Nvidia’s BlueField-4 STX: Turning Storage Into a Context Memory Layer for Agentic AI

Nvidia is reframing a familiar AI performance problem. When an AI agent loses context mid-task or GPUs sit idle waiting on data, the issue is often blamed on the model or orchestration layer. With its new BlueField-4 STX reference architecture, announced at GTC 2026, Nvidia is arguing that the real bottleneck in many agentic AI systems is the storage path — and that the remedy is a dedicated context memory layer between GPUs and traditional storage.

The AI context bottleneck: why storage, not just GPUs, is failing agents

As large language models (LLMs) are wrapped in agents that plan, call tools, and work across multiple sessions, they depend on a growing volume of key-value (KV) cache data. This cache is the internal record of what the model has already processed: intermediate attention calculations and state that let the model avoid recomputing everything at each token step.

Nvidia’s core claim is straightforward: the path between GPUs and that KV cache has become a limiting factor. In many current deployments, cache data is forced to traverse a traditional storage stack optimized for capacity and durability, not for the latency and throughput patterns of agentic inference. The result is slow retrieval, reduced GPU utilization, and agents that appear to “lose the thread” simply because they cannot access their working memory at the pace the GPU can consume it.

From an infrastructure perspective, this recasts familiar symptoms. Dropped context, unpredictable latency, and difficulty scaling multi-step agent workflows are not just model tuning issues. They increasingly reflect the misalignment between general-purpose storage architectures and AI-native workloads that depend on rapid, fine-grained access to context.

What Nvidia BlueField-4 STX actually is (and isn’t)

BlueField-4 STX is not a boxed product you can order. Nvidia is positioning it as a modular reference architecture for partners rather than an end-user SKU. The company is distributing STX to its storage ecosystem so vendors can design systems that expose a dedicated, GPU-adjacent context memory tier.

The performance targets Nvidia cites for this architecture are substantial when compared with traditional CPU-based storage paths:

• Up to 5x token throughput
• Up to 4x energy efficiency
• Up to 2x data ingestion speed

Nvidia has not disclosed the precise baseline configuration used for these comparisons. For architects, that caveat matters: these figures are directional, not yet tied to a published reference system you can benchmark against.

Crucially, STX encompasses both hardware and software guidance. Alongside the hardware reference design, Nvidia is providing a software reference platform built on its DOCA framework, expanded with a new component referred to as DOCA Memo. Together, these are meant to give storage partners a programmable foundation for building “AI-native” storage behavior tuned to agentic workloads.

Inside the STX architecture: BlueField-4, CMX, and DOCA Memo

At the heart of STX is a new storage-optimized BlueField-4 processor. It combines Nvidia’s Vera CPU technology with the ConnectX-9 SuperNIC and is designed to operate on Spectrum-X Ethernet networking. The platform is programmable via Nvidia’s DOCA software, which is already used to offload and customize data center networking and security functions on BlueField DPUs.

On top of this, Nvidia is defining a first rack-scale implementation called the Nvidia CMX context memory storage platform. CMX effectively extends GPU memory with a specialized, high-performance layer dedicated to KV cache operations. Rather than shuttling context through general-purpose NAS or object stores, CMX is designed to keep that cache close and responsive to GPUs.

In practice, this means:

• KV cache data generated during LLM inference is written to and read from a dedicated context memory tier.
• The data path is optimized for low-latency, high-throughput access patterns associated with multi-step agent reasoning.
• Traditional storage remains in play for capacity, durability, and broader data lifecycle management, but is no longer the primary medium for hot agent context.

Ian Buck, Nvidia’s vice president of hyperscale and high-performance computing, framed the issue directly, noting that traditional data centers “provide high-capacity, general-purpose storage, but generally lack the responsiveness required” for agents that must work across many steps, tools, and sessions. STX, paired with an expanded DOCA stack, is intended to give storage vendors the knobs they need to tune behavior for what Nvidia has begun calling the “agentic AI factory.”

