NVIDIA and HPE expand an “AI factory” aimed at enterprise agents, adding new CPU and agent-governance tools

NVIDIA used HPE Discover Las Vegas this week to press its case that enterprise AI is moving beyond chatbots and toward “agentic” systems that can plan, call tools, and run multi-step workflows with tighter controls. The company, together with Hewlett Packard Enterprise, outlined an expanded HPE AI Factory with NVIDIA, positioning the stack for production deployments where governance, data protection, and predictable performance matter as much as raw model quality.

The announcement centers on an updated set of compute, networking, and software components that HPE says are intended to move agentic AI from proof of concept to running workloads in corporate environments. NVIDIA said the expansion includes NVIDIA Vera CPU, the NVIDIA Agent Toolkit for HPE Private Cloud AI, and broader NVIDIA Confidential Computing coverage across the HPE AI Factory portfolio through HPE Services.

A key hardware datapoint is a planned server that pairs HPE’s ProLiant Compute DL394 Gen12 with the NVIDIA Vera CPU. NVIDIA said the configuration is targeted to be available in 2027, bundled with HPE Private Cloud AI as a turnkey AI factory co-engineered with NVIDIA. The goal, according to NVIDIA, is to bring deterministic, low-latency performance to the “agent loop” tasks, including orchestration and real-time data processing required when agents execute tool calls and react to new information.

NVIDIA characterized Vera as the first CPU built specifically for agents, designed around the operational demands of tool calling and orchestration rather than general-purpose server workloads. As part of that broader push, the company said Vera CPU is part of the NVIDIA Vera Rubin platform, which it described as ramping into full production. NVIDIA added that the Vera Rubin NVL72 rack-scale system will be available from HPE and will ship with full-stack NVIDIA Confidential Computing across every chip.

The companies also detailed an agent-focused software layer. NVIDIA said the NVIDIA Agent Toolkit for HPE Private Cloud AI will include NVIDIA NeMoTron open models, NVIDIA OpenShell (described as a secure runtime), and NVIDIA NemoClaw blueprints. NVIDIA said the toolkit is meant to function as an agentic AI “operating system” for monitoring agent behavior, enforcing governance policies, and supporting safer construction and execution of autonomous, long-running multi-agent systems.

Governance is not only implemented in the model layer, NVIDIA said, but also in the control plane around agent execution. HPE Private Cloud AI, the companies said, adds secure local agent registration, enabling customers to approve models, skills, and tools against centralized governance and security policies before agents run. NVIDIA also tied HPE Zerto Software capabilities into the stack by saying the system can detect “rogue” agent actions and use continuous data protection to rewind to a known clean state.

On the data and storage side, NVIDIA said HPE Alletra Storage MP X10000 has achieved the NVIDIA-Certified Storage “foundation level,” and that it can apply metadata and governance policies to unstructured data as it is prepared for AI pipelines. NVIDIA linked that workflow to improving token throughput, an operational metric that reflects how efficiently the pipeline can convert text and other data into inputs for model training or inference.

Networking and secure deployment were also prominent in the technical rundown. NVIDIA said the HPE AI Factory at scale, sovereign AI factory, and HPE Private Cloud AI are available with a full set of NVIDIA components, including NVIDIA RTX PRO 6000 Blackwell Server Edition GPUs, NVIDIA Spectrum-X Ethernet networking, NVIDIA BlueField-3 DPUs, and NVIDIA ConnectX-8 SuperNICs. For “every Vera Rubin NVL72 system,” NVIDIA said the plan is for built-in NVIDIA networking, including Vera BlueField-4 DPUs, ConnectX-9 SuperNICs, and NVIDIA Spectrum-X Ethernet with Spectrum-6 switching, which it described as delivering 1.6 times higher networking performance for AI communication compared with off-the-shelf Ethernet.

NVIDIA and HPE also highlighted confidential computing and zero-trust networking controls as the security backbone for agent workloads that may touch private or sensitive information. NVIDIA said confidential computing protects models and private data during execution for on-premises and sovereign deployments by establishing a chain of trust through cryptographic attestation and encryption. The companies said NVIDIA BlueField DPUs and NVIDIA DOCA provide in-silicon policy enforcement, runtime threat detection, and network encryption without performance tradeoffs, and that specific HPE ProLiant systems have been certified under NVIDIA’s confidential-computing program.

What remains less specific in the announcement is the timeline and packaging details for each component beyond the 2027 Vera CPU and DL394 Gen12 server configuration. While NVIDIA named an early enterprise customer, it did not disclose performance benchmarks, pricing, or capacity targets. For customers evaluating the shift to agentic AI, the next announcement to watch will likely be how HPE and NVIDIA translate these building blocks into reference deployments, including what governance policies can be enforced out of the box, how quickly systems can recover from rogue actions, and whether the agent-optimized CPU roadmap aligns with the scaling demands of frontier-scale model workloads.