FuriosaAI Raises $620 Million To Challenge NVIDIA and Solve Data Center Power Bottleneck

FuriosaAI, a Seoul-based chipmaker, has secured $620 million in funding to develop chips that can handle AI inference, a process that is increasingly becoming the bottleneck in data centers. The round, FuriosaAI's Series D, was led by South Korea's National Growth Fund and included international investors such as Infinitum and Cisco Investments. This brings the company's valuation to an impressive $2.9 billion, four times its previous valuation of $735 million.

FuriosaAI's flagship chip, the RNGD (Renegade), is designed for inference tasks rather than the training of AI models. It stands out with its Tensor Contraction Processor, which reduces the energy-intensive data shuffling typically handled by GPUs. FuriosaAI claims that its 180-watt, air-cooled RNGD chip matches the performance of GPUs that consume twice the power or more, offering a significant energy efficiency advantage.

The chip supports popular AI frameworks like PyTorch and Hugging Face models and can serve as a drop-in replacement for vLLM. It has already been implemented in large-scale deployments, including LG's EXAONE models and Samsung SDS's cloud service. The RNGD chip entered mass production in January, with TSMC manufacturing the chips.

FuriosaAI's success in this $620 million funding round is a testament to the growing demand for efficient AI inference chips. As data centers face increasing power constraints, FuriosaAI's solution could prove crucial for scaling AI applications without the need for additional power infrastructure. The company aims to build 20,000 RNGD chips in 2024 and over 40,000 in 2027, with a third-generation chip in development alongside Broadcom.

While NVIDIA continues to dominate the AI chip market, the growing power limitations could open the door for innovative players like FuriosaAI to make a significant impact.