AI storage firm Vast Data has launched native integration of its operating system available on Nvidia BlueField-4 DPUs in a bid to service inference sessions in the agentic era.

Leveraging those DPUs alongside Spectrum-X Ethernet networking, the release complements the extension of Nvidia’s Rubin Platform in organizing AI (artificial intelligence) infrastructure into compute pods, where Nvidia's Inference Context Memory Storage (ICMS) platform offers AI-native storage infrastructure designed for gigascale inference.

The Nvidia stack sees BlueField DPUs as the storage engine handling context movements close to the GPU, while Spectrum-X networking keeps access predictable under load. Dynamo, meanwhile, coordinates inference behavior, and ICMS defines context as system memory. Vast stores, shares, and reuses the context efficiently across the pod.

Like Nvidia, Vast anticipates inference shifting from isolated prompts to ongoing, multi-turn reasoning across agents. This evolution undermines the assumption that context remains confined to local GPU memory.

The Vast AI OS integration therefore provides a persistent, shared KV cache layer to prevent memory bandwidth bottlenecks, as seen when GPUs stall waiting for context via cache rather than compute. This is helped along by Vast’s parallel Disaggregated Shared-Everything (DASE) architecture where, as Vast Data CEO Renen Hallak recently told SDxCentral, users add nodes to "get more performance, more resilience, and more capacity efficiency."

Such architecture allows each host to access a unified, globally coherent context namespace. This eliminates coordination overhead that creates scale-up bottlenecks, and as a result, context streams efficiently from GPU memory to persistent NVMe storage via RDMA fabrics.

The Vast announcement extends its long history with Nvidia, when Vast Data launched its first joint reference architecture with Nvidia DGX A100 systems in January 2021.

"From the early days, we were there with Nvidia right next to us," Hallak recently told SDxCentral. "We were partnering with them for many years. In fact, the first external vendor that came to visit us when we started vast was Mellanox, which was at the time, separate from Nvidia, but then Nvidia bought them as part of their networking farm. We're good friends with them, and they help us out. We help them out, and our customers tend to, again, wherever they use Nvidia gear, they tend to use our software on top."

Vast and Nvidia's joint push comes after the latter showcased BlueField-4 playing a central role in its new ICMS platform. On display at CES 2026, the offering will benefit from the DPU's 800 Gb/s throughput and tight integration with the DOCA framework and Spectrum-X.

The agreement also comes after Nvidia all but acquired AI chip designer Groq, in a deal that saw it license Groq's tech for a likely $20 billion.

Nvidia CEO Jensen Huang claimed the Groq deal would extend Nvidia's AI factory architecture "to serve an even broader range of AI inference and real-time workloads.”