FuriosaAI to scale RNGD inference chip production following Series D funding

FuriosaAI's Series D values the Seoul-based chip designer at nearly four times its July 2025 valuation, with Korea's state-backed National Growth Fund as the round's biggest investor.

South Korean AI chip designer FuriosaAI has closed a US$620 million Series D funding round at a post-money valuation of US$2.9 billion, as it ramps up production of its RNGD inference accelerator and develops a third-generation chip.

The round drew international investors Cisco Investments, Infinitum, Jericho Capital, NeoFort Holding, QUDA Capital and Worldline Capital, the company said. Korea's National Growth Fund committed KRW370 billion (about US$280 million), while the Export-Import Bank of Korea, UAMCO and Hanwha Asset Management came in as new investors.

FuriosaAI's three seed investors, NAVER (through NAVER Ventures), DSC Investment and KDB Capital, also returned, alongside Korea Development Bank, Nextrans and Kakao Investment.

The National Growth Fund was the round's largest investor, according to Seoul Economic Daily, which reported that demand from investors in Korea and overseas pushed the final amount above the company's initial target. Bloomberg reported in January that FuriosaAI had hired Morgan Stanley and Mirae Asset Securities to advise on the Series D, which was then aiming to raise between US$300 million and US$500 million ahead of a planned listing.

The new valuation is close to four times the US$735 million the company was worth after its US$125 million Series C bridge round in July 2025. Earlier that year, in March 2025, Bloomberg reported that FuriosaAI had turned down a US$800 million takeover offer from Meta Platforms in order to remain independent.

FuriosaAI Series D to fund RNGD production

FuriosaAI said the Series D will go towards two priorities: scaling RNGD production and deployments globally, and developing its Gen 3 chip.

RNGD went into mass production in January 2026. The company said it will produce 20,000 cards this year and has secured enough memory and fab capacity to more than double that in 2027.

The chip ships as a 180-watt, air-cooled PCIe card. Eight of them make up the NXT RNGD Server, a 4U, 3kW system rated at 4 petaflops (FP8) with 384GB of HBM3 memory. FuriosaAI said the server is designed to fit into the racks and power budgets enterprises already have, with no cooling retrofit.

The company claims RNGD matches the throughput of Nvidia's RTX PRO 6000 at the same card count while serving twice as many users per kilowatt. RNGD is aimed at models running at up to about 100 tokens per second per user, which FuriosaAI said is the level that matters most for enterprise workloads.

Samsung SDS has launched an NPU-as-a-Service offering powered by RNGD, and LG Group companies are deploying the chip in their on-premises hybrid cloud environments. FuriosaAI said it is also deploying RNGD with AI cloud providers and sovereign AI initiatives, but did not name those customers.

Infinitum chief investment officer John Yetimoglu said the company's move into volume production, together with its deployments at Samsung SDS and LG Group, gave the firm conviction in FuriosaAI’s ability to scale RNGD globally. "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," he said.

NAVER Ventures general partner Yong Jung Park said the Silicon Valley-based investor sees helping FuriosaAI grow in North America as where it can add the most value. FuriosaAI said it is hiring in Seoul, Silicon Valley, Lisbon and Singapore.

Gen 3 chip with Broadcom

FuriosaAI's third-generation chip is being developed with Broadcom and other partners, and will use a 2nm, multi-die design with HBM4/4E memory. The company said Gen 3 is meant to serve frontier agentic models running in large clusters.

Both generations are built on what FuriosaAI calls its Tensor Contraction Processor (TCP) architecture. According to the company, GPUs and most other AI accelerators break model operations into fixed-size, two-dimensional matrix multiplications. TCP instead treats tensor contraction, a higher-dimensional operation, as the basic unit of work for both the hardware and its software, which the company said lets its compiler plan across a whole model and reduce energy lost to moving data around.

The company has shipped four major SDK releases this year. The latest, SDK 2026.4, raised token generation per watt across 18 models compared with the previous release, including by 138% on gpt-oss-120B, according to FuriosaAI.

Korea's state funding for AI chips

TheElec reported in May that FuriosaAI is the second company to receive backing under what it described as the government's "K-Nvidia" initiative, following rival Rebellions. The National Growth Fund is a KRW150 trillion programme that the Korean government plans to deploy over five years into strategic industries including AI and semiconductors, according to a statement by the Financial Services Commission.

Rebellions raised US$400 million in a pre-IPO round in March that was co-led by Mirae Asset and the National Growth Fund, valuing it at about US$2.34 billion, Reuters reported.