The partner ecosystem: from storage incumbents to AI-native clouds

Nvidia is backing STX with a broad partner list that spans both classic enterprise storage vendors and AI-focused cloud providers.

On the storage side, co-design partners include:

• Cloudian
• DDN
• Dell Technologies
• Everpure
• Hitachi Vantara
• HPE
• IBM
• MinIO
• NetApp
• Nutanix
• VAST Data
• WEKA

Manufacturing partners building systems around STX include AIC, Supermicro, and Quanta Cloud Technology. On the cloud and AI services side, early adopters and supporters include CoreWeave, Crusoe, IREN, Lambda, Mistral AI, Nebius, Oracle Cloud Infrastructure, and Vultr.

For enterprise buyers, this mix is significant. Nvidia is not confining STX to a narrow set of hyperscalers. Instead, it is aiming to make STX a reference standard for any storage infrastructure expected to serve agentic AI workloads. Given the breadth of vendors engaged, many organizations are likely to see STX-aligned offerings show up inside existing strategic relationships rather than as entirely new supplier engagements.

STX-based platforms from partners are expected to become available in the second half of 2026.

IBM and Nestlé: a concrete look at data-layer acceleration

IBM features prominently in this story, both as a storage co-designer for STX and as a provider of infrastructure that Nvidia itself is adopting. Nvidia has selected the IBM Storage Scale System 6000 — certified and validated on Nvidia DGX platforms — as the high-performance storage foundation for its own GPU-native analytics environment.

Separately, IBM and Nvidia announced an expanded collaboration that illustrates the impact of GPU-accelerated data layers. IBM has integrated its watsonx.data Presto SQL engine with Nvidia’s cuDF library, enabling GPU-accelerated analytics. In a production proof of concept with Nestlé, this stack was used to accelerate a data refresh cycle across the company’s Order-to-Cash data mart, spanning 186 countries and 44 tables.

The results reported by IBM:

• Data refresh time reduced from 15 minutes to 3 minutes
• 83% cost savings
• 30x price-performance improvement

This example is drawn from structured analytics rather than agentic inference. It does not directly measure LLM token throughput or KV cache performance. But it supports the shared thesis from IBM and Nvidia: the data layer is frequently the limiting factor in enterprise AI performance, and GPU-accelerating it can generate substantial, measurable gains in production environments.

What STX changes for enterprise AI infrastructure strategy

For AI architects and infrastructure planners, STX is a signal that storage is moving from a background concern to a first-class design decision in AI stacks. Historically, GPU procurement and networking dominated planning conversations, while storage was often treated as an extension of existing NAS or object infrastructure.

Agentic AI stresses that assumption. General-purpose storage systems were not built to deliver KV cache data at the latencies modern LLMs can exploit. The result is an increasingly visible gap between raw GPU capability and end-to-end system throughput. STX-based systems from vendors such as Dell, HPE, NetApp, and VAST Data are Nvidia’s proposed answer: introduce a context memory layer, programmable via DOCA, that is explicitly shaped around AI workloads.

For organizations planning AI infrastructure refreshes, a few practical implications follow:

• Plan for a new tier: Storage architectures will likely evolve toward a three-tier pattern: GPU memory, a context memory layer optimized for KV cache, and traditional capacity storage.
• Include programmability in RFPs: As DOCA and DOCA Memo become part of vendor offerings, the ability to tune storage behavior for specific agentic workflows may become a differentiator.
• Scrutinize performance claims: Nvidia’s 5x/4x/2x gains are relative to unspecified CPU-based baselines. Before those numbers drive investment, enterprises will need clarity on test conditions and alignment with their own workloads.
• Align timelines with product availability: With STX-based platforms expected in the second half of 2026, organizations considering major storage refreshes in the next 12 months should factor potential STX options from existing vendors into their planning cycles.

In short, Nvidia’s BlueField-4 STX frames context memory as an explicit layer of the AI infrastructure stack, not an emergent property of whatever storage happens to sit behind GPUs. For enterprises betting on agentic AI, that shift is likely to make storage architecture and its programmability central to performance discussions over the next planning horizon.