South Korean startup, which walked away from Meta's $800M acquisition bid, partnered with OpenAI to demonstrate the future of sustainable enterprise AI without GPUs (www.techradar.com)

FuriosaAI and OpenAI teamed up to demo a real-time chatbot in Seoul running the open-weight gpt-oss 120B model on two of Furiosa’s RNGD (“Renegade”) inference accelerators, showing large-scale enterprise models can run sustainably without GPUs. The system used MXFP4 precision to cut energy use while preserving enterprise-grade accuracy, and the end-to-end setup fit within standard data-center power budgets—illustrating that open-source LLMs need not rely on power-hungry GPU farms. Technically, Furiosa’s RNGD is a TSMC 5nm inference chip with dual HBM3 and a Tensor Contraction Processor architecture designed to maximize parallelism and eliminate unnecessary computation; it debuted at Hot Chips 2024. The demo, at OpenAI’s new Seoul office, positioned Furiosa as the sole hardware partner invited and underscores a broader shift: specialized accelerators can deliver cost, energy, and infrastructure advantages for growing model sizes. The showcase gains added weight because Furiosa—fresh off a $125M bridge round and an LG AI Research partnership—earlier rejected Meta’s reported $800M buyout offer, signaling an independent route to challenge GPU dominance in enterprise AI deployments.