FuriosaAI, a Seoul, South Korea-based AI semiconductor company, has raised $620 million (KRW 820 billion) in a Series D funding round, bringing its post-money valuation to approximately $2.9 billion.
South Korea’s National Growth Fund was the largest investor in the round, contributing approximately $280 million. International investors committed a combined $110 million, with Infinitum serving as the largest international investor. Other global participants included QUDA Capital Partners, Cisco Investments, NeoFort Holding, Jericho Capital and Worldline Capital.
Existing investors Korea Development Bank (KDB), NAVER Ventures, DSC Investment, KDB Capital, Nextrans and Kakao Investment also participated. The Export-Import Bank of Korea, UAMCO and Hanwha Asset Management joined as new investors.
Founded in 2017 and led by co-founder and CEO June Paik, FuriosaAI develops AI accelerators designed to run artificial intelligence models in data centers. Its technology combines semiconductor hardware, compilers and software through a proprietary Tensor Contraction Processor (TCP) architecture.
The company’s second-generation accelerator, RNGD, is designed for AI inference workloads, including large language models and enterprise AI applications. The 180-watt, air-cooled PCIe accelerator is supported by an integrated software stack intended to simplify deployment in existing data center infrastructure.
FuriosaAI began mass production of RNGD in January 2026. The company plans to produce 20,000 accelerator cards during 2026 and more than double its production capacity in 2027.
The company has also expanded commercial deployments, with Samsung SDS introducing an NPU-as-a-Service offering powered by RNGD and LG Group companies deploying the accelerator in their on-premises hybrid cloud environments.
FuriosaAI intends to use the new capital to increase RNGD manufacturing capacity, expand global deployments, strengthen its software ecosystem and accelerate development of its third-generation AI accelerator.
The next-generation accelerator is being developed with Broadcom and other partners. It is planned to use a 2-nanometer, multi-die design with HBM4/4E memory, extending the company’s existing TCP architecture to larger-scale AI inference workloads.
