FuriosaAI raised $620M at a $2.9B valuation, the Seoul-based inference chip maker said Wednesday.
The round was led by South Korea’s National Growth Fund, which put in $280M under a government program aimed at building up the country’s own chip industry. Cisco Investments, Abu Dhabi’s NeoFort Holdings, QUDA Capital Partners, Jericho Capital and Worldline Capital added $110M. Korea Development Bank, Kakao Investment and NAVER came back from earlier rounds.
The company is nine years old, and its argument is that the constraint on AI data centers has moved. Chips are not scarce; the electricity to run them is. Grids and cooling systems are stretched close to their limits, which is why accelerators tuned for tokens per watt look attractive to operators who have racks but no spare megawatts.
The design detail matters. Nvidia’s GPUs process with a rigid two-dimensional matrix, while Furiosa’s Tensor Contraction Processor treats tensor contraction as the fundamental unit of computation and runs it deterministically, which keeps power-hungry data movement low. Power draw comes in at less than half that of an Nvidia GPU by the company’s account, with support for common tooling including PyTorch and vLLM.
The customer list is domestic so far. LG Group runs its EXAONE models on the silicon, Upstage uses it for live search summaries, and Samsung SDS sells access through a commercial neural processing unit service. Furiosa plans 20,000 chips this year and 40,000 in 2027, and is co-developing a third generation with Broadcom.
A state fund putting up nearly half a round is also a signal about where Seoul wants this industry to land.
