Table of Contents
More than 40 vendors join Storage-Next, but Nvidia owns the stack
In sum – what we know:
- A GPU-first storage tier – Storage-Next targets 512-byte IOPS per GPU under tight power and tail-latency limits, matching how inference workloads actually read data.
- Open-sourcing the interface – Nvidia is publishing cuFile and its supporting stack under XIO-SIG, a new GitHub organization maintained by Google, Intel, Meta, and Nvidia.
- Questions about neutrality – Nvidia controls the core APIs, branding, reference architectures, and certification programs, leaving open whether the standard is truly vendor-neutral.
Nvidia has formally launched its Storage-Next initiative at the Future of Memory and Storage event, an effort to establish open, interoperable standards for GPU-driven storage in large-scale AI systems. The launch put a name and a membership figure on work that had been simmering since GTC 2025, when Nvidia first outlined plans to reorganize storage around GPU-initiated I/O.
The initiative convenes more than 40 vendors spanning storage, flash media, controllers, cooling, orchestration, and standards bodies. That breadth is deliberate. Storage-Next isn’t limited to defining a faster protocol — it’s an attempt to redefine how storage devices behave when the GPU, rather than the CPU, is the primary client, and to codify those behaviors as vendor-neutral specifications rather than proprietary implementations.
Whether “vendor-neutral” holds up in practice is the open question. Nvidia owns the core APIs, the branding, and the reference architectures underneath the initiative, and the industry has seen dominant vendors launch “open” standards before.
What is Storage-Next?
Storage-Next is described by Nvidia and its participants as a collaborative standards effort focused on accelerated data access for large AI datasets, specifically when GPUs initiate and manage storage I/O directly. The pitch is interoperability. Advanced techniques for GPU-centric storage exist today, but they tend to live inside single-vendor stacks. Storage-Next aims to make them portable across the ecosystem, with participants co-defining hardware behaviors, I/O semantics, and performance targets exposed through open APIs.
The technical goal is maximizing 512-byte IOPS per GPU under strict power and tail-latency constraints. That number matters because it reflects what modern AI workloads actually do — retrieving tokens, embeddings, and context slices in small, random bursts rather than streaming large sequential files. Legacy enterprise storage was designed and benchmarked around large-block throughput, which is largely irrelevant to an inference cluster pulling context fragments on the fly. Storage-Next retools drives, controllers, and systems around those smaller I/O units instead.
That’s a genuine departure from how general-purpose enterprise storage has been engineered for decades. It’s also a narrow one — a drive tuned for 512-byte GPU-initiated reads may not be the drive you want under a transactional database. Nvidia is betting that AI workloads are now large enough to justify their own storage category.
Ecosystem and participating vendors
The 40-plus participant roster includes named storage vendors like DDN, KIOXIA, and Micron, each contributing next-generation technologies aligned with Nvidia’s GPU-centric data path. Notably, the group also pulls in thermal design and cooling standards organizations — a signal that tightly coupling GPUs and aggressively accessed storage changes power and cooling profiles at the rack level, not just at the drive level.
On the software side, Nvidia has stood up a new organization called XIO-SIG on GitHub to host the open-source implementation of its GPU-to-storage APIs. The founding maintainers are Google, Intel, Meta, and Nvidia. That lineup is worth pausing on. Hyperscaler involvement gives the effort credibility beyond Nvidia’s own priorities, and Intel’s presence is interesting given that the whole premise of the initiative is demoting the CPU from its traditional role in the storage path.
Storage-Next is also explicitly tied to Nvidia’s broader AI Storage Ecosystem, which bundles hardware reference designs, software stacks, and certification programs for aligned vendors. That’s a double-edged sword. Integration with certification programs gives vendors a clear path to market, but it also means the “standard” and Nvidia’s commercial ecosystem are hard to separate. Vendors targeting alternative accelerators may reasonably worry that tuning hardware primarily for GPU-initiated I/O privileges Nvidia workloads over everything else.
Underlying tech
Storage-Next sits on top of technical pieces Nvidia has been rolling out across several product cycles. The most fundamental is SCADA — Scaled, Accelerated Data Access — which applies NVSHMEM-style semantics to storage. In practice, that means moving storage control logic onto the GPU itself, so I/O operations align with GPU scheduling rather than waiting on a CPU intermediary. SCADA is the framework Storage-Next will build against, so any hardware standard emerging from the group is expected to expose behaviors compatible with its GPU-centric model.
The second piece is cuFile, Nvidia’s API for file-based data access directly from GPUs, bypassing CPU mediation entirely. At FMS 2026, Nvidia announced that cuFile and its supporting software stack are being open-sourced under XIO-SIG. That’s arguably the most significant part of the launch. Publishing the interface and open-sourcing the implementation is a real step toward shared innovation — though it’s worth noting that controlling the reference implementation gives Nvidia considerable influence over how the standard evolves, even under nominally shared governance.
Then there’s STX, a modular storage reference architecture built around BlueField-4 DPUs and Spectrum-X Ethernet, introduced at GTC 2026. STX gives vendors a concrete blueprint — rack-scale configurations, networking topologies, controller behaviors — while Storage-Next generalizes those patterns into standards. Partners are expected to ship BlueField-4 STX systems in the second half of 2026, so the concepts won’t stay theoretical for long.
Skeptics have pointed out that all of this intersects with existing standards work around NVMe, CXL, PCIe, and various fabrics, raising fair questions about how much is genuinely new versus Nvidia-specific layering on top of protocols the industry already maintains. There’s substance to that critique. But the fact is nobody else has published a coherent, end-to-end model for GPU-initiated storage, and the existing standards bodies weren’t moving at the pace AI infrastructure demands.
Data center design
For enterprise IT teams, the practical takeaway is that AI workloads will likely require their own storage tier, separate from the arrays serving traditional databases and file workloads. The two access patterns are different enough that one system serving both well is increasingly unrealistic.
Storage-Next and STX should also accelerate the adoption of BlueField-class DPUs as storage offload engines, handling RAID, compression, encryption, and other data services that would otherwise consume CPU cycles. That shift has been coming for a while — DPUs have been marketed as offload engines since the original BlueField parts — but a standards framework and shipping reference designs give it real momentum. The broader push is toward rack-scale designs that treat GPUs, DPUs, and storage as an integrated AI fabric rather than separate silos, which changes how teams plan capacity, lifecycle management, and procurement.
Physical assumptions change too. Co-locating high-power GPUs with storage devices under constant small-block hammering alters airflow and power-distribution math at the rack level — hence the cooling organizations on the member list.
The hardest question lands on organizations with heavy investments in traditional storage. They’ll face a choice between retrofitting existing infrastructure to support GPU-initiated I/O or deploying entirely new AI-native clusters alongside it, and neither path is cheap. Expect “Storage-Next-aligned” and “GPU-ready” labels to start appearing in RFPs regardless. Whether those labels come to mean genuine interoperability or simply a certification badge for Nvidia-aligned systems will depend on how the governance actually plays out — and on whether anyone builds a credible implementation that doesn’t run through Nvidia silicon.
