FuriosaAI has closed a $620 million Series D, about KRW 820 billion, at a post-money valuation of roughly $2.9 billion. Its last round valued the Seoul company at around $735 million. So the price went up about four times in a year.
That kind of jump doesn’t come from slides. It comes from a chip that’s actually shipping.
Who put money in
South Korea’s National Growth Fund was the largest investor with $280 million, part of the government’s push to build domestic AI compute and a local AI chip industry. International investors added $110 million. Infinitum, a US firm that focuses on AI semiconductors, was the biggest of them. Jericho Capital, Worldline Capital, Cisco Investments, Abu Dhabi’s NeoFort Holding and Singapore’s QUDA Capital Partners also came in. QUDA’s founder, Charlie Foo, ran Asia Pacific enterprise business at NVIDIA and Mellanox and now advises FuriosaAI.
Existing shareholders put in another $125 million. All three seed backers, NAVER, DSC Investment and KDB Capital, joined again; DSC has been in every round since 2017. Korea Development Bank added $22 million. Nextrans and Kakao Investment raised their stakes. New Korean money came from UAMCO with $30 million, Hanwha Asset Management, and the Export-Import Bank of Korea, which picked FuriosaAI for its first ever direct investment in a startup.
What RNGD is
RNGD is FuriosaAI’s second-generation accelerator, built only for inference in data centers. It went into mass production at TSMC in January. The company plans 20,000 units this year and more than 40,000 in 2027.
The design is called a Tensor Contraction Processor. GPUs chop model math into fixed two-dimensional matrix multiplications. FuriosaAI makes tensor contraction, a higher-dimensional operation, the basic unit of work, and runs it deterministically on reconfigurable compute slices. The point is to move less data around, because moving data is where much of the power goes. The card is air-cooled and draws 180 watts. FuriosaAI says it matches the throughput of GPUs that draw twice as much or more.
That matters more than raw speed for a lot of buyers. Plenty of enterprise data centers can’t get more power or liquid cooling into the building. A card that fits the racks they already have is an easier sale than a faster one that doesn’t.
On software, developers can pull models from Hugging Face, write them in PyTorch and use a drop-in replacement for vLLM, with a lower-level language and virtual ISA for anyone who wants to tune further.
Customers, not pilots
LG Group runs its EXAONE models on RNGD. Upstage uses it to serve live search summaries to millions of users. Samsung SDS launched a commercial NPU-as-a-service product on RNGD in July. Those are production deployments, and investors in the round pointed to them directly. Few chip startups get from design to volume production at all.
FuriosaAI famously turned down an acquisition offer of about $800 million from Meta. Selling to everyone instead of one hyperscaler looked like a risky call then. This round makes it look better.
Where the money goes
The company will expand RNGD output, build out its software stack and push its third-generation accelerator, which it’s developing with Broadcom for hyperscale infrastructure. It has offices in the US, Portugal and Singapore, and its new US backers are pitched as help for the North American push. An IPO has been talked about for as early as 2027.
The inference chip field is crowded and well funded, with Groq, Cerebras and Etched all raising large sums. FuriosaAI’s angle is power efficiency inside existing data center limits, and a state-backed home market that wants a national champion. Volume, not valuation, will decide whether that holds.