Furiosa Raises $620M to Scale Sovereign AI Accelerators

By Elena Vasquez

Furiosa announced a $620 million Series D round this week to expand production of its RNGD AI accelerators, targeting about 20,000 units this year and roughly 40,000 in 2027. The round also funds continued development of a third-generation chip that the South Korean company is building with Broadcom.

What does sovereign AI change for South Korea?

Sovereign AI means “AI trained and run on accelerators you own in datacenters within your nation's borders.” For South Korea, “having the most powerful AI accelerator is secondary to having one that is domestically sourced or manufactured,” because access to US policy can change without much warning.

South Korea can import as many GPUs from Nvidia or AMD as it wants, but that purchasing power does not remove its dependence on decisions made in Washington. The report draws a Cold War parallel: developing the atomic bomb was the first step toward still more powerful nuclear weapons, and the same logic now shapes the AI arms race.

Sovereign AI is not unique to South Korea. Buyers across EMEA and the Asia-Pacific region also want alternative accelerators, both to diversify supply chains and to hedge against an AI arms race turning into a second Cold War, with Washington and Beijing restricting access to leading hardware.

What supply advantages does South Korea have?

South Korea can build a domestic accelerator base on top of a dense semiconductor supplier network. SK Hynix and Samsung rank among the world's two largest memory vendors, giving local designers an advantage in securing HBM capacity; Samsung Electronics also operates the world's second-largest leading-edge wafer fab. TSMC and Japan's Rapidus are nearby, shortening the route to advanced manufacturing.

Since the start of the AI boom, several credible South Korean chip startups have appeared and disappeared. Rebellions and FuriosaAI are the two front-runners, while Sapeon's brief run ended when its intellectual property and talent moved into Rebellions in late 2024.

How do Rebellions and Furiosa compare?

Both front-runners have working silicon, although their latest releases are not among the most sophisticated accelerators available. Their pitch does not depend on matching peak chip performance: China's domestic AI buildout shows the market can use large fleets of less powerful chips.

• Rebellions' Rebel100 first appeared in late 2025 and performs roughly on par with Nvidia's three-year-old H200.
• Furiosa's RNGD, examined in mid-2025, delivers about 2.25 times the performance per watt of Nvidia's six-year-old A100 in the tested workloads.

Having proven their designs, Rebellions and Furiosa are now advancing new systems and next-generation chips. The production push centers on Furiosa's RNGD, with the Series D intended to carry output from roughly 20,000 units this year to about twice that level in 2027.

What could Broadcom's rack-scale design contain?

Details of Furiosa's third-generation part remain limited, but a 2 nm process, multi-die chiplets and HBM4 or HBM4e memory indicate a more potent accelerator. Broadcom's participation points toward a rack-scale architecture using Tomahawk 6 or Tomahawk Ultra and/or 3.5D XDSiP packaging, although Furiosa has not confirmed the configuration.

Broadcom's 3.5D XDSiP is an advanced packaging reference architecture designed as a shortcut for XPU designers to bring AMD Instinct-style accelerators to market faster. The report says Meta and OpenAI have adopted the technology for their respective XPU platforms, making it a plausible basis for Furiosa's package design.

Even if Furiosa combines both approaches, the move to 2 nm and HBM4 or HBM4e points to a significantly more capable successor. Furiosa has not disclosed performance figures or launch timing, and its next chip will probably use TSMC's 2 nm process.

Where are commercial deployments heading?

Commercial deployments are beginning across the region. Japanese neocloud ai& plans to deploy up to 100 Rebellions RebelRack systems for its GPU-based compute offering, equivalent to 3,200 Rebel100 accelerators. That is a modest order beside the quarterly Blackwell and Instinct volumes shipped by Nvidia and AMD.

This summer, Furiosa agreed with Equinix to field RNGD accelerators at the colocation operator's LS2 facility in Lisbon. Samsung and LG Electronics are also among Furiosa's biggest patrons, giving the company an early customer base at home.

Why does non-US origin matter?

Last year, the Trump administration tried to cut China's access to Nvidia H20 accelerators, then quickly reversed the decision. Beijing still received the message that US chip designers could not be relied upon, reinforcing interest in accelerators outside the United States.

Rebellions and Furiosa do not offer the strongest chips, but their non-American origin gives buyers a way to reduce exposure to supply disruptions centered on the United States. For South Korean operators, the attraction is a domestic supplier option without giving up access to local HBM and nearby leading-edge foundry capacity.