AI Arms Race Highlights the Importance of Sovereign Supply Chains

South Korea Builds an Independent AI Hardware Base

During the Cold War, developing an atomic bomb was only the first step toward producing more powerful nuclear weapons. A similar dynamic is emerging in the AI arms race. For countries such as South Korea, owning a domestically sourced or manufactured AI accelerator may be more important than having the most powerful device available, particularly if it reduces dependence on the United States.

South Korea is a longtime US ally and can import GPUs from Nvidia or AMD. However, access to those products can change with US policy, creating uncertainty for organizations and governments that depend on them.

South Korea is in a strong position to develop its own AI infrastructure. The country is home to two of the world's largest memory suppliers, SK Hynix and Samsung, giving domestic chip designers an advantage when securing high-bandwidth memory (HBM) capacity. Samsung Electronics is also the world's second-largest leading-edge wafer foundry. Taiwan Semiconductor Manufacturing Co and Japan's Rapidus are located relatively nearby.

Several credible South Korean chip startups have emerged since the beginning of the AI boom. FuriosaAI and Rebellions are the two leading companies. Sapeon was also active for a short period, but its intellectual property and personnel were incorporated into Rebellions in late 2024.

Both companies have produced working silicon, an accomplishment that many chip startups do not reach. Their latest products, however, are not yet among the most sophisticated AI accelerators.

Rebellions and FuriosaAI Develop New Accelerators

Rebellions' Rebel100 accelerator, introduced in late 2025, is approximately comparable to Nvidia's three-year-old H200. FuriosaAI's RNGD accelerator is less powerful, delivering about 2.25 times the performance per watt of Nvidia's six-year-old A100 in the tested workloads.

As China's domestic AI infrastructure effort has demonstrated, accelerators do not necessarily need to be the fastest available if enough of them can be deployed. After proving their designs, Rebellions and FuriosaAI are now working on larger systems and next-generation chips.

FuriosaAI recently announced a $620 million Series D funding round. The company plans to use the funding to expand production of its RNGD accelerators and continue developing third-generation accelerators in collaboration with Broadcom.

FuriosaAI expects to produce approximately 20,000 RNGD accelerators this year and about twice that number in 2027. Samsung and LG Electronics are among the company's largest backers.

Details about FuriosaAI's next accelerator remain limited. The company has said that the chip will use a 2 nm manufacturing process, presumably at TSMC, along with a multi-die chiplet design. Depending on its development schedule, it will incorporate HBM4 or HBM4e.

Broadcom's role in the development suggests a rack-scale architecture that could use technologies such as Tomahawk 6, Tomahawk Ultra, and 3.5D XDSiP for package design.

3.5D XDSiP is a Broadcom-designed advanced packaging reference architecture intended to help XPU developers bring AMD Instinct-style accelerators to market more quickly. Meta and OpenAI have reportedly adopted the technology for their respective XPU platforms.

FuriosaAI has not confirmed that it will use this exact combination of technologies. However, its previous disclosures point in that direction. Moving to a 2 nm process and HBM4 or HBM4e would represent a substantial increase in accelerator capability.

Demand for Non-US AI Hardware Grows

The pursuit of sovereign AI, meaning AI trained and operated on accelerators owned by an organization or government and installed in data centers within its national borders, is not limited to South Korea. Organizations in other regions also want to diversify their supply chains in case geopolitical tensions develop into a new Cold War and the United States or China restricts access to advanced hardware.

The Trump administration's attempt to restrict China's access to Nvidia's H20 accelerators illustrated the risks. Although the decision was later reversed, it reinforced the view in Beijing that US chip designers could not be relied upon indefinitely.

Interest in alternative accelerators has consequently grown across Europe, the Middle East, Africa, and the Asia-Pacific region. Japanese neocloud ai&, for example, announced plans to deploy up to 100 of Rebellions' RebelRack systems for its GPU-based computing service. That represents 3,200 Rebel100 accelerators.

The deployment is small compared with the number of Blackwell and Instinct accelerators shipped by Nvidia and AMD each quarter, but it provides Rebellions with a foothold outside South Korea.

FuriosaAI is also finding opportunities in Europe, where the supply of capable locally developed AI accelerators remains limited. The company reached an agreement with Equinix to deploy RNGD accelerators at the colocation provider's LS2 facility in Lisbon.

Rebellions and FuriosaAI do not currently offer the most powerful AI chips. Their products are nevertheless attractive to organizations seeking to reduce exposure to supply-chain disruptions and dependence on US-based hardware suppliers.