX Context Memory Storage, which it described as an AI-native context tier built on STX for long-context, multi-turn, agentic inference.

Even with the detailed architecture explanations, some specifics remain absent from the blog. NVIDIA did not publish full benchmark methodology, workload profiles, or comparative results across a wider range of storage systems and AI agent behaviors. It also described its secure access approach using a split between “user parts” that remain outside the trusted computing base and a separate privileged component that configures protected access at setup, following standard Linux protocols, but it did not quantify operational overhead or compatibility constraints across different enterprise environments. Those details will likely matter to IT teams evaluating rollout timelines and integration risk.

What to watch next is how quickly cuFile and the related open source components are adopted by storage platforms and application developers, and whether Storage-Next’s standardization effort produces interoperable product updates that can reliably deliver the promised microsecond-class data access. Given the increasing emphasis on data path throughput, the near-term competitive battleground may shift from GPU benchmarks alone to end-to-end AI pipeline performance that includes encryption, compression, verification, and reconstruction under high concurrency.