In addition to announcing details of the Vera Rubin platform, NVIDIA also launched a key technology at GTC 2026 described as "the reinvention of storage space"—the BlueField-4 STX storage architecture. This new modular reference architecture will solve one of the most pressing problems for the emerging "Agentic AI": when AI agents need to perform multi-step reasoning, call tools, and maintain "memory" during long conversations, traditional storage systems simply cannot keep up with the computing speed of GPUs.
Storage bottlenecks in proxy AI: When "memory" becomes a performance killer
As AI evolves from simply generating questions and answers to becoming "agents" capable of autonomous planning and execution, the underlying workflows have undergone fundamental changes. Modern AI agents need to collaborate across multiple steps, tools, and dialogue stages, meaning they must possess "long-term contextual memory." However, the high-capacity general-purpose storage provided by traditional data centers lacks the responsiveness to collaborate with GPUs in real time.
The crux of the problem lies in the "key-value cache" generated during the operation of large language models. When the model processes long contexts or performs multi-turn inference, the key-value cache grows rapidly. However, this type of data cannot be stored in expensive GPU high-bandwidth memory (HBM) for extended periods, nor can it be disposed of to slow traditional enterprise-grade storage. Pushing the key-value cache to traditional storage would cause the GPU to remain idle for extended periods while waiting for data, severely reducing inference throughput and increasing overall holding costs.
Therefore, NVIDIA believes that to perform reasoning and continuous learning within massive contexts, AI systems require a completely new category of storage architecture. The NVIDIA STX architecture design reinvents storage stacking, providing a modular foundation for AI-native infrastructure, enabling AI factories to maintain peak performance.
What is BlueField-4 STX? The birth of the "G3.5" layer.
To address this challenge, NVIDIA announced the BlueField-4 STX, a rack-class storage reference architecture powered by the BlueField-4 DPU. At its core is the new NVIDIA CMX Contextual Memory storage platform, which establishes a new storage tier called G3.5 between the GPU's high-speed memory (G1) and traditional slow storage (G4).
G3.5 is an Ethernet-connected flash memory layer optimized for KV caching, boasting petabyte-level storage capacity. It serves as "long-term working memory" for AI agents, allowing multiple agents to simultaneously share evolving dialogue contexts. When the GPU needs data, the CMX platform can preload the context at high speed via RDMA, preventing pauses in the decoding phase and thus avoiding delays in AI inference.
According to official data, this new architecture delivers three major performance leaps:
• 5x token throughput:Compared to traditional storage, the number of tokens processed per second is increased by up to 5 times.
• 4 times more energy efficient:Compared to traditional high-performance storage CPU architectures, energy efficiency is improved by 4 times.
• 2x data ingestion speed:The speed of acquiring enterprise AI data has been increased by 2 times.
Technical Foundation: The Trinity of Vera Rubin + BlueField-4 + Spectrum-X
The STX architecture didn't emerge out of thin air; rather, it's deeply integrated with NVIDIA's latest hardware product line. Accelerated by the Vera Rubin platform, it utilizes the new BlueField-4 DPU optimized for storage. This chip combines an NVIDIA Vera CPU with an NVIDIA ConnectX-9 SuperNIC, and operates with Spectrum-X Ethernet, the NVIDIA DOCA software framework, and NVIDIA AI Enterprise software.
In terms of specific operation, the BlueField-4 DPU is responsible for managing the hardware-accelerated KV cache placement, eliminating metadata costs, reducing unnecessary data migrations, and ensuring secure and isolated data access for GPU nodes. Meanwhile, the Spectrum-X Ethernet provides low-latency, high-bandwidth RDMA connectivity to ensure the consistency of KV cache data access.
Kevin Deierling, Senior Vice President of Networking at NVIDIA, likened contextual memory to human memory and emphasized that "effectively managing context is key to expanding multi-round, multi-user reasoning. This necessitates a radical change in how contextual memory is managed on a large scale."
The entire ecosystem provides support: from major storage manufacturers to cloud service providers.
Like other NVIDIA technologies, the release of the STX architecture was accompanied by extremely broad industry support. In his keynote address, Jensen Huang showcased a long list of partners, emphasizing that this was not just NVIDIA's effort, but a paradigm shift for the entire storage industry.
Storage vendors that have jointly designed next-generation AI infrastructure using the STX architecture include: Cloudian, DDN, Dell, Everpure, Hitachi Vantara, HPE, IBM, MinIO, NetApp, Nutanix, VAST Data, and WEKA. Hardware partners responsible for manufacturing STX architecture systems include AIC, Supermicro, and Quanta Computer (QCT).
On the end-user side, many leading AI labs and cloud service providers are planning to adopt the STX architecture system as a contextual memory storage solution, including: CoreWeave, Crusoe, IREN, Lambda, Mistral AI, Nebius, Oracle Cloud Infrastructure (OCI), and Vultr.
Timeline for listing
According to official NVIDIA information, all platforms based on the BlueField-4 STX architecture will be supplied through partners in the second half of 2026.
Analysis: Storage is no longer a supporting role, but rather the "second brain" of the AI factory.
The BlueField-4 STX unveiled by NVIDIA at GTC 2026 is significant beyond a simple product line update. It symbolizes that NVIDIA's thinking on AI infrastructure has evolved from "how to accelerate computing" to "how to organize data".
First, this is a precise filling of the gap in the infrastructure requirements for "agent-based AI". While the industry is still discussing how AI agents will change workflows, NVIDIA has already begun to solve the most troublesome engineering problem behind it: where should those conversation histories, tool call records, and intermediate inference steps with millions of tokens be stored so that the GPU can be "on demand" when needed? The STX architecture's answer is: don't cram the GPU to the brim, and don't throw it into a remote hard drive, but instead create a brand new "caching layer" in the data center.
Secondly, the emergence of the G3.5 layer will redefine data center storage budgets. In the past, when IT companies purchased storage devices, they often considered capacity, durability, and backup capabilities. However, in the STX architecture system, the KPIs for storage devices have become "how many tokens can be fed to the GPU per second" and "how many inferences can be generated per watt." This will force traditional storage giants to rethink their product positioning, or they will be marginalized in this wave of AI factory construction.
Furthermore, the ecosystem list reveals NVIDIA's "inclusive" strategy. The list includes not only existing storage solution providers like Dell, HPE, and NetApp, but also next-generation software-defined storage unicorns such as VAST Data and WEKA, as well as object storage providers like Cloudian and MinIO. This indicates that NVIDIA does not intend to dominate the entire market, but rather hopes to allow partners to develop differentiated solutions on a unified BlueField platform through an open architecture. This strategy is consistent with its GPU marketplace strategy—establishing standards and filling them with an ecosystem.
Finally, for companies planning to build large-scale AI services, the STX architecture provides an important direction for thinking: as the GPU computing power arms race gradually reaches its physical limits, the solution to the next performance bottleneck may not lie in the GPU itself, but in how to make the "memory" and "storage" around the GPU smarter.
