FuriosaAI announced on 29 October that it has closed a 620 million dollar Series D funding round at a post-money valuation of 2.9 billion dollars. The South Korean chipmaker said its RNGD inference accelerator is already in mass production, with 20,000 cards to be produced this year and enough memory and fabrication capacity secured to more than double that in 2027.
The chip at the centre of the round
RNGD is a 180-watt, air-cooled PCIe card built specifically for inference, the work of running AI models rather than training them. Eight cards make the NXT RNGD Server, a 4U system drawing 3 kilowatts and delivering 4 petaFLOPS of FP8 compute with 384 gigabytes of HBM3 memory, designed to fit the racks and power budgets enterprises already have.
FuriosaAI said RNGD matches the throughput of Nvidia’s RTX PRO 6000 at the same card count while serving twice as many users per kilowatt of power. That claim goes to the heart of the inference market: training happens once, while models are run billions of times, so power and cost per token decide which hardware wins.
The chip targets models running at up to about 100 tokens per second per user, the level FuriosaAI says matters most for enterprise workloads. It is designed to run the latest open-weight models, including Gemma 4 and Qwen 3.8-27B, and the company said it reached mass production in January 2026 after years of development.
Production deployments give the round its credibility. Samsung SDS has launched an NPU-as-a-Service offering powered by RNGD, while LG Group companies are deploying the chip in on-premises hybrid cloud environments. Those are paying, large-scale customers, not pilot projects, and they separate FuriosaAI from the many chip startups still waiting for their first volume order.
Who is paying, and what comes next
Korea’s National Growth Fund committed 370 billion won, about 280 million dollars, anchoring the round. International investors included Cisco Investments, Infinitum, Jericho Capital, NeoFort Holding, QUDA Capital and Worldline Capital, while the Export-Import Bank of Korea, UAMCO and Hanwha Asset Management joined as new investors.
All three original seed investors returned: NAVER through NAVER Ventures, DSC Investment and KDB Capital, alongside Korea Development Bank, Nextrans, Kakao Investment and other early backers. Bloomberg reported that the new valuation roughly quadruples the previous round, a measure of how far the company has come since its founding in 2017, the year after AlphaGo’s victory in Seoul.
The money has two stated priorities: scaling RNGD production and deployments worldwide, and building a third-generation chip for frontier agentic models in large clusters. Gen 3 is in development with Broadcom and other partners, using a 2-nanometre, multi-die design with HBM4 and HBM4E memory, and the company’s Tensor Contraction Processor architecture.
Software is the other half of the plan. FuriosaAI has shipped four major SDK releases this year, adding hybrid batching, prefix caching and native Kubernetes support, and its latest release raises token generation per watt across 18 models compared with its predecessor. For a challenger chip, software maturity is often the deciding factor for customers weighing a switch.
Why It Matters
Inference is where the AI chip race will be decided. Training grabs headlines, but most compute spending over a model’s life goes to serving it, and power has become the binding constraint in data centres everywhere. A chip that delivers more tokens per watt, with named enterprise customers, is a credible alternative in a market dominated by one supplier.
That dominance is already under pressure from several directions, including hyperscalers opening their clouds to rival silicon, as covered in Oracle’s embrace of rival chips. FuriosaAI’s round shows investors are willing to fund that challenge at scale, and its production milestone shows the challenge is no longer theoretical.
Background
FuriosaAI was founded in Seoul in 2017 to build inference accelerators that balance performance, efficiency and programmability, after AlphaGo’s win made AI’s infrastructure needs unmistakable.
Source: FuriosaAI’s announcement
0 Comments