FuriosaAI, a South Korean maker of chips for running AI models, has raised $620m in a Series D round at a post-money valuation of $2.9bn. The company says it will produce 20,000 of its RNGD inference cards this year and has secured enough memory and chip-making capacity to more than double that in 2027.
The round, announced on FuriosaAI’s blog and on X, mixes foreign investors with a large commitment from the Korean state, and it pays for two things: making more of today’s chip and designing a third-generation one with Broadcom.
A state fund and foreign investors
FuriosaAI names six international investors: Infinitum, Jericho Capital, QUDA Capital, Cisco Investments, NeoFort Holding and Worldline Capital. Korea’s National Growth Fund committed KRW 370bn, about $280m, or close to half the round. The Export-Import Bank of Korea, UAMCO and Hanwha Asset Management joined as new investors. Earlier backers returned, including NAVER through NAVER Ventures, DSC Investment, KDB Capital, Korea Development Bank, Nextrans and Kakao Investment. The company did not name a lead investor.
“Taking an advanced AI chip from design to volume production is an exceptionally difficult undertaking, and few companies have demonstrated the ability to do so successfully,” said John Yetimoglu, chief investment officer at Infinitum.
What RNGD is built for
RNGD is a 180-watt, air-cooled PCIe card designed only for inference, the work of running trained models rather than training them. Eight cards make up FuriosaAI’s NXT RNGD server: a 4U box drawing 3kW, with 4 petaflops of FP8 compute and 384GB of HBM3 memory. The pitch is that it fits racks and power budgets that data centres already have, without new cooling.
FuriosaAI says RNGD matches the throughput of NVIDIA’s RTX PRO 6000 at the same card count and “serves twice as many users per kW of power”. That is the company’s own benchmark; TSN has not seen an independent test. The chip entered mass production in January 2026. Samsung SDS sells an NPU-as-a-Service product built on it, and LG Group companies are deploying it in their own data centres.
Software matters as much as silicon for a challenger. FuriosaAI says its SDK 2026.4, released this month, raised tokens generated per watt across 18 models, by 138% on gpt-oss-120B compared with the previous release.
Gen 3 moves to 2nm with Broadcom
The money is split between scaling RNGD production and building a third-generation chip “to serve frontier agentic models in large clusters”. FuriosaAI says Gen 3 is in development with Broadcom and other partners, and will use a 2nm, multi-die design with HBM4 or HBM4E memory. It gave no date.
Power, not peak speed, is the bet
FuriosaAI is not trying to beat NVIDIA’s biggest training systems. It is betting that power will limit AI deployment more than raw speed, and that many companies will want to run open models on hardware that fits existing racks. The comparison it chose, a workstation-class NVIDIA card rather than a data-centre flagship, shows where it expects to compete.
Production numbers give that bet some weight. Many AI chip start-ups have struggled to move from samples to volume, and 20,000 cards in a year is a real, if modest, figure next to NVIDIA’s output. The harder step is next year’s doubling, which depends on memory supply. High-bandwidth memory is scarce enough that SK hynix and Samsung are posting record profits from it, and Gen 3’s HBM4 will compete for the same supply as everyone else.
The state’s role is notable too. Korea is pushing hard for domestic AI chips, from Mobilint’s robot chips to large compute deals such as Naver’s with Brookfield. Elsewhere, inference chip makers are consolidating, as with Unconventional AI’s plan to buy NeuReality. A $2.9bn valuation gives FuriosaAI room to stay independent, at least through Gen 3.
The next test is whether customers beyond Samsung SDS and LG sign on in public before the 2027 ramp.
Sources
- FuriosaAI, “FuriosaAI closes $620 million Series D to scale efficient, production-ready AI inference”, 29 October 2026
- @FuriosaAI, post on X, 28 October 2026
Related stories
- Mobilint to Build Humanoid Robot AI Chips in Korea’s ₩800bn Programme
- Unconventional AI to Buy NeuReality as It Chases a 1,000x Chip Goal
- SK hynix Tops 100 Trillion Won in Quarterly Revenue for the First Time
Follow @tsnmediahub on X for breaking AI news as it happens.
Follow on X
