The new AI bottleneck isn't GPUs. It's how you feed them

The GPU used to be the easy part of an AI buildout. The hard part is what comes after: the storage stack that has to feed it.

Nvidia made the case at FMS 2026 in Santa Clara this week, and the technical shape of it is specific. Modern AI inference (the work of running a trained model to answer a question or generate a token) lives on data called the KV cache, which stores the model's intermediate memory of what it has already computed. The model pulls small slices of that cache constantly: hundreds of bytes at a time, millions of times per request, scattered across drives. Solid-state drives were designed for the opposite workload. They are tuned for 4KB transfers, the same way a freight railroad is tuned for container cars. Per-request overhead gets paid on the CPU, the part of the system that already has plenty to do. Storage did not become slow; it became the wrong shape.

Nvidia's response, announced at the show, is to take the storage control path off the CPU entirely. Its new framework, SCADA (scaled, accelerated data access), lets the GPU request storage directly. That complements cuFile, the API that already moved the bulk data path off the CPU using direct memory access (DMA), the technique that lets a drive write into a chip's memory without going through the processor. Together the two pieces form a single bet: the place where AI performance and cost are now decided is the path between storage and accelerator, not the accelerator itself.

The company also open-sourced cuFile and the surrounding stack at FMS, putting it under a new governance body called XIO-SIG with Google, Intel, and Meta as inaugural co-maintainers. The organization ships four repositories today: the cuFile API itself, a conformance suite for cross-vendor testing, a user-level implementation called libxFile, and xioLinux, a downstream kernel fork. Continuous integration is delegated to the platform vendors, which means each storage company runs the tests on its own hardware. That choice is the real governance signal. It moves conformance away from a single lab and onto every shop that wants the label.

The same show also formally launched Storage-Next, the storage-and-flash industry coalition Nvidia has been pushing since at least GTC 2025. The announced membership spans more than 40 storage and flash vendors, with DDN, KIOXIA, and Micron named in the rollout. Storage-Next is the part of the announcement the press will treat as the easy, feel-good story. It is also the part that deserves a closer read.

The lineage of SCADA is older than the marketing. It traces back to BaM, the Big Accelerator Memory research published at ASPLOS 2023 by teams at Nvidia, IBM, the University of Illinois, and the University at Buffalo. The earlier NVSHMEM work, which let GPUs initiate network transfers, set the same template: move the request off the CPU, then move the storage too. GROMACS, the molecular dynamics simulator, used NVSHMEM to overlap its halo-exchange communication (the small, frequent neighbor messages between parallel simulations) with compute, hiding a cost that used to be visible. SCADA is the storage analogue of that move. Nvidia's Vera CPU, paired with BlueField-4 STX, is cited at up to 3.21x the throughput of x86 in a two-stage compression-and-encryption pipeline. That number is Nvidia's own benchmark on a pipeline Nvidia designed, so the comparison is real but not independent.

Which leaves the honest question. A storage stack that Nvidia open-sources and Nvidia's closest compute peers co-maintain is, in practice, the storage stack for anyone who wants to stay on the same roadmap as the largest GPU fleet on the planet. That is a real development, not a cynical one. The 512-byte random read problem is real, and BaM is real research. The open-sourcing is also a way to set the de facto interface that future storage hardware will speak. The honest read is to take both at once: the productive-use era of AI infrastructure is a genuine shift, and the same announcement is the platform-anchor play that comes with it.

What to watch next is whether the conformance suite turns into a real certification that storage buyers can demand in writing, or stays a marketing checkbox.