NVIDIA Open-Sources cuFile API and Launches Storage-Next Initiative With 40+ Industry Partners

As AI Increases Demands on Memory, Storage Steps Up

At FMS, NVIDIA shows how accelerated computing enables AI applications to access storage directly -- fast enough to act like memory and secure by design. Surging AI demands are driving the need for massive datasets and context windows that burst past the confines of system memory. But rising needs aren't met by simply adding more storage capacity. What's needed is useful, grounded insights from AI factories and efficient, secure storage architectures that enable those insights. At this week's Future of Memory and Storage (FMS) conference, NVIDIA is unveiling new storage advancements and showcasing how the next leap in AI depends as much on the storage infrastructure feeding accelerated computing as on the computing power itself. The pressure on that infrastructure is intensifying as AI agents consume massive amounts of data -- and GPUs can now initiate storage requests directly, generating thousands of concurrent operations. To serve those requests, storage systems must continuously encrypt, compress, verify and reconstruct data. These critical data services can become bottlenecks when thousands of agents access storage simultaneously. Benchmarks highlighted in this NVIDIA technical blog show that the NVIDIA Vera CPU, part of NVIDIA Vera BlueField-4 STX, delivers up to 3.21x higher throughput than an x86 CPU in a two-stage compression and encryption pipeline. This means that with Vera, storage platforms can absorb the flood of AI data more efficiently -- delivering greater throughput with significantly less compute infrastructure. With accelerated computing, storage stops being a passive place to keep data and becomes an active part of the data path. This upends the old economics of determining when data belongs in memory (where applications can fetch it faster) versus on a storage drive (where it can be held in cheap and plentiful space). The tradeoff was first framed 40 years ago, when the answer was measured in accessing that data in minutes. On today's GPUs, paired with AI storage solutions from NVIDIA and partners, the same tradeoff now plays out in microseconds. Closing the gap between AI's needs and memory shortage depends on extreme codesign across the whole ecosystem, from memory and storage manufacturers to the software built on them. Open Source NVIDIA cuFile APIs Enable Interoperability for Storage Solutions At FMS, NVIDIA announced it is open sourcing its cuFile application programming interfaces (APIs) -- and the vertical storage software stack underneath them -- which let GPUs, not just CPUs, read from and write to storage directly. cuFile is an open source component of NVIDIA GPUDirect Storage. Using hundreds of thousands of GPU threads, fast high-bandwidth memory and other methodologies, cuFile enables securely accessing data from storage in just microseconds. This represents how the industry is unifying a security-first storage stack based on Linux best practices, providing interoperability between GPUs and data. In addition, fast, secure access to data and storage is a foundational element to powering preventive and detective cybersecurity measures. Making cuFile openly available will help make security context, data and storage accessible at the speed AI-powered defenses need. Such open technologies support initiatives such as the new Open Secure AI Alliance. This site is the new home for APIs that are open to contributions -- with Google, Intel, NVIDIA and Meta as inaugural maintainers -- and can be optimized for use across various software and hardware platforms, driving innovation and efficiency for developers and enterprises. NVIDIA and Industry Leaders Advance New Frontier of AI Storage In addition, NVIDIA and storage industry leaders are optimizing memory and storage solutions through an initiative called Storage-Next. The NVIDIA-driven initiative brings together storage makers, controller vendors, thermal design, cooling and orchestration operators, and standards bodies to align on how GPU-driven storage should behave -- then turn these advancements into interoperable, open industry standards. Storage-Next includes over 40 leading storage and flash vendors -- including DDN, KIOXIA and Micron -- each contributing to the next generation of AI storage technologies with NVIDIA. The initiative is grounded in accelerated data access for large AI datasets. To support this, NVIDIA offers SCADA -- short for scaled, accelerated data access -- a framework that lets massively parallel GPUs pull only the data necessary for the application directly from storage into their own high-speed memory. For example, DDN is integrating SCADA with Infinia, its software-defined, AI-native data intelligence platform built to eliminate storage bottlenecks at scale. "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 at 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." Storage-Next and SCADA extend NVIDIA's longstanding work on AI storage infrastructure, including on NVIDIA Vera BlueField-4 STX -- a modular, rack-scale foundation powered by the NVIDIA Vera Rubin platform, NVIDIA Vera BlueField-4 storage processors and NVIDIA Spectrum-X Ethernet networking. Defining a new class of AI-native data platforms, NVIDIA STX uses the unified NVIDIA DOCA security stack to let enterprises enable continuous policy enforcement in the AI data path. Plus, NVIDIA CMX Context Memory Storage provides an AI‑native context tier for long‑context, multi‑turn, agentic AI inference, built on NVIDIA STX. SCADA Enables Fast AI Storage That Stays Secure Speed at the storage layer comes with a catch. Letting an application talk straight to a drive is quick, but done carelessly, it can scribble over other processes' memory -- a security hole, not a feature. NVIDIA SCADA uses a safe, robust method to achieve scaled direct access by splitting the job in two: * The user parts of an application that need raw speed stay outside the trusted computing base. * A separate, privileged component configures protected access between the user application and its approved storage at setup, adhering to standard Linux protocols for security enforcement while efficiently safeguarding data. It's all part of how advancements in fast, massively parallel, efficient, secure AI storage infrastructure can feed better data to applications and AI factories -- so they can produce more useful, accurate, grounded intelligence at scale. Join NVIDIA sessions at FMS, running Aug. 4-6 in Santa Clara, California, and learn more about NVIDIA AI storage. See notice regarding software product information.

