Broadcom’s Next-Gen AI Chip Revolutionizes Data Center Connectivity for Hyperscale AI Training

Broadcom has launched the Jericho4 chip, a cutting-edge AI networking solution purpose-built to transform the connectivity landscape for data centers handling hyperscale AI workloads. Officially shipping since early August 2025, Jericho4 enables ultra-fast, secure, and lossless Ethernet-based connectivity across multiple data centers, effectively addressing the growing demands of AI model training that exceed the capacity of single facilities. This launch marks a pivotal advancement in AI infrastructure, poised to accelerate the AI hardware race among hyperscalers and cloud giants. Jericho4’s standout capability is its ability to interconnect over one million XPUs — encompassing GPUs, TPUs, and other accelerators — spread across geographically distributed data centers up to 60 miles apart. This scalability supports broader AI workload distribution while maintaining unprecedented bandwidth and minimizing latency between GPU clusters, critical for training massive AI models with low communication overhead. The chip supports up to 3.2 Tbps HyperPort interconnect speeds, making it highly competitive with existing solutions like NVIDIA's NVLink scale-up technology but over Ethernet fabrics. Technical innovations embedded in Jericho4 include its support for advanced Ethernet protocols such as RoCEv2 and compliance with Ultra Ethernet Consortium standards, which enable packet loss prevention and ultra-low latency behavior. Additionally, Broadcom has integrated comprehensive security features, such as MACsec encryption on every port delivered at full line speed, safeguarding sensitive AI data transfers without sacrificing performance even under the heaviest network loads. Jericho4’s architecture complements Broadcom’s broader AI networking portfolio, including the Tomahawk 6 switch series, which offers up to 102.4 Tbps bandwidth and supports cluster scaling of over 100,000 XPUs in a more power-efficient two-tier switch design. The inclusion of co-packaged optics technology in these switches further reduces latency and cost by integrating optical components directly into the switch ASIC package. Beyond hardware specs, Jericho4 is designed to tackle practical challenges like network congestion and inter-data center traffic orchestration through advanced load balancing, dynamic congestion control, and real-time telemetry embedded in its Cognitive Routing 2.0 technology. These innovations position Broadcom at the forefront of enabling a new era of distributed AI computing, where training workloads no longer face physical data center size constraints. By bridging hundreds of megawatts of power and countless accelerator units across multiple locations, Jericho4 facilitates AI model training at scales previously impractical. For tech enthusiasts and industry observers, Broadcom’s Jericho4 represents a strategic inflection point: an infrastructure leap that not only meets but anticipates the intensifying compute, bandwidth, and security demands of next-generation AI. As hyperscalers race to build ever-larger models, this chip will be a critical enabler in the quest for faster, more efficient, and more secure AI data center operations.