According to TheRegister.com, HPE announced it will ship a turnkey, rack-scale AI system worldwide sometime in 2026, built on AMD’s Helios reference architecture. The system will be one of the first to market using AMD’s next-generation Instinct MI455X GPUs and 6th-gen Epyc “Venice” CPUs, packing 72 accelerators per rack for 2.9 exaFLOPS of performance. Key to the design is a novel networking approach using a purpose-built Juniper switch based on Broadcom’s Tomahawk 6 silicon to run the UALink protocol over standard Ethernet, creating a 102.4 Tbps fabric. HPE positions this as a direct rival to Nvidia’s DGX GB200 NVL72 systems, which reportedly cost nearly $3.5 million each, and is targeting cloud service providers as the primary customers.
The Ethernet Gamble
Here’s the thing that really stands out: they’re running UALink over Ethernet. UALink is the open consortium’s answer to Nvidia’s proprietary NVLink, meant for tying GPUs together at insane speeds. But instead of needing special, expensive UALink hardware, HPE and Broadcom are basically saying, “We can just make Ethernet do the job.” That’s a huge bet. If it works as promised, it could offer a more flexible, potentially cheaper path to that massive scale-up fabric needed for trillion-parameter models. But it’s also a technical gamble—can standard Ethernet, even supercharged Tomahawk 6 Ethernet, truly match the latency and performance of a purpose-built interconnect like NVLink? HPE’s Rami Rahim is touting the “100 percent open standard” angle to avoid vendor lock-in, which is a powerful message in an Nvidia-dominated market. We’ll have to wait until 2026 to see if the performance lives up to the promise.
Strategy and the Bigger Picture
So why is HPE doing this? Look, it’s a classic play for relevance in the AI infrastructure gold rush. They’re not trying to beat Nvidia at the chip game; they’re leveraging AMD’s silicon and Juniper’s networking (now under the HPE umbrella) to offer an integrated, rack-scale alternative. The timing is everything—2026 lines up with AMD’s next-gen chip launches and gives the market more time to crave alternatives as AI cluster sizes explode. By targeting cloud providers and “neoclouds” first, they’re going straight to the customers with the biggest budgets and the most acute pain from Nvidia‘s pricing and supply constraints. It’s a high-end, low-volume strategy initially. For companies building out serious AI capacity, having a credible second source for this tier of hardware is crucial, and that’s the wedge HPE is trying to drive. It’s also a smart showcase for their full-stack capabilities, from industrial panel PCs and servers all the way up to this mega-rack, positioning them as a full-spectrum infrastructure player.
The Open Rack Factor
Another interesting piece is the chassis itself. Basing it on Meta’s Open Rack Wide spec from the Open Compute Project is a nod to the hyperscaler crowd they’re chasing. It means modularity, liquid cooling support, and efficiency in power-constrained environments—all critical for the folks who will be buying dozens or hundreds of these racks. It basically says, “We’re building for the scale you operate at, with the operational principles you already use.” This isn’t a bespoke, one-off science project; it’s meant to be deployed at scale. But let’s be real: “open” at the rack level is one thing, and delivering seamless, high-performance software integration for AI workloads is another. The hardware is just the table stakes. The real battle in 2026 will be about the software stack and how easily this Helios rack can be slotted into existing AI pipelines. Can it run PyTorch or JAX workloads as smoothly as the incumbent? That’s the billion-dollar question.
