FuriosaAI raises $620M at a $2.9B valuation, four times its last round

South Korean AI chip startup FuriosaAI has closed a $620 million Series D round (KRW 820 billion) at a post-money valuation of about $2.9 billion, according to its Business Wire release. CEO June Paik also discussed the round in a Bloomberg TV interview, which put the valuation at four times its previous level.

Who put up the money

The largest check, $280 million, came from South Korea’s National Growth Fund. Existing shareholders added $125 million in follow-on investments, including $22 million from KDB. International investors contributed $110 million in total. The largest of them is the U.S. firm Infinitum, joined by QUDA Capital Partners of Singapore, Cisco Investments, NeoFort Holding of Abu Dhabi, Jericho Capital and Worldline Capital.

Seed backers DSC Investment, NAVER Ventures and KDB Capital took part again. So did Nextrans and Kakao Investment, which joined the Series C bridge last year. UAMCO invested $30 million, and Hanwha Asset Management also participated. The Export-Import Bank of Korea made what the company describes as its first direct investment in a startup.

FuriosaAI says the proceeds will expand RNGD production, strengthen its software ecosystem and fund its third-generation accelerator, which it is developing with Broadcom for hyperscale infrastructure.

RNGD is already shipping, and volumes are set

RNGD, the company’s second-generation inference accelerator, has been in mass production since January 2026. FuriosaAI describes it as a 180-watt, air-cooled card that uses high-bandwidth memory and delivers throughput comparable to a GPU drawing twice the power or more. That comparison is the company’s own and was not independently verified in the sources we read.

The company plans to produce 20,000 RNGD units in 2026 and more than 40,000 in 2027. On the software side, models can be loaded from Hugging Face, defined in PyTorch and served through a drop-in replacement for vLLM. Developers who want finer control can use the company’s Tensor Contraction Language and a low-level virtual instruction set.

The release names three deployments. LG Group runs its EXAONE models on RNGD, Upstage uses it for live search summarization serving millions of users, and Samsung SDS launched a commercial NPU-as-a-service built on the chip in July. Infinitum’s chief investment officer, John Yetimoglu, pointed to the Samsung SDS and LG Group deployments as evidence that the company has moved past the design stage.

Open questions behind the headline valuation

Several things remain unclear. The release gives no revenue, order backlog or per-chip pricing, so the $2.9 billion figure rests on deployments and volume targets rather than disclosed sales. The performance claim has no named benchmark or workload attached. And 20,000 units is small next to the GPU volumes that hyperscalers buy: Nvidia alone just put $3 billion into Meta’s stock sale as that company’s agent work deepens, and lenders are now syndicating a $3.1 billion Nvidia GPU loan to the wider market.

The size of the round also moved over the year. Bloomberg reported in January that the company was planning a $500 million raise ahead of an IPO, and the $620 million that closed is above that figure. According to the Techmeme summary of Bloomberg’s reporting, the previous valuation was $735 million in 2025, which is where the fourfold jump comes from. The release does not say when an IPO might happen.

The Gen 3 chip is the other unknown. The release says it is being developed with Broadcom for hyperscale infrastructure but gives no specifications or timing. Whether FuriosaAI can win hyperscale buyers, rather than Korean enterprise customers, will depend on that chip and on whether its software keeps pace with the CUDA ecosystem that competing hardware has to match. Naver Ventures, quoted in the release, sees North America as a key growth market.

For now, the round gives one of the better-funded non-Nvidia inference chip makers a large amount of domestic state backing, a strategic investor in Cisco and a production plan that can be checked against shipments next year.