NVIDIA is addressing AI storage bottlenecks through new hardware acceleration, open-sourced software APIs, and industry

AI is driving demand for massive datasets and longer context windows that exceed system memory capacity. However, storage infrastructure cannot simply scale by adding more capacity—what is needed are efficient, secure storage architectures delivering grounded insights from AI factories.

At the Future of Memory and Storage conference this week, NVIDIA unveiled storage advancements centered on a fundamental shift: with accelerated computing, storage becomes an active part of the data path rather than a passive repository. Thousands of AI agents now generate concurrent storage requests, requiring systems to continuously encrypt, compress, verify and reconstruct data. These operations can become critical bottlenecks under load.

NVIDIA highlighted that its Vera CPU, part of the Vera BlueField-4 STX platform, delivers up to 3.21x higher throughput than an x86 CPU in two-stage compression and encryption pipelines. This allows storage platforms to absorb AI data more efficiently with less compute infrastructure.

The economics of memory versus storage have shifted dramatically. Forty years ago, the tradeoff between fetching data from memory versus cheaper storage was measured in minutes. Today's GPU-paired AI storage solutions now execute that same tradeoff in microseconds.

To advance this ecosystem-wide transition, NVIDIA announced it is open sourcing cuFile application programming interfaces and the vertical storage software stack beneath them. cuFile, a component of NVIDIA GPUDirect Storage, enables GPUs to read from and write to storage directly, using hundreds of thousands of GPU threads and high-bandwidth memory to access data securely in microseconds. This open-source foundation now includes Google, Intel, NVIDIA and Meta as inaugural maintainers.

NVIDIA also introduced Storage-Next, an initiative bringing together over 40 storage and flash vendors—including DDN, KIOXIA and Micron—to align on GPU-driven storage behavior and establish open industry standards. The initiative is grounded in accelerated data access for large AI datasets.

Supporting this, NVIDIA developed SCADA, a framework for scaled, accelerated data access that enables massively parallel GPUs to pull only necessary data directly from storage into their high-speed memory. DDN is integrating SCADA into Infinia, its software-defined, AI-native platform. According to DDN's chief technology officer Sven Oehme, this collaboration creates a more direct, efficient connection between GPUs and data, keeping accelerated computing resources productive and speeding time to insight.

These advancements build on NVIDIA's existing AI storage work, including the Vera BlueField-4 STX—a modular, rack-scale foundation using the unified NVIDIA DOCA security stack for continuous policy enforcement in the AI data path. NVIDIA CMX Context Memory Storage provides an AI-native context tier for long-context, multi-turn agentic AI inference on STX platforms.

Direct application access to storage enables speed but creates security risks if not carefully implemented. NVIDIA SCADA addresses this by splitting the job into two parts: user-facing components requiring raw speed remain outside the trusted computing base, while a separate privileged component configures protected access between user applications and approved storage at setup, adhering to standard Linux protocols while efficiently safeguarding data.

These advancements in fast, massively parallel, efficient, secure AI storage infrastructure enable better data delivery to applications and AI factories, allowing them to produce more useful, accurate, grounded intelligence at scale.