• Networking

Prior to the results last week, Broadcom also announced its 51.2Tb/s Jericho 4 Ethernet-based scale across router chip, which can handle clusters beyond 200,000 compute nodes across multiple data centers. It is designed to connect data centers around 100km apart. This announcement follows on from previous announcements for Tomahawk 5 for scale-up and Tomahawk 6 for scale-out.

Scale across is a relatively new term coined for scaling between data centers, (i.e. it is another term for Data Center Interconnect or DCI). Clusters are typically limited to around 100,000 XPUs each due to power/real estate reasons and so hyperscalers are building individual clusters set around 100km apart. These data centers need to be networked together. The Jericho 4 chip has been designed to target this market.

A few months ago, Broadcom announced its Scale-Up Ethernet (SUE) standard which is designed for hyperscale data centers developing rackscale AI training clusters using Ethernet-based AI fabrics and enables scale-up and scale-out. Combining scale-up and scale-out in a 𝐬𝐢𝐧𝐠𝐥𝐞 networking technology will bring advantages as the same Ethernet toolchain (ops/telemetry/silicon optics roadmap) can be used for both. And using Ethernet reduces risk and vendor lock-in. It could therefore be as threat to Nvidia’s NVLink and Infiniband/Spectrum-X technologies.

• Non-AI Semiconductors

Although revenues at $4 billion were flat sequentially, demand is still slow to recover, except for broadband, which showed strong sequential growth, while enterprise networking and server storage were down sequentially. Broadcom expects low double digit growth sequentially for Q4 to $4.6 billion.

Analyst Viewpoint

Broadcom expects growth in FY2026 to accelerate compared to FY2025 and a major part of this growth will be from its AI semiconductors, particularly XPUs. The company looks set to continue gaining market share with its original three XPU customers as they transition to next-generation XPUs - plus now a fourth XPU customer for which the company just received a $10 billion production order!

Broadcom benefits from its XPU business in another way as many of its XPU contracts also bring networking wins as customers opt to buy its Ethernet-based switch silicon products. The company now offers a full suite of Ethernet-based switches: Tomahawk Ultra for scale-up, Tomahawk 6 for scale-out and now Jericho 3/4 router chips for scale across. In addition, it also offers the SUE networking technology, which combines scale-up and scale-out in a single Ethernet-based networking fabric.

Although SUE is not a direct replacement for NVLink, there is clearly a market opportunity here for Broadcom depending on the use case. According to Broadcom, SUE is designed for distributed AI training for applications with 𝐦𝐨𝐝𝐞𝐫𝐚𝐭𝐞𝐥𝐲 high interconnect bandwidth needs. Even with its various enhancements, SUE still has higher latency and lower bandwidth than Nvidia’s NVLink. As a result, whether it is a threat or not to NVLink really depends on the use case. For example, it could be a threat for AI training where ultra-tight GPU coupling is less important and where customers prefer open standards and want to reduce their dependency on Nvidia. However, for tightly-coupled workloads where AI workloads are deeply integrated with Nvidia’s CUDA, Nvidia will reign supreme. Similarly, it could also be a threat for Infiniband and Spectrum-X, particularly for non-Nvidia accelerators. But again, Nvidia will likely dominate for ultra-low latency, multi-mode training in CUDA-heavy clusters.