HPE this week unveiled four ProLiant Gen 13 servers targeting dense AI inferencing and agentic workloads, with the flagship 10U ProLiant DL585a supporting up to eight double-wide GPUs and two AMD Epyc CPUs with 256 cores each when it ships in March 2027.

The launch lands against a backdrop of unprecedented infrastructure spending. Analysts at Bain & Co predict annual AI infrastructure spending could reach $1.5 trillion by 2031, covering datacenter buildouts, compute capacity, GPU upgrades, memory and networking. Power capacity in the most advanced AI datacenters, now approaching 1 gigawatt, could double by next year and exceed 9 GW by the end of the decade.

Capital spending from Microsoft, Amazon, Google, Oracle and Meta is expected to hit $780 billion this year — roughly five times the level of three years ago, according to the Bain report released in late September.

What do the new systems offer?

All four Gen 13 systems run AMD's 6th Gen Epyc 9006 SP7 "Venice" CPUs, released by the chipmaker in July. Common improvements span compute, memory and networking, alongside upgraded management and security software in HPE's iLO 8.

ProLiant DL585a — 10U, air-cooled, available March 2027:

• Up to eight double-wide GPUs from Nvidia, AMD or Intel
• Two Epyc CPUs, up to 256 cores each
• Sixth-generation PCI connectivity
• Built for inferencing, retrieval augmented generation (RAG) and agentic AI token throughput

"Most of our customers who are using a PCIe-type GPU, what they're really telling us is, 'We're running out of power long before we run out of space, so make it bigger,'" said John Carter, vice president of product management for HPE Compute. "We want to keep air cooling, so that's really where this is positioned. This is more for the enterprise, for the mid-tier service providers."

ProLiant DL525 — 1U, single-socket, available next month:

• Single 256-core AMD Epyc chip
• 16-channel high-bandwidth memory, DDR5 and MRDIMM support
• Targets AI inference, EDA and fraud detection in standard 19-inch racks

Carter framed the DL525 as a sign of shifting workload profiles: "Now we're shifting from GPU-focus to CPU-focus. We're expecting the ratio to shift from two to eight CPUs to GPUs, now more like one to one." As agentic AI scales, he said, customers will need dense CPU platforms to offload actions performed on their own processors.

What about the XD series?

Two additional systems, the XD245 and XD285, ship next year as follow-ons to HPE's Apollo 2000/XD 2000 series. Both use a 2OU chassis in an open 21-inch ORV3 rack infrastructure with shared power and shared cooling between systems.

• XD245: four half-width dual-socket nodes, liquid-cooled
• XD285: two half-width dual-socket nodes, air-cooled
• Both support 256-core AMD CPUs, 16 memory channels per CPU and MRDIMM

"What you really get here is the ultimate density in compute," Carter said. "I can drive many, many, many cores in the same data space." The systems target operators of modernized datacenters with high power delivery to the rack and either liquid or high-efficiency air cooling.

How does HPE address AI security?

Security features in iLO 8 respond to risks HPE associates with AI workloads. "That kind of scale and performance brings new security risks," Carter said. "We're talking about companies taking their most core critical IP and data and now putting it into these kinds of LLMs and simulations."

Key capabilities include:

• Self-encrypting HDD and SSD drives that encrypt and decrypt data from the moment the system boots
• Multi-party authorization requiring two authorized human approvers for high-impact AI or agent actions, such as key provisioning and secure erase
• Expanded post-quantum cryptography (PQC) within the management construct

The PQC work addresses the "harvest now, decrypt later" threat, in which actors steal encrypted data now and plan to decrypt it once quantum computing becomes available to them. HPE is "ensuring that we've got both quantum cryptography algorithms and capabilities already in the management construct," said Chris Bradley, director of mainstream computer customer advocacy and technical enablement at HPE. Multi-party authorization rules are embedded into individual nodes within iLO.

Is the buildout sustainable?

The Bain analysts who documented the spending surge also raised a caution. "The unprecedented speed and scale of the AI buildout, with billions flowing into chips, data centers, networks, and power systems, have focused attention on the challenge of building capacity," they wrote. "But the more important question may be whether enough economic value can be created to justify it."

For now, HPE is betting that demand holds. The Gen 12-to-Gen 13 transition pairs denser silicon with the security and management layer the company argues becomes more critical as enterprises feed proprietary data into large language models and simulations.