Broadcom's custom ASIC biz adds South Korea's FuriosaAI to its empire - The Register

Broadcom has expanded its custom ASIC business to include South Korean startup FuriosaAI, aiming to co-develop high-performance AI inference chips using Broadcom's Broadzilla advanced packaging techno... Read original →

Broadcom’s integration of FuriosaAI isn’t just an acquisition—it’s a strategic pivot toward modular AI inference via advanced packaging. By leveraging 3.5D XDSiP to disaggregate compute, HBM4e, and I/O into chiplets, Broadcom slashes design complexity and yield risk at the 2nm node, directly challenging TSMC’s CoWoS dominance. Geopolitically, while FuriosaAI’s South Korean origin offers a partial hedge against U.S.-China tech decoupling, escalating export controls could still trigger secondary supply chain scrutiny. In response, NVIDIA may accelerate NVLink ecosystem openness, while AMD pushes MI300 with UALink to standardize heterogeneous interconnects. Over the next 18 months, HBM4e will become the decisive battleground for mid-power AI accelerators—control over memory-packaging co-design will determine who dominates inference in air-cooled and edge datacenters.