HPE showed an open AMD Helios switch tray at OCP 2026, and it carries two Broadcom Tomahawk 6 chips rated at 102.6T each. The unit, built with HPE-Juniper as the QFX5252, sat open on the show floor. It's a double-wide tray with side-by-side switches and, unusually, very little bare internal PCB.

Why Tomahawk 6 and not Tomahawk Ultra?

The chip choice looks odd at first. You'd expect Broadcom's Tomahawk Ultra, which suits scale-up architectures. But the current-generation Ultra is a 112G PAM4 SerDes switch — 512 lanes of PAM4-112G. That's not enough for Helios, which runs 224G PAM4 SerDes. So the QFX5252 uses Tomahawk 6 instead.

Two big switch chips in a 1OU tray (one open-unit of rack height) generate enough heat that liquid cooling is mandatory. Internal liquid cooling manifolds distribute coolant to the various blocks.

What's with the retimers?

AMD and HPE-Juniper did not simply mount Tomahawk 6 chips on PCBs and call it a day. The cabled connections to the rear high-density fabric connectors — which mate with the Helios rack's cable cartridges — sit quite far from the switch chips. Some are very far.

The liquid cooling blocks between those cable connectors and the switch chips are apparently retimers (chips that regenerate signals to extend copper reach). AMD leans heavily on retimers. NVIDIA has worked to remove them from its designs.

Why so much copper?

The tray's cable bundle illustrates exactly why the industry will eventually move to co-packaged optics (CPO, optics integrated directly onto the switch package). In the QFX5252, retimers extend reach so the system can reliably scale up within the rack. That still requires an enormous array of copper cables. CPO would simplify this design considerably.

What does this change for rack design?

The tray differs from normal rackmount switches in one fundamental way: it's designed to scale up within a rack, not out through front-panel connectors. That drives everything else in the mechanical layout.

Notable specifics:

• The back panel is long and must blind-mate (connect without visual alignment) into the rack, so insertion force is significant. The cam levers on the tray are correspondingly very long.
• The front latch carries a large HPE QFX5252 label.
• Fans and airflow guides sit over heatsinks for other components. Unlike NVIDIA's push to remove fans from the AI rack entirely, AMD Helios still uses air cooling for some parts of the tray.

Final words

The QFX5252 previously appeared at HPE Discover 2026, but the OCP 2026 display marked the first time the assembled tray was shown on a show floor in this configuration.

One aside from the booth: Juniper co-founder and former CTO Pradeep Sindhu walked over and examined the switch tray alongside the HPE-Juniper team. His last appearance in this publication's coverage was a Fungible DPU-based storage cluster hands-on in 2020, during the pandemic.

A lot is going on in this tray, and almost none of it resembles a conventional rackmount switch. Between the 224G SerDes requirement, the retimer-heavy signal path, and the copper cable mass, it's a working argument for where rack-scale networking is heading — and for why CPO keeps coming up.