'AI success will be defined not by how much infrastructure organizations own, but by how productively they use it': Nvidia lays out its thoughts on how storage has become the next frontier of AI

* Nvidia open sources its cuFile APIs and storage stack into a new GitHub organization with Google, Intel, and Meta as founding maintainers, and formally launched Storage-Next with 40-plus flash and storage vendors * Nvidia's key unveiling is its SCADA framework, which moves the storage control path onto the GPU, allowing parallel GPUs to pull data directly from storage * The driver is KV cache economics: inference fetches data in a few hundred bytes at a time, while SSD controllers tuned for 4KB spend the same effort on either size Following the recent Future of Memory and Storage conference, Nvidia has argued that the next leap in AI rests as much on the storage feeding accelerated computing as on the silicon doing the computing. The company open sourced its cuFile APIs and the storage stack beneath them, and formally launched an industry initiative called Storage-Next with more than 40 flash and storage vendors. It also put a name to SCADA, the framework that lets GPUs pull data from drives without the CPU brokering every request. A 512-byte problem that comes into focus as storage becomes key Nvidia's approach here is not new. cuFile was introduced in 2019 as the interface component of GPUDirect Storage and has shipped since 2021; its role is to take the CPU out of the data path. Bytes move by Direct Memory Access (DMA) straight between the drive and GPU memory, with no bounce buffer in host RAM. What stayed on the CPU was the control path: host software still decided what to fetch and issued every request, with the GPU as the DMA target rather than the initiator. That split is invisible at large transfer sizes. A one-megabyte read essentially amortizes the per-request cost. At 512 bytes, the ratio inverts, the fixed cost dominates, and the CPU saturates long before the drives do. SCADA is the piece that moves the control path onto the GPU, letting it construct and complete its own storage requests and absorb per-operation latency, just as it already absorbs memory latency by keeping hundreds of thousands of operations in flight. The two are complementary rather than successive: cuFile for bulk transfers, SCADA for high volumes of small random reads. The problem it solves is one that enterprise SSDs exhibit, having been tuned around 4KB random reads for nearly a decade to match virtualization and databases. As a result, most controllers do the same amount of work to serve 512 bytes of data as they would a 4KB read request. AI inference access patterns are considerably smaller than that: embeddings run a few hundred bytes, and the KV cache blocks sit well under a kilobyte. Serving them from 4K-tuned drives imposes something like eightfold read amplification, and at tens of terabytes of small objects, that amplification decides whether flash works as a memory tier at all. The KV cache, essentially the attention state for every token already processed, grows with context length, and agentic deployments run thousands of concurrent conversations. It quickly outgrows GPU memory, and recomputing evicted entries costs more GPU time than reading them back, so the standard design spills from GPU memory to system memory and flash, and refills through high volumes of small random reads. Serving that from flash rather than DRAM increases the context length and the concurrent user count each GPU can support, which in turn affects the per-user cost of serving a model. This also explains why Nvidia is focusing on 512-byte IOPS rather than raw bandwidth, since inference performance depends heavily on the former. Opening up cuFile is a departure; this layer has historically lived inside CUDA, and the logic is not charity. A GPU-initiated storage interface only pays off if drive, controller and array vendors build to it, and vendors do not build to a proprietary interface owned by the company whose GPUs they are feeding. Publishing the interface, open-sourcing the implementation, and convening 40-plus vendors to standardize the underlying hardware behavior is how Nvidia aims to make GPU-initiated storage the industry default. It also stands to benefit the most from that outcome, since it sells most of the GPUs in question. StorageReview notes Intel's participation as a significant development; a leading supplier of the x86 silicon currently sitting in storage controllers has signed on to maintain software designed to remove that silicon from the I/O path. Google and Meta co-maintaining a layer that standardizes how accelerators reach storage, while building their own accelerators, points in the same direction. Nvidia's take is a coherent, well-argued push at a real bottleneck, with unusually credible partners attached. Partner systems from DDN, Dell, HPE, IBM, VAST Data and WEKA are due in the second half of 2026. Kioxia's XL-Flash drives built for 512-byte access are in development under Storage-Next. Nvidia's roadmap calls for Gen7 SSDs sustaining 100 million IOPS each, which is a target controller vendors are designing toward rather than a product anyone can buy. Storage-Next itself has been discussed publicly since GTC 2025; this week, it acquired both a membership number and a framework to build against. Follow TechRadar on Google News and add us as a preferred source to get our expert news, reviews, and opinion in your feeds.

