FuriosaAI just raised $620 million to prove that the next AI infrastructure war is fought on the power bill, not the spec sheet.
The Series D values the South Korean startup at $2.9 billion. South Korea's National Growth Fund led with $280 million as part of a government-backed push to build domestic semiconductor capability. Investors including Cisco Investments, Abu Dhabi's NeoFort Holdings, QUDA Capital Partners, Jericho Capital and Worldline Capital added $110 million, with existing backers including Korea Development Bank, Kakao Investment and NAVER covering the rest.
What the money is buying
The funds go toward scaling production of FuriosaAI's RNGD chip, built on its Tensor Contraction Processor architecture. The company plans to produce 20,000 chips this year and 40,000 in 2027, with a third-generation chip already in co-development with Broadcom.
The core claim is simple: Nvidia's chips are power-hungry. FuriosaAI's chips are built for one job—running AI inference—and they do it at less than half the power draw. That gap matters now because the binding constraint in AI data centers has shifted. Energy grids and cooling systems are at their limits. Adding more Nvidia GPUs does not solve a power problem; a chip that uses less than half the power can help.
The commercial traction is real. LG Group is running its EXAONE models on FuriosaAI chips. Upstage is powering live search summaries for millions of users on them. Samsung SDS sells access to the silicon as a commercial cloud service. NAVER Ventures, which has backed the company since its pre-seed in 2017, says FuriosaAI has cleared the hardest hurdle in semiconductors: turning architecture into proven, shipping hardware. Full details on the round are in the SiliconAngle report.
The broader shift
This is the same bet Groq has been making with its inference chip. It rhymes with Cerebras and Etched, both of which are building inference silicon as an alternative to the GPU. And it fits the pattern of sovereign capital—South Korea here—treating domestic chip capability as strategic infrastructure.
The common thread across all of it: a bet that raw compute is no longer the scarce resource in AI, and energy is. Investors are now pricing that in, and operators still planning their infrastructure around GPU availability alone are planning around the wrong constraint.
