The AI infrastructure race is no longer just about buying faster GPUs.
Those processors need a constant supply of data. As AI agents work with larger context windows, company records, stored memories and thousands of simultaneous tasks, the systems sitting behind the GPUs are starting to feel the strain.
NVIDIA used the Future of Memory and Storage conference in Santa Clara to outline how it plans to tackle that problem. The company announced that it is opening its cuFile application programming interfaces, expanding industry work around GPU-driven storage and pushing storage closer to the role traditionally handled by system memory.
The pitch is fairly direct: storage can no longer behave like a quiet warehouse for data. It needs to become an active part of the AI computing pipeline.
AI Agents Are Creating a Storage Bottleneck
Traditional software may make a limited number of predictable storage requests.
AI agents do not always behave that way.
One agent can retrieve company documents, access persistent memory, reuse cached information, run external tools and save newly generated results. Multiply that across thousands of agents and the number of storage operations climbs quickly.
The data also needs work before it reaches the model. Storage systems may encrypt files, compress them, verify their integrity or reconstruct missing information. Each operation uses processing resources and can slow the broader AI workload when too many requests arrive at once.
Adding more drives will not solve the whole problem. Faster storage is useful, but the surrounding processors, security services and data paths must keep pace with it.
NVIDIA argues that storage infrastructure now needs to be designed alongside GPUs, networking and AI software rather than treated as a separate layer added later.
NVIDIA Opens cuFile APIs to Developers
One of the biggest announcements is NVIDIA’s decision to open source its cuFile APIs and the storage software stack behind them.
cuFile forms part of NVIDIA GPUDirect Storage. The technology allows GPUs to read from and write to storage directly instead of routing every operation through a traditional CPU-controlled path.
That shorter route matters when an AI system is handling huge datasets.
NVIDIA says cuFile can use large numbers of GPU threads and high-bandwidth memory to access stored data within microseconds. The APIs are now available for outside contributions, with Google, Intel, Meta and NVIDIA named as the inaugural maintainers of the project.
Opening the technology could also make it easier for storage providers to support different hardware and software environments instead of building around a closed NVIDIA-only implementation.
There is a security angle here too. Faster access is not particularly helpful when it opens an unsafe route into protected memory. NVIDIA says the software stack follows established Linux security practices and keeps access controls in place while allowing GPUs to communicate more directly with approved storage.
Storage-Next Brings More Than 40 Vendors Together
NVIDIA is also expanding an industry programme called Storage-Next.
The initiative includes more than 40 storage and flash companies, including DDN, KIOXIA and Micron. Controller manufacturers, cooling specialists, orchestration providers and standards organisations are involved as well.
The group is working on common approaches for storage systems built around GPU-heavy AI infrastructure.
That sounds technical because it is. The practical issue is interoperability.
AI data centres often combine GPUs, networking equipment, storage arrays, security software and management tools from several vendors. If every component handles direct data access differently, the infrastructure becomes harder to deploy and maintain.
Storage-Next is intended to align those systems before incompatible approaches become deeply embedded across the market.
SCADA Lets GPUs Pull Only the Data They Need
Another part of NVIDIA’s storage strategy is SCADA, short for scaled, accelerated data access.
The framework allows massively parallel GPUs to retrieve selected data directly from storage and move it into high-speed memory. Instead of loading an entire dataset, an application can request only the portion needed for a specific task.
That could prove useful for retrieval systems, AI agents and large-scale inference workloads where models repeatedly search through extensive collections of enterprise data.
DDN is integrating SCADA with Infinia, its software-defined data intelligence platform. The aim is to reduce storage delays and keep expensive GPU infrastructure working instead of waiting for data to arrive.
Idle GPU time is not a small concern. Companies are spending heavily on accelerated computing, and any bottleneck that leaves those processors underused can weaken the economics of an AI deployment.
NVIDIA Vera Targets Storage Processing Work
NVIDIA also highlighted performance results for the Vera CPU used in its Vera BlueField-4 STX storage processor.
According to NVIDIA’s own benchmarks, Vera delivered up to 3.21 times the throughput of the x86 processor used in a two-stage compression and encryption test. The company also reported gains across encryption, data recovery, integrity checking, compression and decompression workloads.
The results are worth reading carefully.
These were controlled processor benchmarks rather than full production storage-system tests. NVIDIA said the measurements excluded areas such as file input and output, disk performance and networking unless otherwise specified. End-to-end testing would still be needed to measure the effect on a complete AI platform.
Still, the numbers show where NVIDIA sees the next pressure point.
AI storage processors are being asked to do far more than move files. They must secure, validate, compress and recover data without slowing the GPUs waiting on the other side.
Security Remains Part of the Direct-Access Design
Giving applications a faster route to storage can create its own risks.
An improperly designed system could allow one process to write into memory or storage assigned to another workload. That turns a performance feature into a serious security problem.
NVIDIA says SCADA separates high-speed application functions from a privileged component responsible for configuring protected access. The privileged layer connects an application only to the storage resources it has been approved to use.
The approach keeps ordinary application processes outside the trusted computing base while relying on standard Linux security controls for enforcement.
NVIDIA is also positioning fast storage as part of AI-powered cybersecurity. Security agents need rapid access to logs, threat intelligence and system context. A slow storage layer could delay the same automated detection and response tools companies are adopting to protect AI infrastructure.
AI Storage Is Becoming Part of the Compute Stack
For years, the dividing line was simple. Frequently used information stayed in expensive memory. Everything else went to slower but cheaper storage. That line is beginning to blur.
Direct GPU access and faster storage processing will not replace high-bandwidth memory. They may, however, allow AI applications to reach much larger pools of information without treating every dataset as if it must remain inside scarce system memory.
This matters most for agentic AI. Agents need persistent context, enterprise records, cached model data and a steady flow of new information. The model may run on the GPU, but the quality and speed of its work increasingly depend on what happens outside the accelerator.
NVIDIA’s message from FMS is that the storage layer can no longer be left behind. The GPU remains the star of the AI factory. Storage is becoming part of the machinery that keeps it running.
