Broadcom and FuriosaAI push Ethernet fabrics for AI inference

Broadcom and FuriosaAI's partnership is a strong infrastructure signal because inference scaling is starting to look different from training. Data Center Knowledge reports that the companies are developing a rack-scale inference platform using Furiosa's third-generation Tensor Contraction Processor architecture with Broadcom networking, advanced packaging, PCIe, interconnect technologies, and Ethernet fabrics. Furiosa says the architecture will use a 2nm compute die with dual-layer HBM4/4E memory and aims to scale inference clusters across thousands of nodes. The important direction is heterogeneous, Ethernet-based token-generation infrastructure as an alternative to proprietary GPU fabrics. Furiosa's earlier RNGD accelerator is already in mass production on TSMC 5nm and deployed by Samsung SDS and LG AI Research. Watch whether the next system can deliver enough performance-per-watt and software support to attract hyperscale inference workloads.

Key details: Broadcom, FuriosaAI, May 27, 2026, Ethernet AI fabrics, third-generation Tensor Contraction Processor, 2nm compute die, HBM4/4E, PCIe.

Continue swiping for more AI Brief stories.