HPE Introduces the First Next-Generation ProLiant Servers for AI Workloads

The rapid adoption of artificial intelligence continues to reshape data center infrastructure, affecting compute capacity, networking, memory, power consumption, and operating costs. Those effects are expected to become more significant over the coming decade.

A Bain & Co. report released in September projects that annual spending on AI infrastructure could reach $1.5 trillion by 2031. The estimate includes new data center infrastructure and compute capacity, as well as GPU upgrades, memory, and networking equipment. Power capacity in the most advanced AI data centers is approaching 1 gigawatt. It could double by next year and exceed 9 GW by the end of the decade.

Hyperscaler infrastructure spending is also increasing. Bain analysts expect capital spending from Microsoft, Amazon, Google, Oracle, and Meta to reach $780 billion this year, approximately five times the level of three years ago.

Four ProLiant Gen 13 systems

Hewlett Packard Enterprise is responding with the first four systems in its ProLiant Gen 13 family. The servers are designed for enterprise AI inference and agentic workloads, with increased compute, memory, and networking capabilities. HPE is also emphasizing security, performance, and reliability, which become especially important as AI systems process sensitive data and operate across large numbers of servers.

John Carter, vice president of product management for HPE Compute, said the first four 13th-generation ProLiant systems target dense computing and GPU workloads. Potential users include traditional enterprises seeking higher infrastructure performance and organizations developing agentic AI systems or large-model simulations.

As workloads expand, Carter said, they introduce additional security risks and management complexity. Organizations are placing core intellectual property and sensitive data into large language models and simulations, increasing the importance of hardware-level security and a silicon-based root of trust.

The new servers are accompanied by updates to the software used to manage and protect them. All four systems use AMD’s 6th Gen Epyc 9006 SP7 “Venice” processors, which AMD released in July.

ProLiant DL585a

The ProLiant DL585a is a 10U server that supports up to eight double-wide GPUs from Nvidia, AMD, or Intel. It can house two AMD Epyc processors, each with up to 256 cores, and includes sixth-generation PCI connectivity.

Scheduled to become available in March 2027, the system is designed for the power and thermal requirements of current and future AI workloads. HPE is targeting applications including inference, retrieval-augmented generation (RAG), and agentic AI, which require increased token throughput and PCIe capacity.

The DL585a is intended primarily for enterprise and mid-tier service-provider environments. HPE says many customers using PCIe-based GPUs reach power limits before exhausting available rack space. The system therefore focuses on expanding capacity while retaining air cooling.

ProLiant DL525

The ProLiant DL525 is a single-socket, air-cooled 1U server scheduled to become available next month. It contains one 256-core AMD processor, 16-channel high-bandwidth memory, and support for both DDR5 and MRDIMM memory.

The system is designed for high-throughput workloads such as AI inference, electronic design automation (EDA), and fraud detection. It fits a standard 19-inch rack.

HPE positions the DL525 for workloads that are shifting from a primarily GPU-focused architecture toward greater CPU use. In agentic AI systems, CPU resources can execute or coordinate many of the actions surrounding GPU processing. HPE expects the CPU-to-GPU ratio in some environments to move from roughly two to eight CPUs per GPU toward a more balanced one-to-one ratio.

XD245 and XD285

The two new XD systems are scheduled for release next year. According to HPE, they continue the company’s earlier Apollo 2000 and XD 2000 product lines.

Both systems are designed for open 21-inch ORV3 rack infrastructure. The wider, specialized racks provide shared power and cooling capabilities, allowing greater compute density. The systems are intended for modernized data centers with high power density per rack and support for liquid cooling or high-efficiency air cooling.

The XD245 and XD285 each use a 2OU chassis. The XD245 contains four half-width, dual-socket nodes, while the XD285 contains two such nodes. Each system supports AMD processors with up to 256 cores, 16 memory channels per processor, and MRDIMM memory. The XD245 uses liquid cooling, while the XD285 uses air cooling.

Management and security updates

Enhanced management and security capabilities are included across the new systems, including updates to HPE iLO 8. The changes address security concerns associated with AI and agentic AI workloads.

With iLO 8, supported HDD and SSD storage drives automatically encrypt and decrypt data stored on the hardware from system startup. The software also adds multi-party authorization for high-impact operations performed by AI systems or agents. Actions such as key provisioning and secure erase require approval from two authorized human approvers. HPE has implemented these controls at the individual-node level as rules within iLO.

HPE has also expanded its post-quantum cryptography capabilities. These features are intended to help protect data against future quantum-computing attacks. Existing encryption methods, including AES-256, are effective against current threats, but data protected with today’s methods could become more vulnerable if attackers gain access to sufficiently powerful quantum computers.

One related concern is the “harvest now, decrypt later” strategy. In this scenario, threat actors collect encrypted data today and retain it until quantum computing allows them to decrypt it.

HPE says its management framework already includes post-quantum cryptography algorithms and related capabilities.

The new ProLiant systems expand the infrastructure options available to enterprises and high-performance computing organizations building AI data centers. However, the scale of investment also raises questions about the economic returns from the AI buildout. Bain analysts noted that the rapid expansion of chips, data centers, networks, and power systems has concentrated attention on building capacity, while the more important question may be whether sufficient economic value will be created to justify that investment.