Beirit Harvey

On December 2, 2025, AMD and Hewlett Packard Enterprise (HPE) announced a deeper collaboration aimed at making rack-scale AI infrastructure more “open by design”.

AMD + HPE Double Down on Open, Rack-Scale AI With “Helios”

On December 2, 2025, AMD and Hewlett Packard Enterprise (HPE) announced a deeper collaboration aimed at making rack-scale AI infrastructure more “open by design”; not just in software, but in how the rack is built, wired, cooled, and networked. The headline: HPE plans to be among the first OEMs to bring AMD’s “Helios” rack-scale AI architecture to market globally in 2026, pairing it with a purpose-built Ethernet scale-up fabric developed with Broadcom.


Why This Announcement Matters

Most conve

On December 2, 2025, AMD and Hewlett Packard Enterprise (HPE) announced a deeper collaboration aimed at making rack-scale AI infrastructure more “open by design”; not just in software, but in how the rack is built, wired, cooled, and networked. The headline: HPE plans to be among the first OEMs to bring AMD’s “Helios” rack-scale AI architecture to market globally in 2026, pairing it with a purpose-built Ethernet scale-up fabric developed with Broadcom.


Why This Announcement Matters

Most conversations about AI infrastructure still revolve around GPU counts and headline FLOPs. This announcement is about something bigger: standardizing the rack as the unit of scale, and pushing the ecosystem toward interoperable, Ethernet-based scale-up networking. Creating a roadmap that’s intended to work across research labs, cloud providers, and enterprise deployments.


The Core Idea: Helios as a Rack-Scale “Full Stack”

AMD positions Helios as a rack-scale platform that pulls together:

 * EPYC CPUs
 * Instinct GPUs
 * AMD Pensando™ advanced networking
 * ROCm software

…so customers can deploy large clusters with fewer custom integrations and more predictable building blocks.


Ethernet Scale-Up, Built With HPE Juniper + Broadcom

A major focus is the scale-up network inside (and between) racks. HPE says its Helios systems will integrate a purpose-built HPE Juniper Networking scale-up switch and associated software, co-developed with Broadcom (including Broadcom’s Tomahawk 6 in HPE’s description). The intent is to deliver high-bandwidth, low-latency connectivity over Ethernet using the Ultra Accelerator Link over Ethernet (UALoE) approach, framing this as an “open standards” path rather than a proprietary interconnect story. 


Performance + Rack Details

Across the two releases, the companies describe a single rack configuration that, at the top end, connects 72 AMD Instinct MI455X GPUs per rack and targets up to 2.9 AI exaFLOPS (FP4) per rack. HPE also calls out 260 TB/s aggregate scale-up bandwidth and 31 TB of HBM4 in that rack concept.

AMD also notes Helios is built around OCP Open Rack Wide design, and references Pensando Vulcano NICs for scale-out networking alongside the ROCm ecosystem.


A Second Thread: “Herder” Supercomputer in Germany

The announcement also highlights “Herder,” a planned supercomputer for HLRS (University of Stuttgart) built on HPE Cray GX5000, using AMD Instinct MI430X GPUs and next-gen AMD EPYC “Venice” CPUs. The stated timeline is delivery in the second half of 2027, with service expected by the end of 2027.


What to Take Away

If you read between the lines, this isn’t just “another GPU rack.” It’s a signal that:

 1. Rack-scale integration is becoming the product, not a bespoke engineering exercise.
 2. Ethernet-based scale-up is getting serious vendor backing (OEM + silicon + standards positioning).

The industry push for “open” AI infrastructure is expanding from software into rack form factors and network fabrics.


What This Means for TensorWave

For TensorWave, this is a strong tailwind for the “open, Ethernet-first, AMD-native” story: when a major OEM like HPE productizes an AMD rack-scale architecture (and emphasizes open rack standards plus Ethernet-based scale-up), it validates the idea that AMD clusters aren’t a science project; they’re becoming a repeatable, ecosystem-backed way to build and scale serious training and inference. 

Practically, it also raises the bar (and the opportunity) for TensorWave to win by doing what hyperscalers and big OEM stacks often don’t: deliver hands-on migration help, tuned software patterns, and predictable operations on AMD Instinct GPUs. With TensorWave, customers get the benefits of this broader momentum without needing to become their own systems integrator.


About TensorWave

TensorWave is the AI neocloud purpose-built for performance. Powered exclusively by AMD Instinct™ Series GPUs, we deliver high-bandwidth, memory optimized infrastructure that scales with your most demanding models–training or inference. 

Ready to get started? Connect with a sales engineer. 

AMD Instinct™ MI355X GPUs Now Available

AMD + HPE Double Down on Open, Rack-Scale AI With “Helios”

Subscribe to our Blog

Stay ahead of the curve with the latest in AI, AMD accelerators, and all things TensorWave.

Product

Resources

Company

© 2025 TensorWave Inc. - All rights reserved.