Broadcom unveils Jericho3-AI Ethernet switch with 800G ports, enabling AI cluster interconnect at new bandwidth density.

Broadcom's Jericho3-AI Ethernet switch targets massive AI fabrics with up to 800G ports and purpose-built congestion control for GPU-heavy clusters. The silicon is designed to connect tens of thousands of GPUs with lossless-like behavior over Ethernet fabrics, competing directly with InfiniBand-based approaches.

Jericho3-AI focuses on congestion control and packet spraying techniques to ensure large AI training jobs see predictable tail latency even when the network is heavily utilized. David Li, a senior product director at Broadcom, described Jericho3-AI as a "purpose-built engine for AI Ethernet fabrics" in the launch material, underlining the narrow focus of the design.

At the hardware level, Jericho3-AI supports up to 800G Ethernet ports, enabling dense top-of-rack or spine switches that can feed modern accelerators such as H100 or custom ASICs without becoming the bottleneck. The silicon targets fabrics with more than 32,000 GPUs connected, pushing Ethernet deeper into territory once considered InfiniBand-only. The chip's power envelope remains within standard high-density switch design constraints, a critical factor when operators stack dozens of units in a single aisle.

Jericho3-AI arrives in a market split between traditional InfiniBand and newer cloud-scale Ethernet fabrics. Broadcom positions the chip as a way to deliver InfiniBand-like congestion management and fairness on Ethernet, which many hyperscalers already use throughout their data centers. The silicon implements fine-grained traffic steering that keeps short AI synchronization messages flowing smoothly alongside bulky parameter updates and data streams. For operators, this means less jitter in training times and more predictable utilization when scheduling jobs in busy clusters.

Jericho3-AI sits alongside Broadcom's standard Jericho3 switches and Tomahawk line, but its feature set is narrowly tuned for collective operations common in AI workloads. In launch material, vice president Ram Velaga emphasized that customers requested Ethernet-based alternatives that still delivered low tail latency and efficient job completion times on massive GPU farms—indicating the product responds to existing hyperscale demand rather than speculative need.

In practice, Jericho3-AI will not be purchased by small enterprises but by hyperscalers, major cloud providers, and large AI research institutions. These buyers typically integrate the chip into custom switch designs or work with OEM partners who produce systems tuned to their specific rack power and cooling profiles. Once deployed, operators interact with Jericho3-AI primarily through switch operating systems and fabric controllers, tuning parameters like queue depths, spraying policies, and congestion response thresholds to match specific training workflows.

Jericho3-AI reflects Broadcom's strategy to remain a central supplier for high-performance networking as AI spending shifts from pilot projects to large-scale infrastructure. For investors, the product underscores how much of Broadcom's growth thesis now rests on providing infrastructure plumbing for compute-intensive workloads.