Why Broadcom’s Jericho3-AI keeps showing up in big AI clusters

When hyperscalers quietly order thousands of switches, Broadcom’s Jericho3-AI is usually somewhere in the mix. The 800G merchant switch silicon targets massive AI fabrics with deterministic performance, high radix, and congestion control that aims to keep GPUs busy instead of waiting.

Reviewed: ad hoc news Software & Services desk. Edited and checked on 2026-06-18, 01:00. Details in the imprint.

Broadcom’s Jericho3-AI switch chip is one of those components you never see in an AI data center tour, yet it quietly decides whether thousands of GPUs are crunching numbers or just waiting on the network. This 5 nm 800G device is designed to stitch together truly enormous AI clusters with predictable latency and brutal bandwidth efficiency.

What Jericho3-AI is built to do

On paper, Jericho3-AI is a merchant switch silicon for AI and high-performance computing networks, but the intent is very specific: keep GPU clusters busy at scale. Broadcom positions the chip as capable of building fabrics that connect up to 32,000 GPUs with tightly controlled tail latency and minimal congestion.

The device supports 800G ports, which operators typically break out into multiple high-speed links to connect GPU nodes and spine layers. That bandwidth is paired with a deep packet buffer and telemetry features meant to absorb traffic bursts and react before queues get ugly, as Broadcom highlights in its AI networking material.

Inside the silicon, in everyday operation

In a running AI cluster, Jericho3-AI never shows up on a dashboard screenshot, yet its scheduling and congestion algorithms decide whether gradients and parameters flow smoothly between racks. Engineers describe a fabric built on chips like this as "deterministic" when the network behaves the same way even under load spikes.

That matters when training multi-trillion-parameter models, where a microburst on one link can expand into seconds of idle GPU time if packets are dropped or delayed. Jericho3-AI’s deep buffering and traffic management are designed to smooth those microbursts so that expensive accelerators do not spend their time waiting on straggler nodes.

How it differs from earlier generations

Compared with earlier Jericho devices used in routing and carrier networks, Jericho3-AI is tuned for lossless or near-lossless fabrics rather than best-effort internet traffic. The focus shifts from thousands of mixed flows to huge, coordinated AI workloads where collective operations dominate the pattern.

This shows up in support for features like precision congestion control and fabric telemetry, which help operators understand exactly where hot spots form in a training job. In practice, that should translate into more consistent step times across a training run and fewer mysterious performance cliffs when scaling from hundreds to thousands of GPUs.

Strengths that appeal to hyperscalers

Jericho3-AI’s biggest selling points are scale and predictability. Hyperscalers want to reuse existing operational knowledge, so a merchant switch that fits into familiar tooling, yet still drives 800G ports and AI-specific telemetry, has a clear appeal for large data center operators.

The chip also gives buyers leverage. Because it is not tied to a single GPU vendor, large cloud providers can mix and match accelerators and still lean on a proven networking base. For Broadcom, that merchant stance keeps Jericho3-AI in the conversation whenever a customer evaluates options for a new AI region.

Where the trade-offs show

For smaller operators, Jericho3-AI can feel like overkill. Building a fabric sized for thousands of GPUs demands serious capital, specialized optics, and teams that are comfortable debugging congestion issues at scale, which already narrows the customer pool to hyperscalers and top-tier enterprises.

There is also competition from tightly vertically integrated AI systems where the GPU vendor controls both accelerators and the switch silicon. Those platforms promise simplified deployment at the cost of some flexibility, so Broadcom has to win on performance, openness, and the comfort buyers have with its networking software stack.

Why software still makes or breaks it

Even with sophisticated silicon, AI networks live or die by their software stack. Jericho3-AI is typically paired with routing and telemetry software that needs to integrate into the customer’s fabric manager and AI orchestration tools, plus higher-level frameworks that schedule jobs and allocate GPUs.

Operators want to see clean hooks into their chosen AI frameworks and monitoring platforms, not another silo. When the plumbing works, the switch chips fade into the background and engineers judge the network on one metric: whether jobs finish on time with the expected efficiency.

Context and where the stock sits

Jericho3-AI underlines how far Broadcom has moved beyond its consumer legacy toward infrastructure that only a handful of customers ever touch directly, but that underpins much of the current AI build-out. The company’s fortunes now lean heavily on whether this kind of high-end networking silicon stays central to hyperscaler roadmaps.

Shares of Broadcom (US11135F1012) trade on Nasdaq under the ticker AVGO, giving investors a liquid way to participate in the broader demand for AI infrastructure hardware and software rather than any single GPU design.

Key facts on Broadcom Jericho3-AI

• Product: Jericho3-AI switch silicon
• Manufacturer: Broadcom Inc.
• Category: Software/Service/Subscription - AI networking platform component
• Launch: Announced as part of Broadcom’s AI networking portfolio in the current AI build-out phase
• RRP / Price: Not publicly disclosed, negotiated directly with hyperscale and enterprise customers
• Availability: Sold via Broadcom’s networking OEM and direct sales channels to hyperscalers and large data center operators worldwide
• Target group: Cloud providers, telecom operators, and enterprises building large-scale AI training and inference clusters
• Highlight / USP: 800G-capable switch designed for deterministic, large-scale AI fabrics connecting up to tens of thousands of GPUs with tight control over congestion and latency

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.