NVIDIA is pushing storage deeper into the AI performance conversation, arguing that bigger datasets and longer context windows make storage an active part of the compute path rather than a passive place to keep files.
At the Future of Memory and Storage conference, the company highlighted systems designed for AI factories where GPUs can initiate large numbers of storage requests directly. Those requests still need encryption, compression, verification and reconstruction, which can become bottlenecks when many agents or workloads access data at once.
NVIDIA says its Vera CPU, part of the Vera BlueField-4 STX platform, delivered up to 3.21 times higher throughput than an x86 CPU in a two-stage compression and encryption pipeline cited in its technical material. The company’s broader claim is that accelerated computing changes the old tradeoff between keeping data in memory and keeping it on storage.
The practical message for infrastructure buyers is that AI performance is no longer only about GPUs. Feeding those GPUs securely and quickly may require rethinking storage, networking and data services together.