South Korean inference chip developer FuriosaAI has closed a $620 million funding round that values the company at $2.9 billion, according to SiliconANGLE. The money is earmarked for scaling production of its latest processors.
South Korea's National Growth Fund led the round with $280 million, part of a government-backed effort to strengthen the country's domestic semiconductor industry. A group of international investors, including Cisco Investments, NeoFort Holdings, QUDA Capital Partners, Jericho Capital and Worldline Capital, contributed another $110 million. Existing backers Korea Development Bank, Kako Investment and NAVER supplied the remainder.
FuriosaAI and its investors argue that the limiting factor in AI data centers is shifting from available compute to available electricity. Grids and cooling systems are under strain, and the company positions its chips, designed to maximize tokens per watt, as a response.
The current product is the Tensor Contraction Processor, the second generation of inference chips on the RNGD architecture. Rather than building around the two-dimensional matrix operations typical of GPUs, it treats tensor contraction as the basic unit of computation and uses deterministic execution to limit data movement. The report says this lets the chips draw less than half the power of Nvidia GPUs while remaining compatible with PyTorch and vLLM. That comparison comes from the company's framing and has not been independently verified here.
Deployments so far are concentrated in South Korea. LG Group uses the chips for its EXAONE models, Upstage runs live search summaries on them, and Samsung SDS offers them through a neural-processing-unit-as-a-service product launched this summer.
FuriosaAI plans to produce 20,000 chips this year and 40,000 in 2027, with further growth afterward. A third-generation processor is in development with Broadcom. A NAVER Ventures partner said the company has moved from proving its architecture to proving its business, citing volume production and customer deployments.
Why it matters: Power availability is increasingly a gating factor for AI capacity, so efficiency-focused inference silicon could matter to operators with constrained sites. The claims remain largely company-reported, though, and production of tens of thousands of units is small next to Nvidia's installed base, so real-world efficiency and supply will decide how far this goes.
TwinStack automated analysis, based on reporting by SiliconANGLE. Check the original for full details.