On May 28, 2026, FuriosaAI, a South Korean AI chip company, unveiled a partnership with Broadcom aimed at developing a third-generation AI inference accelerator. Samples of this cutting-edge chip are anticipated to be available in the first half of 2028. Leveraging a state-of-the-art 2nm manufacturing process, the chip will incorporate separate I/O dies and HBM4(E) memory stacks. Within a single rack, multiple chips will be interconnected in a fully connected topology via Broadcom's SUE technology, facilitating rack-level system deployments.

FuriosaAI's Tensor Contraction Processor (TCP) architecture is specifically tailored for AI computations, emphasizing high-bandwidth data transfer and extensive tensor operations. This collaboration will merge Furiosa's TCP architecture with Broadcom's XPU technology, IP platform, and Ethernet vertical scaling prowess, with the goal of overcoming significant hurdles in large-scale agent artificial intelligence.