HPE is all in on AI, according to the message coming from CEO Antonio Neri. AI agents are now running alongside end users in enterprise infrastructure, changing how workloads move across networks and what compute and storage must deliver. At HPE Discover 2026 in Las Vegas this week, Neri used the opening keynote to detail the company's response across its full stack. Key announcements include:

• Networking: HPE extended AI connectivity from GPU racks to the inference edge with new QFX switches, the PTX 12,000 routing platform for data center interconnect, the SRX 4700 quantum-safe firewall, and the MX 301 edge router, while Marvis Actions comes to Aruba Central and Aruba CX switching to HPE Mist.
• Compute: Private Cloud AI now scales to 256 GPUs with multi-node inference across three AI Factory tiers, backed by the new ProLiant DL 394 Gen 12 built for agentic workloads.
• Storage: The Alletra MPX 10,000 becomes the Private Cloud AI storage layer, unifying file and object storage on a single architecture with native MCP support and Nvidia Certified Storage validation.
• Agentic operations: Private Cloud AI gains agentic governance controls including zero-code agent registration across any framework and a new three-tier identity model, backed by Nvidia Open Shell for isolated policy-enforced agent runtimes, NeMo Cloud for governed workflow blueprints, and Zerto for clean-state rollback when agents make errors.
• Cloud: HPE CloudOps consolidates virtualization, data protection and cloud management into a single hybrid operating layer, with the Unleash AI program now covering more than 60 validated partners.

"Today, we are witnessing one of the largest technology platform shifts in history," Neri said during the keynote. "Workloads and applications are moving from being driven by end users [to] now being driven by both end users and AI agents."

The network comes first

For HPE, the network is the AI foundation. "Every byte, every token, every decision, all of it crosses the network," Neri said. Much of Neri's keynote discussion about networking revolved around the expansion and integration of Juniper Networks as part of the broader HPE portfolio. HPE structured its portfolio across four layers: scale-up within a rack, scale-out across GPU clusters, data center interconnect, and edge inference routing. New QFX switches address the first two layers, while the PTX 12,000 handles data center interconnect with 800G routing, the SRX 4700 delivers quantum-safe firewall throughput at 1.44 Tbps in a single rack unit, and the MX 301 brings the MX platform to the inference edge on Juniper's sixth-generation Trio silicon.

Fundamentally, it's about speed and what that means in the AI era. Neri put the cost of latency at training scale in plain terms: "Multiply[ing] a small delay across hundreds of thousands of GPUs over weeks of training in your network can mean the difference between training a new model in 90 days or 30 days," he said. "It is the difference between chasing a breakthrough or making one."

The network portfolio also includes new software capabilities. Marvis Actions, the AI-driven network operations engine from HPE Aruba Networking, is now integrated into Aruba Central and Aruba CX switching platforms through HPE Mist. This allows IT teams to automate remediation of common network issues, such as misconfigurations or security policy violations, without manual intervention. The emphasis on automation aligns with the broader goal of reducing operational complexity as AI workloads multiply.

HPE scales compute for the agentic era

While networking connects systems, those compute systems are still needed and they are being organized and optimized for AI. HPE organizes its compute portfolio into three AI Factory tiers for enterprise, service provider, and sovereign deployments. "AI today is about moving faster from ambition to outcome, accelerating time to token, reducing execution risk, and ensuring your environments are ready to perform from day one," Neri said.

The new ProLiant DL 394 Gen 12 is built for agentic AI and long-context workloads. It leverages the latest AMD EPYC processors and supports up to 8 NVIDIA H200 GPUs in a compact 2U form factor. At the AI Factory at Scale tier, new configurations deliver AI training with one-quarter the GPUs required by the prior Blackwell-generation platform and inference at one-tenth the cost per million tokens, Neri said.

Private Cloud AI configurations now scale to 256 GPUs with multi-node inference. A unified gateway provides a single API for frontier and open-source model access. Shared cache reduces the cost per first token. "Private cloud AI can now serve larger models across multiple systems with multi node inference, so capacity grows with the math," Neri said. This multi-node capability is particularly important for enterprises running large language models (LLMs) that exceed the memory capacity of a single GPU node, enabling distributed inference without sacrificing performance.

The AI Factory tiers are designed to match different customer maturity levels. The Entry tier offers pre-validated reference architectures with 8 to 32 GPUs, suitable for proof-of-concept projects and departmental AI. The Enterprise tier scales to 128 GPUs with integrated networking and storage, targeting production workloads. The Service Provider tier supports up to 256 GPUs with full redundancy and multi-tenant isolation, aimed at cloud providers and large-scale AI factories.