Nvidia open sources cuFile API, accelerating GPU read/write capability for high-speed storage

As artificial intelligence applications become ever hungrier for faster access to data, Nvidia Corp. today announced it is open-sourcing the application programming interface for its powerful cuFile vertical data storage stack, enabling millisecond data access. The company also announced a large-scale industry initiative with technology leaders in optimizing memory and storage solutions with Storage-Next. The initiative aims to bring storage makers, controller vendors, thermal design, cooling, and orchestration providers together with standard bodies to find opportunities for superior graphics processing unit-driven storage. During Nvidia's announcement at the Future of Memory and Storage conference, the company said the cuFile API enables secure access from storage in just milliseconds. CuFile was first officially launched into general availability in July 2021 alongside the company's GPU access CUDA Toolkit 11.4. Prior to the launch, it acted as a core component of the company's GPUDirect Storage software. It provides direct, high-speed data access between local distributed storage and GPU memory; this allows GPUs to bypass the central processing unit and system main memory, which greatly reduces memory access delays. Using direct memory access, or DMA, it can move data directly from storage devices such as NVMe drives into GPU memory. The primary source of AI inference happens on GPUs and specialized cards that interact primarily with GPU memory; refreshing that memory is often completed by another software layer that uses the GPU to orchestrate and make decisions. CuFile cuts out the "middle man," so to speak, providing a direct access path and enabling GPU orchestration, meaning it can rapidly pull more data. This allows the synchronization of massive data sets at rates that can keep up with elite GPU speeds. If the bandwidth to this data is delayed or pinched off by a bottleneck, it causes something called the "GPU starvation" loop, where some GPUs in a distributed array sit idle waiting for data to be ready. Increased speeds and low millisecond delays allow massive datasets from retrieval-augmented generation and agentic AI to run faster by providing them a low-latency "superhighway" between deep storage and the GPU itself. Bringing open standards to GPU accessible storage Storage-Next includes 40 leading storage and flash memory vendors, including DataDirect Networks, Inc., Kioxia Corp. and Micron Technology Inc. Each will contribute to the next generation of AI storage technologies with Nvidia. The company will build its grounded, high-speed data access for storage off SCADA, short for scaled, accelerated data access. This solution provides the groundwork for massively parallel GPUs to pull data at high bandwidth and extremely low delay. Modern AI training requires massive data access all at once and AI inference now includes colossal mixture-of-experts models, which "think" extremely quickly, rapidly pulling data and firing tools for agentic workflows. "AI success will be defined not by how much infrastructure organizations own, but by how productively they use it," said DDN Chief Technology Officer Sven Oehme. "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." Nvidia said SCADA allows high-speed storage layers to operate securely. Technologies such as SCADA allow direct access to storage media by the AI; this punches through safeguards designed to prevent clobber - what happens when two processes attempt to write to the same data at once, with one scribbling over the other's work. This is a security problem, not a proper feature. The other side of the coin means that this might bypass security safeguards such as encrypted and privileged access. This opens up another security hole that malicious parties could exploit. SCADA overcomes this challenge by scaling direct access by splitting access across different jobs: the user parts of the application receive raw speed but stay outside the encrypted, secure and trusted computing base; and a separate privileged computing component configures protected access following Linux protocols to prevent clobber.

NVIDIA Opens cuFile and Builds an SSD-to-GPU Storage Ecosystem

NVIDIA has opened the source code for the cuFile API and associated GPU-direct storage components. The change gives storage vendors and software developers greater access to the technology used by GPUDirect Storage to move information between NVMe storage and GPU memory without routing every transfer through a conventional CPU-managed system-memory buffer. The announcement does not literally convert an SSD into graphics memory. HBM and GDDR remain considerably faster and provide much lower latency. Instead, cuFile creates a more efficient path for loading model weights, datasets, textures and other information that cannot remain permanently resident in GPU memory. NVIDIA is also organizing a Storage-Next initiative involving approximately forty storage and flash companies, including Kioxia, Micron and DataDirect Networks. The objective is to develop lower-latency storage capable of supporting retrieval-augmented generation, agentic AI, large inference datasets and other GPU-heavy workloads. Consumer-game benefits remain speculative. Direct storage access could eventually help engines stream assets with less CPU overhead, but software integration, drive latency and PCIe bandwidth will determine actual results. Existing technologies such as Microsoft DirectStorage already pursue related objectives on Windows. The development is more immediately relevant to AI servers, where an expensive accelerator can remain underused while waiting for data. Even a small reduction in those idle periods can improve the economics of large GPU clusters. NVIDIA storage developmentDetails Opened componentcuFile API and supporting storage stack Data pathNVMe storage to GPU memory through DMA Primary technologyGPUDirect Storage Storage-Next participantsApproximately 40 companies Initial workloadsAI inference, RAG, agents and large datasets Replacement for HBM/GDDRNo; operates as a slower capacity tier