FuriosaAI Lands $620M at a $2.9B Valuation to Build More 180-Watt RNGD Inference Cards and a 2nm Successor

Most of the money in AI hardware still follows training clusters. FuriosaAI is betting that the bigger, longer-lasting bill is inference, and that power, not raw compute, is what limits it. On October 29, 2026, the company announced a $620 million Series D at a post-money valuation of $2.9 billion, saying its RNGD accelerator is now in mass production.

The company names Infinitum, Jericho Capital, QUDA Capital, Cisco Investments, NeoFort Holding and Worldline Capital as its international investors in the round. Closer to home, it says Korea’s National Growth Fund committed KRW 370 billion, about $280 million. The Export-Import Bank of Korea, UAMCO and Hanwha Asset Management joined as new investors, and its three original seed backers, NAVER (through NAVER Ventures), DSC Investment and KDB Capital, put in money again, along with Korea Development Bank, Nextrans and Kakao Investment.

A chip designed around a power budget

FuriosaAI was founded in 2017, a year after AlphaGo beat Lee Sedol in Seoul. The company frames its goal as three requirements that usually pull against each other: the most tokens from every chip and rack, low power draw, and programmability that does not depend on months of hand-tuned kernels.

Its announcement compares the approach to Formula One, where the rules fix engine displacement and engineers chase performance from a set amount of fuel. RNGD is a 180-watt, air-cooled PCIe card. Eight of them make up the NXT RNGD Server, which the company specifies at 4U, 3 kW, 4 petaFLOPS of FP8 compute and 384 GB of HBM3. It is designed to fit the racks and power budgets enterprises already have, with no cooling retrofit.

FuriosaAI says RNGD matches the throughput of NVIDIA’s RTX PRO 6000 at the same card count while serving twice as many users per kilowatt. That comparison is the company’s own. It also says the chip targets models running at up to about 100 tokens per second per user, which it considers the range that matters most for enterprise workloads, and that it runs open models such as Gemma 4 and Qwen 3.8-27B.

Tensor contractions instead of matrix tiles

The architecture behind the card is what FuriosaAI calls the Tensor Contraction Processor. According to the company, GPUs and most other accelerators break model operations into fixed-size two-dimensional matrix multiplications. TCP instead makes tensor contraction, a higher-dimensional operation, the basic unit of work for both hardware and software.

The company says that lets its compiler plan across a whole model and, because execution on the silicon is deterministic, cut energy-wasting data movement. Reconfigurable compute slices are meant to keep the chip busy at both prefill and decode, and at large or small batch sizes.

On software, FuriosaAI says it has shipped four major SDK releases this year, adding hybrid batching of prefill and decode requests, prefix caching and native Kubernetes support. It says SDK 2026.4, released this month, raised token generation per watt across 18 models compared with SDK 2026.3, including by 138% on gpt-oss-120B. Developers can deploy through PyTorch as they would on GPUs, or write kernels in the company’s Tensor Contraction Language and Virtual ISA.

From 20,000 cards to a third generation

RNGD reached mass production in January 2026. FuriosaAI says it will make 20,000 cards this year and has secured the memory and fab capacity to more than double that in 2027. Named deployments include Samsung SDS, which has launched an NPU-as-a-Service offering on RNGD, and LG Group companies running it in on-premises hybrid cloud environments.

The company says the Series D has two priorities: scaling RNGD production and deployments globally, and building a third-generation chip to serve frontier agentic models in large clusters. That chip is in development with Broadcom and other partners, and will use the TCP architecture in a 2nm multi-die design with HBM4 or HBM4E memory. No release date was given.

Investors framed the round around the move from prototype to product. “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. NAVER Ventures general partner Yong Jung Park said the firm sees FuriosaAI moving “from proving its technology to proving its business,” and that helping it grow in North America is where NAVER Ventures can add the most value.

FuriosaAI did not disclose revenue or the size of its customer base. It is hiring in Seoul, Silicon Valley, Lisbon and Singapore.