OCP Global Summit 2025: Broadcom on Networking for AI Scaling
Introduction
In its OCP Global Summit 2025 talk, Broadcom focused on the networking foundation for the AI scaling era. With a single company announcing more than 26 GW of data center projects in one go, infrastructure demand for AI training and inference has reached an unprecedented scale. Broadcom's core thesis: only Ethernet can support scale-up, scale-out and cross-data-center scale-across at the same time.
Key Takeaways
Broadcom opened with the current landscape: one company announced 26 GW of data center build-outs in just 4 weeks, equivalent to roughly 15 million XPUs (GPUs, TPUs or custom accelerators). AI compute is inherently distributed, so hundreds of thousands of accelerators must be stitched together over the network.
Why Ethernet Is the Only Option
Three years ago at OCP, Broadcom declared that "Ethernet is the only choice for AI networking." Compared with proprietary protocols, Ethernet offers openness, flexibility and cost efficiency, and spans every scale-up, scale-out and scale-across scenario.
Scale-up: High-Bandwidth Demand Across XPUs
As HBM bandwidth climbs from 40TB/s to 100TB/s, XPU-to-XPU interconnects need extremely high bandwidth and low latency. Broadcom's Ethernet for Scale-up Networking (ESUN) lets heterogeneous XPU environments compete freely instead of being locked into proprietary protocols.
Scale-out: Fewer Network Tiers, Lower Cost
Broadcom showed a comparison: in a 128K-GPU cluster design, 100Tb/s switches reduce the number of network tiers and optical modules, cutting both latency and cost. For mega-clusters costing billions of dollars, these design differences directly affect return on investment.
The Role of Optics and CPO
As bandwidth doubles, optics becomes critical. Broadcom is now on its third generation of CPO (Co-Packaged Optics), and stressed that CPO reliability is even better than pluggable optics.
Ultra Ethernet and a New-Generation NIC
Broadcom announced Thor Ultra — the world's first true 800G NIC, fully UEC-compliant and designed for enhanced RDMA. Its headline feature is being "truly open": it works with any switch, cable or XPU.
Conclusion
Broadcom's core message was clear: The network is the computer, and Ethernet is the only way to scale AI infrastructure. Ethernet is not only the most cost-effective choice but also the only standard able to support future AI scaling.
STT Perspective
Technology impact: Progress in CPO shows that integrating Ethernet with silicon photonics is the necessary path to break through bandwidth and power bottlenecks.
Supply chain: Broadcom's Thor Ultra NIC, Marvell's DSPs and Credo's AEC/LRO form competition at different layers, and silicon photonics module makers such as Lumentum and Innolight also stand to benefit.
Market trend: In a 100K-GPU cluster, networking accounts for less than 3% of total CAPEX yet has an outsized impact on performance and payback period, making Ethernet a must-have infrastructure investment.
Comments