At CES in January 2026 NVIDIA announced an Inference Context Memory Storage Platform powered by its BlueField-4 data processor. At GTC in March it launched BlueField-4 STX, “a modular reference architecture” for accelerated storage; the first rack-scale implementation includes the NVIDIA CMX context memory storage platform.

BlueField-4 combines NVIDIA’s Vera CPU with a ConnectX-9 SuperNIC (network card), connected over Spectrum-X Ethernet and programmed with NVIDIA’s DOCA software. NVIDIA names twelve storage partners, among them Dell, HPE, IBM, NetApp, Nutanix, DDN, VAST Data and WEKA, and early adopters that are “planning to adopt”, among them CoreWeave, Crusoe, Lambda, Mistral AI, Nebius and OCI.

When a model answers, it builds a key-value (KV) cache: the intermediate results for every token of the conversation so far, so it does not recompute them for each new token. Long documents and multi-step agents make that cache very large. It is fastest in the GPU’s own memory (HBM), but HBM is small and expensive, and NVIDIA says “a KV cache cannot be stored on GPUs long term, as this would create a bottleneck for real-time inference”.

Without a fast tier below HBM, the cache is thrown away and recomputed when a user returns or an agent resumes, which spends GPU time on work already done. STX is NVIDIA’s design for that tier.

Think of three tiers. Hottest is the GPU’s HBM, holding the cache for requests being answered now. Next is the context storage tier: flash storage in the same pod, reached over the Spectrum-X Ethernet fabric, managed by BlueField-4 processors that move cache blocks in and out without loading the GPU servers’ own CPUs. Coldest is the general data store, for training data and files.

The inference software (NVIDIA cites its Dynamo framework elsewhere) decides which cache blocks to keep, evict or reload. The design only pays off when the same context is reused, for example by long chats, shared documents or agents that come back to a task.

IT Infrastructure

Inference clusters gain a storage tier sized for context, not for datasets. Capacity planning needs a view of how much context is reused, which depends on the workloads the operator expects to sell.

Networking

Context moves over the scale-out Ethernet fabric between GPU servers and storage; that traffic must be included when the fabric is sized.

Operations

A new tier to monitor: cache hit rates and reload latency become service metrics alongside GPU utilisation.

1. AnnouncedDescribed publicly; no date for customers to buy it.
2. Planned (current stage)A stated availability window that has not arrived yet.
3. ShippingIn production and reaching early customers or partners.
4. Generally availableOrderable by any customer through normal channels.
5. DeployedRunning in customers’ production environments, confirmed by them or the manufacturer.

NVIDIA said in March 2026 that “STX-based platforms will be available from partners in the second half of this year”, on a when-and-if-available basis. That window is now open, but no partner shipment has been confirmed for this entry.

• The building blocks: storageThe storage tiers in an AI data centre.
• Anatomy of an AI requestWhere the KV cache is built during serving.

1. NVIDIA Launches BlueField-4 STX Storage Architecture With Broad Industry Adoption(opens in a new tab)NVIDIA Newsroom · 16 March 2026 · Manufacturer
2. NVIDIA BlueField-4 Powers New Class of AI-Native Storage Infrastructure for the Next Frontier of AI(opens in a new tab)NVIDIA Newsroom · 5 January 2026 · Manufacturer

Figures are as the manufacturer states them unless marked otherwise. None has been independently tested by BSS/OSS Academy.