South Korea is seeking to have AI accelerators designed and manufactured on its own soil, to reduce its dependence on the United States. Its two main start-ups in the sector, FuriosaAI and Rebellions, have already produced functional chips and are preparing new generations.
For a long-time ally of Washington, the issue is not the ability to buy GPUs from Nvidia or AMD, which nothing prevents. The problem, according to the author, lies in the reliability of this supply in the face of a US government they describe as often capricious. What is granted can be withdrawn with little or no notice, and Seoul appears to be well aware of this.
The country has industrial strengths. It is home to two of the world’s largest memory manufacturers, SK Hynix and Samsung, giving local designers an advantage in obtaining HBM, the high-bandwidth memory essential for accelerators. Samsung Electronics also operates the world’s second-largest advanced chip fabrication plant, while TSMC and Rapidus, in Japan, are nearby.
Chips Already in Production
Several South Korean start-ups have emerged since the beginning of the AI boom, and some have disappeared. Sapeon, whose intellectual property and teams were integrated into Rebellions at the end of 2024, is one such case. FuriosaAI and Rebellions remain the two frontrunners.
Rebellions’ Rebel100, which appeared in late 2025, is roughly on par with Nvidia’s H200, a three-year-old accelerator. Furiosa’s RNGD is not the most powerful: according to tests cited in the article, it offers about 2.25 times the performance per watt of Nvidia’s A100, itself six years old, on the workloads tested. However, the author notes that, as with the Chinese strategy, unit power matters less when you have a large number of chips.
Furiosa announced a $620 million Series D funding round, aimed at increasing its production of RNGD and continuing the development of its third-generation accelerator, designed with Broadcom. The company plans to produce about 20,000 RNGDs this year, then roughly double that in 2027. Samsung and LG Electronics are among its key backers.
Information about the next chip remains limited. It is expected to be manufactured on a 2 nm process, likely at TSMC according to the article, based on a multi-die chiplet architecture and incorporating HBM4 or HBM4e memory. Broadcom’s involvement suggests a rack-scale architecture and 3.5D XDSiP packaging, a reference design by Broadcom that Meta and OpenAI have reportedly adopted. The author notes they cannot confirm this.
Demand Extends Beyond Seoul
The quest for sovereign AI, trained and operated on accelerators owned in national data centers, is not unique to South Korea. The article cites the Trump administration’s decision to cut China’s access to Nvidia’s H20, quickly reversed, which showed Beijing that US designers could no longer be considered reliable.
Japan’s ai& neocloud announced its intention to deploy up to 100 Rebellions RebelRack systems, which, according to the article, equates to 3,200 Rebel100 accelerators. This volume remains modest compared to the quarterly shipments of Nvidia’s Blackwell and AMD’s Instinct, but it exists. Furiosa, for its part, has struck a deal with Equinix to install its RNGDs in the LS2 center in Lisbon.
These chips may not have the best performance, but their non-American origin is enough, according to the article, to make them attractive to players who want to protect their AI investments from supply chain disruptions.
