FuriosaAI and Broadcom team up for 2nm AI chip by 2028

South Korean AI chip company FuriosaAI announced a partnership with Broadcom to develop its third-generation AI inference accelerator, targeting sample production in the first half of 2028. The chip will use 2nm manufacturing, separate I/O die, and HBM4(E) memory stacks, with multi-chip connections via Broadcom's SUE Ethernet for rack-level systems.

FuriosaAI's TCP architecture is optimized for AI math, focusing on high-bandwidth data transfer and large-scale tensor operations rather than managing many small threads. Broadcom's semiconductor chief said inference performance now depends more on data reuse and communication efficiency, and the collaboration aims to solve key bottlenecks in large-scale AI agents.