Hardware purpose built for agents
At HPE Discover in Las Vegas, Hewlett Packard Enterprise and NVIDIA expanded their joint AI Factory offering with a clear thesis. The next wave of enterprise AI is agentic, autonomous software that observes, decides, and acts, and the infrastructure to run it has to be designed differently. The centerpiece is the HPE ProLiant Compute DL394 Gen12, a server built on the NVIDIA Vera CPU. NVIDIA describes Vera as the first CPU built for agents, designed for the tool calls, orchestration, and real time data processing that define what the company calls the agent loop. The server is slated for availability in 2027.
The framing matters because it reflects a genuine shift in workload characteristics. Training large models is a throughput problem suited to massive parallel GPU clusters. Running agents in production is different. It involves frequent, latency sensitive operations as agents call tools, retrieve data, and chain decisions together. That pattern stresses the CPU, memory, and data path in ways that traditional AI infrastructure, optimized for training, handles poorly. By building silicon specifically for the agent loop, NVIDIA and HPE are betting that the economics of production agents will hinge on hardware tuned for orchestration, not just raw model compute.
A full stack, not just chips
HPE was careful to position this as more than a hardware refresh. HPE Private Cloud AI, the company's turnkey AI platform, gains support for the NVIDIA Agent Toolkit, which includes NVIDIA Nemotron open models, NemoClaw blueprints, and the NVIDIA OpenShell secure runtime. The intent is to give enterprises a pre integrated path from infrastructure to deployed agents, rather than leaving them to assemble the pieces themselves. For organizations that have struggled to move AI pilots into production, the appeal of a vendor assembled stack is that it collapses a long integration project into a supported product.
The supporting cast fills out the picture. The offering adds the NVIDIA Vera Rubin NVL72 rack scale system and the HPE Compute XD700 with NVIDIA HGX Rubin, along with NVIDIA RTX PRO 6000 Blackwell Server Edition GPUs and a networking and storage stack spanning Spectrum-X Ethernet, BlueField DPUs, and HPE Alletra Storage. HPE CEO Antonio Neri summarized the rationale plainly. As AI becomes more autonomous, organizations need a new architecture to run it securely, govern it responsibly, and scale it economically, he said. NVIDIA CEO Jensen Huang put it more sweepingly, arguing that every layer of the computing stack is being reinvented for the age of AI agents.
Security and governance move to the center
The most telling theme is the emphasis on security and governance, which signals that enterprise AI buyers have moved past the question of whether agents work to the harder question of whether they can be trusted in production. HPE is extending NVIDIA Confidential Computing across all AI Factory solutions in the fourth quarter of 2026, protecting data and models even while they are being processed. HPE Zerto software is being used to detect rogue agent actions and to handle secure local agent registration with governance enforcement, an acknowledgment that autonomous software introduces new categories of risk.
This focus reflects a maturing market. An agent that can act autonomously is, by definition, software that can do things without a human in the loop, and that autonomy is exactly what makes governance non negotiable. Enterprises need to know what their agents are permitted to do, to monitor what they actually do, and to detect when one behaves anomalously or is compromised. By building monitoring, confidential computing, and registration into the platform rather than bolting them on, HPE is responding to the concern we hear most from technology leaders evaluating agentic AI. The capability is no longer the obstacle. Control is.
The AI Factory as a competitive concept
The AI Factory framing is itself a strategic move. By packaging hardware, software, security, and governance into a branded, supported offering, HPE is competing not on individual components but on the promise of a complete, production ready system. This is a deliberate counter to the do it yourself approach, in which enterprises buy GPUs, assemble open source tooling, and shoulder the integration burden themselves. For organizations without deep AI infrastructure expertise, a turnkey factory is attractive precisely because it transfers that burden to a vendor with the engineering depth to carry it.
The risk for buyers is the familiar one of lock in. A tightly integrated stack spanning HPE hardware and NVIDIA software is convenient, but it concentrates dependence on two vendors whose roadmaps and pricing the customer does not control. The counterargument, which HPE and NVIDIA are making implicitly, is that the alternative of self assembly is slower, riskier, and more expensive in practice, and that the time to value of a supported platform justifies the dependence. Where an organization lands on that tradeoff depends on its own capabilities, but the choice is now explicit rather than theoretical.
Early validation and the road ahead
HPE pointed to early customer interest to validate the approach, naming the New York Stock Exchange as an organization exploring the Vera CPU with the HPE ProLiant DL394 Gen12. A reference customer in financial markets infrastructure carries weight, because few environments are more demanding on latency, reliability, and regulatory compliance. If a system can meet the bar for agentic workloads at an exchange, that is a meaningful signal about its production readiness. HPE also expanded its Unleash AI partner program with roughly a dozen new software partners, broadening the ecosystem around the platform.
The timeline is the practical caveat. The Vera based server is not available until 2027, and several capabilities, including Confidential Computing and additional agentic observability features, arrive in the fourth quarter of 2026. So while the vision is concrete, the full stack is still partly a roadmap. For technology leaders, the announcement is best read as a statement of direction from two of the most influential infrastructure vendors. They are betting heavily that production agentic AI is the next major enterprise workload, and they are building specialized infrastructure, with governance at its core, to capture it. Whether the timeline holds is the thing to watch.
