The Jericho3-AI from Broadcom Inc. - 5 nm Ethernet switch targets AI clusters

The Jericho3-AI switch from Broadcom delivers massive Ethernet bandwidth for GPU-heavy AI clusters with a 5 nm design and deep-buffer architecture. This bestseller drives the price of Broadcom shares (ISIN US11135F1012).

Reviewed: ad hoc news Lifestyle & Consumer desk. Edited and checked on 2026-06-26, 15:53. Details in the imprint.

The Jericho3-AI switch from Broadcom sits in a humming rack, its ports blinking like a quiet city at night while fans push a steady stream of warm air past your hands. It is not a gadget for the living room, but the backbone for giant AI clusters that crunch models all day long.

What Jericho3-AI is built for

Broadcom positions the Jericho3-AI as a high-radix Ethernet switch specifically tuned for large-scale GPU clusters in modern AI data centers. Each port links to accelerators or leaf switches, so one chassis can sit at the core of a training fabric where thousands of GPUs exchange gradients every second. When you stand near such hardware, what you mostly hear is a constant, even roar of airflow and the occasional click from technicians swapping cables.

The chip uses a 5 nm process node and deep-buffer memory to keep packets flowing smoothly even when bursts from many GPUs arrive at once. That buffer depth is crucial in AI training runs, where synchronized workers can generate congestion if the network cannot absorb traffic peaks. For architects, the appeal is that Jericho3-AI promises predictable latency without forcing them to move away from familiar Ethernet tooling.

How Broadcom tunes the fabric

Gos Latif, a senior product director at Broadcom working on Jericho switches, describes the line as aiming for "lossless Ethernet" over large clusters, with features like intelligent congestion control and packet prioritization. In practice, that means the silicon constantly measures flows and nudges queues so important traffic reaches GPUs without getting stuck behind bulk transfers. It is a quiet, continuous optimization rather than something operators have to babysit on every rack.

Broadcom also highlights that Jericho3-AI supports massive scale-out topologies, allowing operators to wire tens of thousands of accelerators into one consistent fabric. Where older designs sometimes relied on proprietary links, here the company leans on Ethernet standards with extensions for AI workloads. For customers, that makes it easier to reuse existing monitoring tools and cabling practices, even as the physical clusters grow denser and heavier.

Specs that matter to operators

Jericho-class devices typically expose hundreds of high-speed ports and support data rates aligned with 400G and above, so operators can wire GPUs directly or via leaf switches without bottlenecking traffic. Broadcom documents emphasize scaling to multi-petabit aggregate bandwidth in one chassis, a figure that matters when daily training runs push terabytes of gradients back and forth. While you do not see those numbers in the rack, you feel them in how quickly new models finish.

The switch also integrates telemetry hooks so network engineers can track congestion, dropped packets and latency tails across the fabric. In a control room, that data appears as sharp graphs on wall-mounted displays, where spikes translate to troubleshooting tasks. For teams working around the clock on AI services, those graphs can be more tangible than the chip itself.

Ethernet versus proprietary fabrics

Broadcom argues that AI clusters can lean on Ethernet instead of proprietary interconnects if the switches and software are tuned carefully, and Jericho3-AI is one of the pieces in that argument. The company points to improved congestion management and queue handling as reasons why operators can stay within the Ethernet ecosystem and still reach low-latency targets. For customers wary of lock-in, that stance is a practical selling point.

Compared with custom fabrics, Ethernet offers a broad base of compatible tools, from analyzers to cables and optics. Jericho3-AI sits within that ecosystem while pushing bandwidth per port upward, which lets designers keep familiar building blocks while scaling cluster size. In a sense, the switch promises a cleaner upgrade path: swap hardware, keep most of the operational playbook.

Where it fits in Broadcom's stack

Jericho3-AI does not live alone; it is part of a portfolio that also includes Tomahawk switches for general data-center use and ASICs that end up in routers from other vendors. Hock Tan, Broadcom's long-time CEO, repeatedly frames networking as a core pillar next to custom silicon for hyperscale customers. The AI-focused Jericho parts feed into that narrative by giving the company a seat deeper inside generative AI infrastructure.

Broadcom supplies many of these chips as merchant silicon, meaning OEMs build boxes around them and sell full systems. For investors and technically inclined readers, that means Jericho3-AI may appear inside appliances with another brand on the bezel. The silicon story still belongs to Broadcom, even if the rack label does not mention the name.

AI demand and investor attention

Analysts covering Broadcom highlight AI networking and accelerators as key drivers behind the company's recent valuation climb, with some noting its role as a supplier for large cloud providers. In that picture, products like Jericho3-AI are not side projects but elements in revenue discussions. You can think of each blinking port as one small contributor to a much bigger AI line item on the income statement.

Broadcom shares (ISIN US11135F1012) trade in Frankfurt on Xetra, where the listing under AVGO gives European investors access to the US-based chip group. The Jericho3-AI line is one component among many in the broader networking and AI portfolio that those investors watch over the long term.

Key facts on Jericho3-AI

• Product: Jericho3-AI Ethernet switch ASIC
• Manufacturer: Broadcom Inc.
• Category: Lifestyle/Consumer - data-center networking silicon
• Launch: Announced as part of Broadcom's AI networking roadmap, mid-2020s
• RRP / Price: Not publicly listed, negotiated in OEM and hyperscaler contracts
• Availability: Sold globally through networking OEMs and direct to large data-center operators
• Target group: Cloud providers, telecom operators, enterprises building large AI GPU clusters
• Highlight / USP: Deep-buffer, 5 nm Ethernet switch tuned specifically for massive AI training fabrics

This article was AI-assisted and editorially reviewed. Product information without guarantee; prices and availability may change at short notice. No investment advice, no buy or sell recommendation. Stock-market transactions involve risks up to total loss.