NVIDIA Storage-Next: Turning Storage into Active AI Infrastructure
NVIDIA is pushing storage beyond its traditional role as a passive data repository and toward becoming an active component of the AI compute pipeline.
At the Future of Memory and Storage (FMS) conference, NVIDIA demonstrated an accelerated storage architecture built around its Vera CPU and Vera BlueField-4 STX platform. In two-stage compression and encryption workloads, the architecture reportedly delivers up to 3.21× the throughput of x86 CPUs.
The underlying strategy is straightforward: as AI workloads become increasingly data-intensive, moving computation closer to storage can eliminate unnecessary CPU bottlenecks and reduce data movement across the system.
NVIDIA is extending this approach through the open-sourcing of its cuFile API and vertical storage software stack, while simultaneously building an industry ecosystem around its Storage-Next initiative and SCADA (Scalable Accelerated Data Access) framework.
⚡ Accelerated Computing Moves Onto the Storage Data Path #
Traditional storage architectures treat storage primarily as a destination for data. The CPU retrieves that data, performs operations such as compression or encryption, and then passes the processed results to accelerators or applications.
That model becomes increasingly inefficient as AI systems process larger datasets and serve thousands of concurrent workloads.
NVIDIA’s Vera BlueField-4 STX architecture takes a different approach by placing accelerated processing directly into the storage data path.
In the demonstrated two-stage compression and encryption workflows, NVIDIA reports throughput of up to 3.21× that of x86 CPUs.
Vera CPU and BlueField-4 STX #
The Vera CPU provides the general-purpose processing component within NVIDIA’s accelerated infrastructure strategy, while BlueField-4 STX extends data-processing capabilities closer to storage.
The important architectural shift is not simply replacing one CPU with another. It is moving storage-related computation toward specialized accelerated infrastructure so that general-purpose host CPUs do not become the default execution layer for every data-management operation.
For AI clusters, this can matter when storage operations are executed concurrently with inference, training, retrieval, and other compute-intensive workloads.
Operations such as compression and encryption consume compute cycles even though they are not themselves the primary AI workload. Offloading them can therefore preserve host and accelerator resources for higher-value computation.
🔓 NVIDIA Open-Sources cuFile and Its Storage Software Stack #
NVIDIA also announced plans to open-source the cuFile API and its vertical storage software stack.
cuFile is designed around direct GPU-to-storage data movement, allowing applications to bypass unnecessary CPU-mediated copies and move data between storage and GPU memory more efficiently.
This architecture becomes particularly relevant when GPUs have enormous parallel-processing capacity and high-bandwidth memory, while traditional storage paths remain constrained by CPU processing and memory-copy overhead.
Direct GPU Access to Storage #
With GPU-direct data paths, storage data can be delivered to GPU memory without requiring the CPU to act as the intermediary for every transfer.
NVIDIA’s architecture leverages the massive parallelism of modern GPUs, including hundreds of thousands of GPU threads, together with high-bandwidth memory.
The goal is to reduce latency and improve data availability for workloads that repeatedly stream large datasets between storage and accelerators.
NVIDIA describes the approach as enabling secure data access within microsecond-scale timeframes, depending on the workload and system configuration.
Why Open Source Matters #
Opening the cuFile API and associated software stack could have implications beyond NVIDIA’s own hardware.
A more open storage software layer can encourage storage vendors, controller developers, infrastructure providers, and Linux ecosystem participants to implement compatible accelerated data paths.
This is particularly important for AI infrastructure, where storage hardware is increasingly heterogeneous and organizations expect software interfaces to remain portable across different storage technologies and deployment environments.
🤖 Agentic AI Is Exposing Traditional Storage Bottlenecks #
The shift toward agentic AI is increasing pressure on storage infrastructure.
Large language models increasingly operate with massive datasets, long context windows, retrieval pipelines, persistent state, and tool-driven workflows. Instead of a small number of predictable jobs, infrastructure may need to serve thousands of agents simultaneously.
That changes the role of storage.
A storage system must not only provide capacity and bandwidth. It may also need to continuously perform data-processing operations such as:
- Compression
- Encryption
- Checksum verification
- Data reconstruction
- Data movement
- Retrieval preparation
- Data integrity processing
If these operations remain concentrated on general-purpose CPUs, they can become bottlenecks when large numbers of AI workloads access the same storage infrastructure concurrently.
