FuriosaAI, a South Korean designer of chips that run artificial intelligence models, has raised $620 million in a Series D round that values the company at $2.9 billion after the investment, giving Asia one of its best-funded challengers to Nvidia in the market for AI inference.

The Seoul-based company said on Oct. 28 that the money will go to scaling production of its RNGD accelerator and to building a third-generation chip with Broadcom. For investors in Asia, the round is a test of whether the region's AI build-out can support home-grown silicon alongside the Nvidia graphics processors that dominate data centers today, and of how far Seoul is willing to use state capital to back that bet.

National Growth Fund leads Korea's bet on FuriosaAI

Korea's National Growth Fund committed 370 billion won (about $280 million), FuriosaAI said. SiliconANGLE reported that the fund led the round as part of a government-backed plan to build up the country's domestic chip industry.

Foreign investors joined too. Infinitum, Jericho Capital, QUDA Capital, Cisco Investments, NeoFort Holding and Worldline Capital took part, according to the company. SiliconANGLE, which described NeoFort as Abu Dhabi-based, put the global investors' combined contribution at about $110 million. The Export-Import Bank of Korea, UAMCO and Hanwha Asset Management came in as new investors. All three original seed backers returned: NAVER through NAVER Ventures, DSC Investment and KDB Capital, along with Korea Development Bank and Kakao Investment, among others.

The mix matters. With public money supplying close to half the round, the deal ties FuriosaAI's fortunes to Korea's industrial policy at a moment when the country's memory makers are booking record profits from AI demand. SK Hynix this quarter posted a record 83 trillion won operating profit, and Samsung Electronics said its HBM4 sales tripled. A domestic accelerator maker would let Korea capture more of the value in each AI server, not only the memory.

John Yetimoglu, chief investment officer at Infinitum, said in the company's statement that few companies have shown they can take an advanced AI chip from design to volume production.

Power, not chips, is the pitch

FuriosaAI is selling efficiency. Its RNGD card draws 180 watts and is air-cooled, the company said. Eight cards make up a server that uses 3 kilowatts, offers 4 petaflops of FP8 compute and carries 384 gigabytes of HBM3 memory. The company said the server is built to slot into the racks and power budgets that enterprises already have, without a cooling retrofit, and that it serves twice as many users per kilowatt as an Nvidia RTX PRO 6000 setup at the same card count.

That argument lands in Asia, where many new data center projects are competing for scarce grid power. SiliconANGLE framed the round around the same point: grids and cooling systems are stretched, so chips that produce more tokens per watt have room to compete even against far larger rivals.

The chip is designed only for inference, the work of running trained models, rather than for training them. RNGD entered mass production in January 2026, according to the company. Its named customers are mostly Korean. Samsung SDS offers access to RNGD through its NPU-as-a-service product, and LG Group companies use the chips in their own hybrid cloud setups. SiliconANGLE reported that LG uses them to run its EXAONE models and that Korean startup Upstage uses them for live search summaries. The first large deal abroad is in Europe: in August the company said it would supply more than 8,800 RNGD chips to a 15-megawatt data center being built in Stockholm with I/ONX HPC and Velox, starting with 1,800 chips, The Korea Herald reported.

FuriosaAI production plans and the Broadcom chip

The company plans to produce 20,000 RNGD cards this year and said it has secured memory and foundry capacity to more than double that in 2027. In a market where advanced packaging and high-bandwidth memory are tightly allocated, that supply commitment may matter as much as the cash.

Its third-generation chip, under development with Broadcom and other partners, will use a 2-nanometer, multi-die design with HBM4 or HBM4E memory and is aimed at large clusters running frontier agentic models, FuriosaAI said. That puts it in the same supply chain as the hyperscalers' own custom chips, drawing on leading-edge foundry capacity and the latest generation of Korean memory.

Co-founder and Chief Executive June Paik said the company is “entering a new phase” of bringing its technology to the global AI infrastructure market at scale, according to SiliconANGLE.

Southeast Asia is next on the sales map. In September the company set up FuriosaAI Singapore Pte. Ltd. as its hub for business development, sales and technical support across the Asia-Pacific region, led by Addison Chi, vice president of APAC sales. The company also has offices in Silicon Valley and Lisbon, and NAVER Ventures said it would focus on helping FuriosaAI grow in North America.

The scale gap with Nvidia remains wide. FuriosaAI's 20,000 cards this year compare with hyperscaler capital budgets now measured in tens of billions of dollars a quarter. The figure to watch is whether the 2027 target of more than 40,000 cards turns into more named customers outside Korea, starting with the Singapore office's first regional deals.