FuriosaAI raises $620 million at $2.9 billion valuation

FuriosaAI announced a $620 million Series D at a $2.9 billion post-money valuation, giving the South Korean AI chipmaker new capital to expand production of its RNGD inference accelerator and develop its next chip. South Korea’s National Growth Fund made the round’s largest investment, committing KRW 370 billion, which FuriosaAI described as about $280 million.

Cisco Investments, Infinitum, Jericho Capital, NeoFort Holding, QUDA Capital and Worldline Capital joined the round. FuriosaAI said international investors contributed $110 million and that Infinitum was the largest international investor.

The company also named the Export-Import Bank of Korea, UAMCO and Hanwha Asset Management as new investors. Returning participants included NAVER through NAVER Ventures, DSC Investment, KDB Capital, KDB, Nextrans and Kakao Investment.

FuriosaAI said it will use the proceeds to expand RNGD production and global deployments, strengthen its software ecosystem and accelerate development of a third-generation accelerator. An inference accelerator is specialized hardware designed to run trained AI models.

The production and customer figures remain company claims. FuriosaAI said it plans to produce 20,000 RNGD accelerator cards in 2026 and has secured memory and fabrication capacity for more than 40,000 cards in 2027. It also said Samsung SDS launched an RNGD-powered NPU-as-a-Service offering and that LG Group companies are deploying RNGD in on-premises hybrid-cloud systems. The available evidence did not independently verify those capacity or deployment statements.

FuriosaAI describes RNGD as a 180-watt, air-cooled PCIe card. It says an eight-card NXT server draws 3 kilowatts, delivers 4 petaFLOPS at FP8 precision and includes 384 GB of HBM3 memory. The company also claims RNGD matches the throughput of Nvidia’s RTX PRO 6000 at equal card counts while serving twice as many users per kilowatt. Those performance comparisons were not independently validated.

The company separately claims its SDK 2026.4 software improves token generation per watt across 18 models compared with version 2026.3, including a 138% improvement for gpt-oss-120B. The available evidence included no third-party replication or independent benchmark methodology for that result.

For its next product, FuriosaAI says it is working with Broadcom and other partners on a third-generation accelerator planned as a 2-nanometer, multi-die design using HBM4 or HBM4E memory. Those specifications are roadmap statements, not features of a delivered product, and the sources provide no delivery date.

FuriosaAI previously said RNGD entered volume production in January 2026, when it reported receiving an initial 4,000 units from TSMC and ASUS. That figure is based on the company’s account rather than independently audited data. The latest financing follows Tenstorrent’s $1.5 billion Series E for manufacturing and deployments, another recent capital raise by an AI-chip developer.