Capacity Alone Is No Longer the Full Solution #
Adding more storage capacity addresses one constraint but does not automatically solve the data-movement problem.
The relevant metric for AI infrastructure is increasingly the ability to move, transform, and serve data at the rate demanded by accelerators.
This creates an imbalance when GPU compute capacity scales faster than the surrounding storage pipeline.
A GPU may be capable of processing enormous volumes of data, but its utilization ultimately depends on how quickly the system can deliver usable data to it.
NVIDIA’s Storage-Next strategy is designed around closing that gap.
🗄️ Storage-Next Targets an Accelerated Data Infrastructure #
NVIDIA is leading the Storage-Next initiative to address the growing disconnect between AI compute requirements and storage infrastructure.
The initiative brings together more than 40 companies and organizations spanning storage systems, flash memory, controllers, thermal and cooling infrastructure, and industry standards.
Participants cited by NVIDIA include DDN, KIOXIA, and Micron.
Rather than treating accelerated storage as an isolated NVIDIA feature, Storage-Next is intended to establish a broader ecosystem around interoperable accelerated data access.
SCADA: Scalable Accelerated Data Access #
At the center of the initiative is SCADA, or Scalable Accelerated Data Access.
The framework is intended to formalize GPU-driven storage behaviors through open standards and interoperable interfaces.
That distinction matters because AI storage bottlenecks are not limited to a single component.
The end-to-end data path can include:
- Storage media
- Storage controllers
- Network or fabric connectivity
- CPU and system memory
- GPU memory
- Compression and encryption engines
- Application-level data pipelines
Optimizing only one component can leave another part of the pipeline as the limiting factor. SCADA’s broader objective is to define how accelerated data access should work across this ecosystem.
🔗 From GPU Direct Storage to an Industry-Wide Architecture #
NVIDIA’s announcements at FMS indicate a broader evolution in accelerated computing.
The company’s earlier focus was primarily on accelerating compute. Increasingly, the strategy extends that acceleration to the movement and transformation of data surrounding the compute engine.
That means storage infrastructure is becoming part of the computational architecture rather than an isolated subsystem.
The model can be viewed as a shift from:
Storage → CPU → Memory → GPU
toward a more heterogeneous architecture in which:
Storage → Accelerated data path → GPU memory
can become a first-class execution path for selected workloads.
The practical advantage is reducing unnecessary data copies and minimizing the amount of CPU work required to prepare data for accelerators.
🧩 Why This Matters for Future AI Infrastructure #
AI scaling is increasingly constrained by the entire system rather than by accelerator performance alone.
Faster GPUs can increase compute throughput, but if storage cannot provide data quickly enough—or if CPUs spend too much time processing data before it reaches the GPU—the accelerator can remain underutilized.
NVIDIA’s FMS announcements attack that problem from several directions:
- Vera CPU and BlueField-4 STX target accelerated infrastructure-side processing.
- cuFile provides a software interface for more direct GPU-to-storage access.
- Open-sourcing the storage stack can encourage broader ecosystem adoption.
- Storage-Next coordinates hardware and infrastructure vendors.
- SCADA aims to establish interoperable standards for accelerated data access.
Together, these initiatives represent an attempt to make storage a more active participant in AI execution.
🚀 NVIDIA’s Storage Strategy Points Beyond Faster SSDs #
The significance of Storage-Next is not simply about building faster storage devices.
The larger architectural trend is toward computational storage infrastructure, where data can be processed, transformed, secured, and delivered closer to the accelerator that ultimately consumes it.
For conventional workloads, CPU-mediated storage operations may remain sufficient. For large-scale AI and agentic systems, however, the sheer volume of concurrent data movement changes the economics.
When thousands of AI agents compete for datasets, model context, retrieval results, and persistent state, every unnecessary copy and CPU-side processing stage adds pressure to the system.
NVIDIA’s approach is to make the entire data path more accelerator-aware.
If Storage-Next and SCADA gain broad industry adoption, storage could increasingly evolve from a passive capacity layer into an active component of AI infrastructure—one that participates directly in data processing, security, movement, and delivery.
That would represent a meaningful architectural shift: the future AI bottleneck may no longer be determined solely by how fast the GPU computes, but by how efficiently the entire system can feed it.