AI build-out puts Broadcom Jericho3-AI switch at the center of cloud data centers

Edited by ad hoc news New Releases & Launches Desk. Reviewed before publication on 06/16/2026 at 2:24 AM ET. Details in the imprint.

Broadcom’s Jericho3-AI Ethernet switch chip is emerging as one of the most closely watched pieces of silicon in the race to build out AI data center infrastructure, promising cluster sizes up to tens of thousands of GPUs over standard Ethernet fabrics. The chip targets AI training and inference networks that need ultra-low congestion and high utilization, positioning Broadcom as a central supplier for hyperscale cloud operators that prefer Ethernet over proprietary interconnects. According to Broadcom, Jericho3-AI enables large-scale AI clusters using distributed, lossless Ethernet with dedicated features to optimize traffic for GPU workloads on the official product page.

What Broadcom’s Jericho3-AI switch is built to do

Jericho3-AI is part of Broadcom’s StrataDNX family and is specifically designed for AI and high-performance computing networks, combining deep buffers, advanced congestion management and programmability to support large-scale training clusters. The chip is built to sit in leaf and spine switches that link together racks of GPU servers, where it must keep thousands of high-bandwidth flows balanced to avoid idle accelerators and wasted power in multimillion-dollar clusters. Industry coverage notes that Broadcom is pitching Jericho3-AI as a way for cloud providers to run AI workloads over Ethernet while approaching the performance typically associated with specialized fabrics, a key argument as operators weigh cost and interoperability against raw speed as reported by The Information.

The architecture centers on extremely deep packet buffers and traffic-management algorithms that are tuned to the bursty, all-to-all communication patterns of AI training jobs rather than classic enterprise traffic. Ethernet switches used for web and storage traffic can struggle when thousands of GPUs synchronize gradients at once, because synchronized barriers can amplify small congestion events into large slowdowns across an entire cluster. Jericho3-AI addresses this by combining massive on-chip and external memory buffers with advanced scheduling, so it can absorb transient bursts without dropping packets and can prioritize flows in ways that keep GPU utilization high even under heavy load. For hyperscalers, higher utilization translates directly into better return on capital for expensive accelerators and power infrastructure, turning the switch silicon into a lever for both performance and economics.

Broadcom positions Jericho3-AI as supporting clusters that can scale to tens of thousands of accelerators built with standard Ethernet optics and cabling rather than requiring proprietary networking cards or exotic topologies. That matters because many cloud operators already run very large Ethernet estates and have operational tooling, staff skills and procurement channels optimized around Ethernet rather than alternative fabrics. By staying within the Ethernet ecosystem, they can reuse optical modules, switches and management frameworks while still targeting AI-scale performance, reducing both up-front integration complexity and long-term operating costs compared with adopting a separate interconnect technology just for GPU clusters. This strategy also gives Broadcom a channel to sell complementary products such as Tomahawk and Ramon switch chips and the company’s own merchant optics into the same racks.

Another design emphasis is on programmability and telemetry so that operators can monitor, debug and fine-tune AI fabrics in production. Large AI clusters are sensitive to subtle network pathologies, from misconfigured queues to hot spots around particular racks, which can cut effective throughput even when headline bandwidth looks sufficient. Jericho3-AI supports granular flow monitoring and telemetry hooks that expose congestion, latency and link-level behavior, allowing network teams to feed that data into automated controllers and software-defined networking stacks. This tight observability loop is increasingly critical as clouds roll out ever-larger training runs with strict service-level goals and limited maintenance windows, and it aligns with Broadcom’s broader push into integrated hardware-plus-software networking platforms for large operators.

While Broadcom does not break out Jericho3-AI revenue line-by-line, AI networking is one of the company’s fastest-growing segments as hyperscale data centers upgrade both compute and fabric to handle larger models and rising user demand. Recent market commentary around Broadcom has repeatedly highlighted the company’s role as a key supplier of AI accelerators and networking chips, with Ethernet switching a central pillar of that story as cloud providers design their next-generation clusters according to TradingKey’s coverage of Broadcom’s AI segment. Shares of Broadcom (US11135F1012) trade on NASDAQ in US dollars, and recent analyst discussions of the stock have focused heavily on the growth outlook for AI-related infrastructure revenue rather than legacy semiconductor lines.

This article was a.i.-assisted and editorially reviewed. Product information without warranty; prices and availability may change at short notice. Not investment advice and not a buy or sell recommendation. Trading involves risk up to and including the total loss of invested capital.