Nvidia is open-sourcing its cuFile APIs and rallying more than 40 storage and flash vendors behind a new initiative called Storage-Next, an attempt to reshape how AI workloads pull data off drives rather than out of memory. The announcements, made at the Future of Memory and Storage conference in Santa Clara on August 4, position storage as an active part of the AI data path rather than a passive tier. Nvidia's own benchmarks claim the Vera CPU inside the Vera BlueField-4 STX delivers up to 3.21x higher throughput than an x86 CPU in a two-stage compression and encryption pipeline.
The pitch is that AI agents are now generating thousands of concurrent storage operations directly from GPUs, and traditional CPU-mediated data paths cannot keep up. Storage systems have to encrypt, compress, verify, and reconstruct data continuously, and those services choke when thousands of agents hit them at once. Nvidia's argument is that pushing those tasks onto BlueField data processors frees the storage layer to absorb the flood with less compute infrastructure behind it.
The economics Nvidia is invoking are 40 years old. When engineers first drew the line between memory and disk, the access-time gap between the two was measured in minutes, and the tradeoff dictated what workloads could run at all. On modern GPUs paired with direct-storage software, that same tradeoff now plays out in microseconds, which is fast enough that storage can behave like a memory extension for large context windows and multi-turn agents.
Key facts
- 01Nvidia's Vera CPU, part of Vera BlueField-4 STX, delivers up to 3.21x higher throughput than x86 in a two-stage compression and encryption pipeline.
- 02Nvidia is open-sourcing its cuFile APIs and the vertical storage software stack, letting GPUs read and write to storage directly in microseconds.
- 03The new Storage-Next initiative pulls in over 40 flash and storage vendors, including DDN, KIOXIA, and Micron.
- 04The Open Secure AI Alliance launches with Google, Intel, Meta, and Nvidia as inaugural maintainers for shared storage-security APIs.
- 05FMS runs August 4-6 in Santa Clara; Nvidia GTC Berlin follows October 20-22.
cuFile is the software mechanism that makes that possible. It is an open-source component of Nvidia GPUDirect Storage, and it lets GPUs — not just CPUs — read from and write to storage directly using hundreds of thousands of GPU threads and high-bandwidth memory. By open-sourcing the APIs and the vertical storage stack underneath them, Nvidia is trying to make GPU-to-storage access a Linux-native standard rather than a proprietary shortcut.
The move dovetails with the Open Secure AI Alliance, which Nvidia disclosed alongside Google, Intel, and Meta as inaugural maintainers. AI Chat Daily covered the alliance's SAFE guidelines for agentic AI incident sharing earlier this month; the cuFile release now gives that group a concrete API surface to optimize across hardware and software platforms. The stated goal is interoperability, but the effect is that Nvidia's storage assumptions become the reference implementation the rest of the industry codes against.
Storage-Next is the hardware-side complement. The initiative brings together storage makers, controller vendors, thermal and cooling designers, and standards bodies to align on how GPU-driven storage should behave, then push those alignments into open industry standards. DDN, KIOXIA, and Micron are among the more than 40 vendors participating, and DDN is integrating Nvidia's SCADA framework — scaled, accelerated data access — into its Infinia data intelligence platform.
“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.”— Sven Oehme, Chief Technology Officer, DDN
SCADA is Nvidia's answer to the security problem that direct-storage access creates. Letting an application talk straight to a drive is fast, but done carelessly it can overwrite memory belonging to other processes. Nvidia's design splits the job in two: the speed-critical user portions stay outside the trusted computing base, while a separate privileged component configures protected access at setup and enforces standard Linux security protocols. It is a design choice that matters because it determines whether direct-storage access can be sold into regulated enterprises.
The broader hardware context is Nvidia's Vera Rubin platform, which pairs Vera CPUs with BlueField-4 storage processors and Spectrum-X Ethernet networking into a modular rack-scale system. Layered on top is Nvidia CMX Context Memory Storage, a context tier aimed specifically at long-context, multi-turn, agentic inference. The message to buyers is that Nvidia sells not just the GPU but the storage architecture that keeps the GPU fed, with the DOCA security stack enforcing policy continuously along the way.
The skeptic's read on Storage-Next is that it is a standards body Nvidia convened to standardize on Nvidia's assumptions. The 40-vendor roster gives it credibility, but interoperability written to a single reference implementation tends to lock the ecosystem to that reference. Vendors outside the initiative — or with competing accelerator roadmaps — will need to decide whether to adopt cuFile as written or fork it. The open-source license lowers that friction but does not eliminate it.
The commercial logic behind this push is straightforward. Every hyperscaler and neocloud building out AI capacity is running into the same wall: memory is expensive, context windows keep growing, and GPUs are idle whenever storage cannot keep pace. If Nvidia can make its storage stack the default plumbing for those buildouts, it captures margin far beyond the accelerator itself and makes the switching cost of moving off Nvidia infrastructure meaningfully higher. Storage-Next is less a research project than a moat, priced in microseconds.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
