NVIDIA has introduced a new line of BlueField data processing units that pack security directly into the silicon, aiming to protect the massive AI computing clusters the company calls AI factories. The chips, powered by NVIDIA's DOCA software framework, are built to handle agentic AI workloads — autonomous AI agents that act on their own — at a scale the company says is unprecedented.
What the DPUs do
Data processing units, or DPUs, sit between the CPU and the network, offloading tasks like packet processing, storage, and security so the main processor can focus on compute. The new BlueField line embeds security policies at the hardware level — what NVIDIA calls in-silicon protection. That means threats get blocked before they ever reach the operating system or application layer. For AI factories running thousands of GPUs and petabytes of data, that speed matters. Agentic AI workloads are especially demanding because multiple AI agents need to communicate and share data in real time. A breach or slowdown in one agent can cascade across the system.
Why AI factories need hardware-level security
AI factories are the massive data centers NVIDIA designs for training and running large-scale AI models. They link hundreds or thousands of accelerators with high-speed networking. The scale creates new attack surfaces: compromised firmware, malicious data streams, side-channel exploits. Software-only defenses can lag behind the pace of data movement. By baking security into the DPU's silicon, NVIDIA can enforce policies at wire speed without consuming CPU cycles. The company argues that as AI becomes more autonomous — with agents making decisions without human intervention — the security backbone has to be just as automated.
DOCA's role
DOCA is NVIDIA's software development kit for the BlueField family. It gives developers a consistent set of APIs to program the DPU's hardware accelerators. With this announcement, DOCA has been updated to support the new in-silicon security features. The framework handles encryption, firewalling, and access control directly on the DPU. That abstraction layer means data center operators can write once and deploy across different BlueField generations without rewriting low-level firmware.
NVIDIA hasn't said exactly when the new BlueField DPUs will ship to customers or what they'll cost. The company typically integrates its DPUs into reference architectures for server makers and cloud providers. For now, the announcement signals that security is moving from the software stack into the hardware itself — a shift that reflects the growing scale and autonomy of AI workloads.
