Introduction

On May 13, 2025, Broadcom introduced Jericho3-AI, its next-generation Ethernet switch chip designed specifically for AI workloads in data centers. With 800G port support and 26 billion transistors, Jericho3-AI targets the rapidly growing need for scalable, low-latency interconnects in GPU and accelerator-heavy environments.

“AI isn’t just about compute—it’s about communication,” said Ram Velaga, SVP and GM of Core Switching at Broadcom.¹ “Jericho3-AI creates deterministic network behavior for massive AI clusters, something InfiniBand dominated until now.”

The chip features a new load-adaptive routing algorithm, sub-microsecond latency, and congestion management optimized for distributed AI model training. It supports RoCEv2 and is backward-compatible with existing Broadcom SDKs, simplifying integration into hyperscaler fabrics.

Why it matters now

• Ethernet is starting to challenge InfiniBand in large AI training clusters.
• Interconnects are the next bottleneck for trillion-parameter model scaling.
• Cloud providers need open, programmable switching platforms to match AI demand.

Call-out: AI is shifting the balance of power from compute to interconnect

Broadcom claims Jericho3-AI can handle over 100,000 GPUs in a single-layer fabric—eliminating the need for spine-leaf topologies in many AI scenarios.

Business implications

Network engineers and AI infrastructure architects should evaluate whether Ethernet-based fabrics can meet the latency and congestion thresholds for large-scale distributed training. Jericho3-AI could lower the cost and complexity of AI clusters, especially in co-lo or hybrid environments.

Switch and fabric design is also becoming strategic again: vendors that offer programmable telemetry, congestion visibility, and lossless transport will win enterprise and cloud AI budgets.

Looking ahead

Meta, Microsoft, and Baidu are reportedly testing Jericho3-AI in pilot deployments. Nvidia is expected to counter with its own Ethernet roadmap via Mellanox. Meanwhile, Broadcom plans to release AI-optimized NICs to complement the switch platform later this year.

Gartner forecasts that by 2030, Ethernet will dominate 70% of AI training interconnect deployments, up from just 20% in 2024.

The upshot: Jericho3-AI represents the most serious challenge yet to InfiniBand’s grip on AI networking. As compute density rises, the real disruption may lie not in the silicon crunching the numbers—but in the fabric connecting them all.

––––––––––––––––––––––––––––
¹ Ram Velaga, Broadcom AI Switching Briefing, May 13, 2025.