FuriosaAI is a South Korean semiconductor startup founded in 2017 by engineers with backgrounds from Samsung, AMD, and Qualcomm. The company specializes in developing high-performance, energy-efficient AI accelerators designed to meet the demands of large language models (LLMs), multimodal applications, and data center environments.
At the heart of FuriosaAI's offerings is its second-generation accelerator, RNGD (pronounced "Renegade"). Built on a custom Tensor Contraction Processor (TCP) architecture, RNGD departs from traditional matrix multiplication-based designs, enabling more efficient tensor contraction operations. This innovation allows for significant performance improvements in AI inference tasks. The RNGD chip boasts impressive specifications, including 512 TFLOPS of FP8 performance, 48GB of HBM3 memory, and a thermal design power (TDP) of just 180W, making it suitable for air-cooled data centers.
FuriosaAI's software ecosystem complements its hardware, offering tools like the Furiosa SDK, which includes compilers, quantizers, and runtime environments optimized for their NPUs. The platform supports popular AI frameworks such as PyTorch and facilitates seamless deployment through Kubernetes and cloud-native integrations.
In 2025, FuriosaAI secured a significant partnership with LG AI Research, supplying RNGD chips for LG's EXAONE platform. This collaboration aims to enhance AI inference performance across various sectors, including electronics, finance, telecommunications, and biotechnology. Notably, FuriosaAI declined an $800 million acquisition offer from Meta, choosing instead to expand its presence in the U.S., Middle East, and Southeast Asia.
With a growing team of over 140 employees and strategic partnerships with industry leaders like TSMC, ASUS, and SK Hynix, FuriosaAI is poised to challenge established players in the AI hardware market.
