Furiosa Raises $620M Series D as Korea Builds Sovereign AI Chips

FuriosaAI has closed a $620 million Series D round to scale production of its RNGD accelerators and develop a third-generation chip with Broadcom, built on a 2 nm process with HBM4 or HBM4e memory. The round, announced this week, is the largest concrete bet yet on South Korea's push for AI silicon it does not have to import from the United States.

The company expects to produce roughly 20,000 RNGD accelerators this year and about double that volume in 2027, according to its own disclosures. Samsung and LG Electronics rank among its largest customers. Those two names matter more than the funding figure: they anchor Furiosa's revenue in the same industrial ecosystem that supplies its memory.

Why is South Korea building its own accelerators?

The logic is Cold War arithmetic. During the original arms race, the atomic bomb was the first step toward more powerful weapons. South Korea has applied the same reasoning to AI: owning the most powerful accelerator matters less than owning one that is domestically designed or manufactured.

As a longtime US ally, South Korea can buy as many Nvidia or AMD GPUs as it wants. Ability is not the issue. Trust is. Washington has shown, repeatedly and with little warning, that what it grants it can revoke. The Trump administration's attempt last year to cut off China's access to Nvidia's H20 accelerators was reversed quickly, but Beijing had already drawn its conclusion: American chip designers cannot be counted on.

South Korea holds unusual cards. It hosts two of the world's largest memory vendors, SK Hynix and Samsung, which gives domestic chip designers an edge in securing HBM capacity. Samsung Electronics is also the world's second-largest leading-edge wafer fab operator, and TSMC plus Japan's Rapidus sit a short hop away.

How competitive are the chips themselves?

Frankly, they are not. Rebellions' Rebel100 accelerator, which first appeared in late 2025, 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 tested workloads.

The comparison is unflattering but arguably beside the point. China's domestic AI infrastructure push demonstrated that chips need not be fast if you have enough of them. Rebellions and Furiosa have both shipped working silicon, which is more than most chip startups can claim, and both are developing next-generation systems.

A handful of credible Korean startups have come and gone since the AI boom began. FuriosaAI and Rebellions are the frontrunners; Sapeon existed briefly before its IP and talent folded into Rebellions in late 2024.

What do we know about the next-generation chip?

Furiosa has disclosed three specifications for its third-generation accelerator: a 2 nm process, presumably at TSMC; a multi-die chiplet architecture; and HBM4 or HBM4e memory depending on timing. Broadcom's involvement points toward a rack-scale design using Tomahawk 6, Tomahawk Ultra networking, or 3.5D XDSiP packaging, possibly a combination of both.

3.5D XDSiP is Broadcom's advanced packaging reference architecture, a shortcut that lets XPU designers bring AMD Instinct-style accelerators to market faster. Meta and OpenAI have both adopted the technology for their respective XPU platforms. Furiosa has not confirmed its exact approach, but its prior disclosures suggest the two elements combined. A 2 nm process plus HBM4 points to a significantly more potent accelerator.

Who is buying outside Korea?

Interest in non-American accelerators is growing across EMEA and Asia-Pacific. Japanese neocloud ai& recently announced plans to deploy up to 100 of Rebellions' RebelRack systems, which works out to 3,200 Rebel100 accelerators. That is a modest number next to the Blackwell and Instinct volumes Nvidia and AMD ship each quarter, but it is a start.

Furiosa, meanwhile, is capitalizing on the shortage of competent European AI accelerators. This summer it landed a deal with Equinix to field RNGD accelerators at the colocation provider's LS2 facility in Lisbon.

Rebellions and Furiosa do not have the most powerful chips. The simple fact that they are not American is enough to make them attractive to anyone seeking to insulate AI investments against supply chain disruption. With 2 nm silicon and HBM4 on the roadmap and volumes doubling toward 2027, both startups now have the funding to test whether sovereignty sells at scale.