South Korea’s FuriosaAI announced the closing of a colossal Series D financing totalling US$620 million. The injection pushes the company’s post-money valuation to roughly US$2.9 billion, positioning it as a serious contender in the global AI-silicon arena, a space traditionally dominated by the likes of Nvidia.
The round was led by a blend of national and overseas backers. Korea’s National Growth Fund contributed the lion’s share of KRW370 billion (about $280 million), while the Export-Import Bank of Korea, UAMCO, and Hanwha Asset Management joined as new participants. Existing seed investors – NAVER Ventures, DSC Investment and KDB Capital – refreshed their stakes, and additional support flowed from Korea Development Bank, Nextrans and Kakao Investment. International players such as Silicon Valley’s Infinitum, New York’s Jericho Capital, Singapore’s QUDA Capital, Cisco Investments, Abu Dhabi’s NeoFort Holding and Worldline Capital rounded out the syndicate, underscoring the global appetite for deep-tech AI hardware.
Production surge for the RNGD inference accelerator
With the fresh capital, FuriosaAI plans to accelerate output of its second-generation inference chip, the RNGD. The device entered mass production in January 2026 and is engineered for the most demanding data-centre workloads where efficiency cannot be compromised. The company has set a target of 20,000 RNGD units for 2026 and intends to double that number to over 40,000 units in 2027. CEO and co-founder June Paik explains that inference is the true cost driver because a model is trained once but executed billions of times in production. RNGD is tuned for workloads that demand roughly 100 tokens per second per user – a sweet spot for enterprise-scale applications.
The accelerator’s architecture couples a proprietary Tensor Contraction Processor (TCP) with high-bandwidth memory, delivering high throughput while staying within typical data-centre power and cooling envelopes. Early adopters report that RNGD can run cutting-edge open models such as Gemma 4 and Qwen 3.8-27B with minimal latency, and its programmable fabric allows customers to fine-tune kernels for bespoke workloads.
Strategic customers and the path to a third-generation chip
Major Korean conglomerates have already incorporated RNGD into their AI services. LG Group leverages the accelerator for its EXAONE platform, while Upstage uses it to power live-search summarisation for millions of daily queries. Samsung SDS launched a commercial NPU-as-a-service offering built on RNGD in July, signalling confidence in the chip’s reliability and scalability.
Beyond immediate production, FuriosaAI is collaborating with Broadcom to design a third-generation AI accelerator aimed at “frontier agentic models” – the next wave of large language models that exhibit more autonomous reasoning capabilities. The partnership will marry Broadcom’s advanced packaging expertise with FuriosaAI’s TCP core, targeting hyperscale data-centre deployments that demand both raw performance and extreme energy efficiency.
The combined momentum of a robust funding package, expanding customer base, and a clear road-map toward a next-gen silicon solution positions FuriosaAI as a rising challenger to entrenched GPU vendors. As the AI inference market matures, the company’s focus on sustainable, high-performance compute could reshape the economics of large-scale model deployment.