Storage: Making data ready for AI

Agents are only as capable as the data behind them. On the storage front, the Alletra MPX 10,000 is now the storage layer for Private Cloud AI, unifying file and object storage on a single architecture. It adds real-time metadata enrichment and native MCP support, enabling agents to retrieve data across structured and unstructured sources. HPE cited 7 to 12 times faster time to value compared to custom-built environments.

"Your AI agents are only as smart as the data you use to train them," Neri said. "Traditionally, that data required custom preparation for every use case and months of building the right AI data pipelines, but not anymore."

The Alletra MPX 10,000 uses a disaggregated architecture that separates compute and storage controllers, allowing independent scaling. It supports NVMe over Fabrics for low-latency access and integrates with Nvidia's Magnum IO SDK for optimized data movement between GPUs and storage. The system also includes data reduction technologies such as inline deduplication and compression, which can reduce storage capacity requirements by up to 5x for typical AI datasets. Additionally, the storage platform is validated under the Nvidia Certified Storage program, ensuring compatibility with Nvidia AI Enterprise software and GPU clusters.

HPE also announced that the Alletra MPX 10,000 now supports object storage via the S3 API, making it suitable for data lakes and model artifact repositories. This unification eliminates the need for separate file and object storage silos, simplifying data management and reducing total cost of ownership.

Toward an agentic enterprise

Running on top of all that networking, compute and storage gear are AI agents. That's another area that HPE is looking to help. "Agents now reason across data, applications, models, and workflows. They help you make decisions, automate processes, and are increasingly taking action on your behalf," Neri said.

Agents are proliferating across enterprises, often in the hands of developers and small teams outside formal IT oversight, creating governance and scale challenges that traditional IT management was not built to handle. "Agentic AI demands a new set of enterprise requirements," Neri said. HPE's answer is a governed agent layer built into Private Cloud AI. Enterprises can register agents built in any framework, applying security controls on API calls, identity, and encryption with zero code changes required. A three-tier identity model verifies the user, governs the agent, and requires human approval for sensitive actions.

The agent governance framework includes several key components. Nvidia Open Shell provides isolated, policy-enforced runtimes for agent execution, preventing unauthorized data access or resource usage. NeMo Cloud offers governed workflow blueprints that allow organizations to define standard operating procedures for common agent tasks, such as data retrieval or report generation. Zerto, HPE's disaster recovery solution, is integrated to enable clean-state rollback when agents make errors, allowing administrators to revert to a known good state without losing data.

Zero-code agent registration is a standout feature. IT administrators can use a web-based console to register agents built with frameworks like LangChain, LlamaIndex, or custom Python scripts, applying consistent security policies without modifying the agent code. The three-tier identity model works as follows: the first tier authenticates the end user initiating the agent call; the second tier validates the agent's own identity and permissions; the third tier requires explicit human approval for actions that modify data or access sensitive systems. This layered approach mirrors the principle of least privilege, reducing the risk of unauthorized actions.

Power, research and what comes next

With all the promise of AI and all the infrastructure that goes with it, there is a key constraint that Neri warned about, and that's power. "Every model, every workload, every agent depends on power, because at its core, an AI factory is doing one thing: turning electrons into tokens," he said. He noted that the U.S faces a 19 gigawatt power gap by 2028, with data centers projected to account for nearly half of US electricity demand through 2031. "As AI scales, the future will not be defined by compute alone," Neri said. "It will be defined by how efficiently we can power it, cool it, and connect it."

HPE is addressing power constraints on multiple fronts. The company has expanded its liquid cooling portfolio, including direct-to-chip and immersion cooling solutions for high-density GPU clusters. New power management features in HPE OneView enable dynamic power capping based on workload demand, and HPE is working with utility providers to co-locate AI factories near renewable energy sources. Additionally, HPE's research division is exploring novel cooling techniques, such as two-phase immersion cooling and advanced thermal interface materials, to improve energy efficiency.

Beyond power, HPE is investing in AI research through collaborations with academic institutions and industry partners. The HPE AI Labs, announced earlier this year, focuses on areas such as efficient model architectures, federated learning, and AI safety. Neri hinted at future announcements in quantum computing and neuromorphic hardware, though no specific products were unveiled at Discover 2026.

HPE also highlighted the importance of sovereign AI—the ability for nations and enterprises to control their own AI infrastructure and data. The three AI Factory tiers include sovereign deployments that meet local data residency and regulatory requirements. This is particularly relevant in regions such as Europe, where GDPR compliance and digital sovereignty are top priorities for government and enterprise customers.

Source: Network World News