AI hardware news drops onto our digital doorsteps so often now that it’s getting hard to tell one big announcement from the next. Faster GPUs, bigger racks, new interconnects… oh my! But HPE’s support for AMD’s Helios platform hints at something bigger.
Competition is opening up, Ethernet is finally getting its moment, and the channel is slowly moving toward AI infrastructure that feels more open and a lot less locked down.
HPE and Helios extend their longtime partnership into new GPU play
At HPE Discover 2025 in Barcelona, HPE said it will be “one of the first OEMs” to adopt AMD’s Helios rack-scale platform. Basically, this means that Helios is built to make a full rack of 72 GPUs act like one massive accelerator. It’s based on Meta’s Open Rack Wide design and pairs AMD’s upcoming Instinct MI455X GPUs with new EPYC “Venice” CPUs and Pensando networking tech.
“HPE has been an exceptional long-term partner to AMD, working with us to redefine what is possible in high-performance computing,” said Dr. Lisa Su, chair and CEO of AMD. “With ‘Helios’, we’re taking that collaboration further, bringing together the full stack of AMD compute technologies and HPE’s system innovation to deliver an open, rack-scale AI platform that drives new levels of efficiency, scalability, and breakthrough performance for our customers in the AI era.”
The real standout, though, is the networking. HPE’s Juniper team built what it calls the “first scale-up switch to deliver optimized performance for AI workloads over standard Ethernet.” Developed with Broadcom and powered by the Tomahawk 6 chip, the switch uses Ultra Accelerator Link over Ethernet (UALoE) instead of a proprietary interconnect, providing the entire system with an open, Ethernet-based backbone specifically designed for AI.
Complicated, eh?
Rami Rahim, president and general manager of HPE’s networking business, summed it up clearly, saying, “This is an industry first scale-up solution using Ethernet, standard Ethernet. So that means it’s 100 percent open standard and avoids proprietary vendor lock-in.”
For cloud providers and neoclouds, the folks this rack is built for, an open, standards-based approach is a pretty big deal.
The Helios rack specs you need to know
One Helios rack has up to 2.9 exaflops of FP4 compute, 31 TB of HBM4, and 260 TB/s of bandwidth. In plain (ish) English, it’s built for the very biggest AI training jobs and trillion-parameter models. It’s liquid-cooled, modular, and built for dense environments where efficiency super matters.
HPE and Broadcom’s work on scale-up Ethernet closes a pretty massive gap that has historically kept Ethernet from being competitive with alternatives like NVLink. Putting UALoE at the heart means that HPE is trying to give cloud providers a path to high-performance AI infrastructure without locking them into a single vendor’s ecosystem.
What this means for the channel
Helios isn’t a broad channel play yet, but the trend it represents is worth paying attention to. There’s definitely a clear pull toward more open, standards-based AI infrastructure.
For the channel, that shift really matters. It paints a picture where partners won’t have to deal with as many proprietary pieces or locked-down systems. As more vendors pursue open networking and modular designs, building and integrating AI environments should become easier. Competition will be stiff, though.
Where this is (probably?) going
HPE expects Helios to be available globally in 2026, marking the first time AMD’s rack-scale design actually hits the market. It’s not a replacement for Nvidia, but it does give hyperscalers and neoclouds a solid alternative built around familiar, open Ethernet.
Basically, the AI infrastructure world is trying to get more open, more flexible, and a little less proprietary. Huzzah.
While Helios boosts performance inside the rack, Nvidia’s Spectrum-XGS update from August tackles the gaps between racks and even across data centers. By making Ethernet smarter over long distances, Nvidia aims to help distributed GPUs act as a single system. Both efforts show how quickly AI infrastructure is being reworked, from the hardware stack to the network that connects it.