Broadcom ships new AI chip to connect big data centers

Broadcom has begun shipping its Jericho4 ethernet fabric router, designed to connect large clusters of AI accelerators across multiple data centers.

The company said the Jericho4 is built to scale over one million XPUs, processors used in AI and high-performance computing, beyond the limits of a single facility.

The system offers features such as deep buffering, MACsec encryption, and RoCE transport over distances exceeding 100 kilometers.

It is manufactured using a 3nm process and supports Broadcom’s 3.2T HyperPort technology, which combines several high-speed links into a single logical port.

Broadcom said the router is compliant with Ultra Ethernet Consortium standards, allowing interoperability with other network devices that follow the same specifications.

Jericho4 is currently available in sample form to customers.

🔗 Source: Broadcom

🧠 Food for thought

1️⃣ Power constraints are forcing AI infrastructure beyond single data centers

Broadcom’s Jericho4 addresses a fundamental scaling challenge that’s reshaping how companies build AI infrastructure.

The shift from centralized to distributed computing architectures has become necessary as AI workloads demand increasingly high performance in bandwidth and latency that single facilities cannot provide 1.

Modern AI clusters are now outgrowing the power and physical footprint limitations of individual data center buildings, forcing cloud providers to distribute these systems across multiple locations and diverse geographies 1.

This infrastructure transformation explains why Jericho4’s ability to maintain lossless connectivity over 100+ kilometer distances has become commercially critical. It enables companies to treat multiple data centers as a single, unified computing resource.

2️⃣ AI infrastructure capabilities are becoming a source of global competitive advantage

The technology behind products like Jericho4 sits at the center of an emerging global divide in AI capabilities.

Currently, only 32 countries worldwide host specialized AI data centers, with the United States and China controlling over 90% of global data center infrastructure 2.