Nvidia has just made a strategic move that reshapes how GPUs access data. At the recent Future of Memory and Storage conference, the company open-sourced the cuFile APIs and the entire storage stack beneath them, setting up a new GitHub organization with Google, Intel, and Meta as founding maintainers. Simultaneously, it officially launched Storage-Next, an industry initiative involving over 40 flash and enterprise storage vendors. However, the core of the announcement is SCADA, the framework that moves the storage control path directly onto the GPU, enabling GPUs to fetch data from drives without the CPU brokering every request.
This evolution stems from a concrete problem that emerges when AI inference becomes the dominant workload. GPUs need to access large amounts of small objects, such as KV cache blocks and embeddings, often smaller than 512 bytes. Traditional SSD controllers are tuned for 4KB random reads, a legacy of decades of virtualization and databases. This mismatch creates a read amplification of about eight times, which can compromise overall system performance. SCADA bridges this gap by allowing the GPU to issue I/O requests independently, keeping hundreds of thousands of operations in flight, just as it already handles memory latency.
Sponsored Protocol
The 512-byte problem and the economics of KV cache
The main driver behind this initiative is the KV cache economics. During inference, the model must maintain the attention state for every token already processed. With ever-longer contexts and agentic deployments handling thousands of concurrent conversations, the KV cache quickly outgrows GPU memory. The standard approach is to spill the cache from GPU memory to system memory and flash, with continuous refills through small random reads. Serving these requests from flash rather than DRAM increases the number of concurrent users each GPU can support, reducing the per-user cost. That is why Nvidia focuses on 512-byte IOPS rather than raw bandwidth.
Sponsored Protocol
Open source as a lever to standardize hardware
The decision to open cuFile is a significant departure. This layer has always been part of CUDA and was introduced in 2019 with GPUDirect Storage, but now Nvidia makes it available to everyone. The goal is clear: a GPU-initiated storage interface only pays off if drive, controller, and array vendors build to it. And vendors will hardly build around a proprietary interface owned by a company that produces the very GPUs they feed. By publishing the interface and open-sourcing the implementation, Nvidia aims to make GPU-initiated storage the industry default, with the benefit of selling most of the GPUs involved.
Intel's participation is especially significant, as StorageReview notes. Intel, which supplies most of the x86 chips currently sitting in storage controllers, has signed on to maintain software designed to remove those chips from the I/O path. Google and Meta, while building their own accelerators, co-maintain a layer that standardizes how accelerators reach storage, confirming the direction toward an open and collaborative ecosystem.
Sponsored Protocol
Storage-Next and systems expected in H2 2026
Storage-Next is not entirely new; it has been discussed publicly since GTC 2025, but now it acquires a concrete membership and a framework to build against. Partner systems from DDN, Dell, HPE, IBM, VAST Data, and WEKA are due in the second half of 2026. Kioxia is developing XL-Flash drives optimized for 512-byte access, while Nvidia's roadmap calls for Gen7 SSDs sustaining 100 million IOPS each, a target controller vendors are designing toward. The initiative fits into a broader context where storage becomes the new competitive frontier of AI, as recent events in the industry show, such as the Buc-ee's lawsuit against a small business that highlights how even brands must defend themselves. For developers building AI applications, understanding how to optimize data access can make a difference, just as for those managing Java containers, where slim images and JVM tuning reduce startup times. Nvidia's openness represents a step forward toward a more efficient infrastructure, where the productivity of resource usage prevails over the amount of hardware owned.
Sponsored Protocol
