Nvidia unveils Vera CPU and open-sources cuFile as GPU-direct storage gains active role

At this week’s FMS, Nvidia focused not on a new graphics card but on the pipeline that feeds it. AI agents generate a thousand parallel storage requests directly from the GPU, and existing systems that handle encryption, compression and verification simultaneously become bottlenecks before the model even starts computing. The message is clear: without storage that runs at accelerator speed, the GPU sits idle.

The shift is fundamental: GPUs now issue I/O requests themselves, bypassing the CPU. This creates a load of a thousand concurrent operations that require immediate data handling, encryption, compression, verification and recovery. When a thousand agents hit storage together, those services stall and throughput collapses. Nvidia frames this as a 40-year economic shift: the classic trade-off between memory (expensive, fast, minute-scale latency) and storage (cheap, slow) was measured in minutes; today, with AI Storage solutions, the same trade-off plays out in microseconds.

The hardware answer is the Vera CPU, part of the Vera BlueField-4 STX platform. In a benchmark published on Nvidia’s technical blog, Vera delivered up to 3.21× higher throughput than an x86-based CPU in a two-stage compression-encryption pipeline. Practically, storage platforms can absorb AI data streams with far less compute infrastructure, turning storage from a passive file repository into an active component of the data path.

In parallel, Nvidia is releasing cuFile, the API layer of GPUDirect Storage, as open source, along with the underlying vertical stack. cuFile enables the GPU to read and write storage directly, using hundreds of thousands of GPU threads and wide-bandwidth HBM, achieving microsecond-scale access times. The move aligns with a security-first approach built on Linux best practices and feeds initiatives such as the Open Secure AI Alliance, whose initial contributors include Google, Intel, Nvidia and Meta. The goal is to make security and data pathways fast enough for AI-driven protection systems.

The third pillar is the Storage-Next initiative, led by Nvidia together with storage manufacturers, controller vendors, thermal and cooling engineers, orchestration operators and standards bodies. Its aim is to define how GPU-directed storage should behave and then codify those definitions into open, interoperable standards. In short, Nvidia is pushing the entire ecosystem toward extreme co-design—from memory and storage through software to standards—because without it even the most powerful GPU merely burns power while waiting for data.