Nvidia, SK Telecom and Rebellions have agreed to build a heterogeneous AI computing system that wires the Korean startup’s neural processors directly into Nvidia’s GPU racks. The three companies signed a memorandum of understanding on the sidelines of Nvidia’s AI Day Seoul event on Monday, and the system will underpin SK Telecom’s AI data centers. The tie-up runs on Nvidia’s NVLink Fusion program, which lets outside chip designers plug their own processors into Nvidia’s rack-scale infrastructure, and it carries the backing of South Korea’s science ministry.
What the three companies agreed
Under the memorandum, the companies will cooperate on a heterogeneous system that connects Nvidia’s graphics processors with neural processing units designed by Rebellions. The combination is aimed at the data centers SK Telecom is building for AI workloads, where the two kinds of chip would work side by side in the same racks rather than in separate systems.
The work is built on NVLink Fusion, Nvidia’s program for letting hyperscalers and custom chip designers integrate their own AI processors into its rack-scale systems. That matters because the interconnect, not just the chip, has become the battleground: whoever controls how processors talk to each other inside a rack shapes what a data center can run.
The agreement was signed during Nvidia’s two-day AI Day Seoul showcase, where the company laid out its sovereign-AI pitch for the country. The event, held at the COEX convention center in southern Seoul, drew the Korean government, carriers and chip startups into one room.
Why Seoul, and why now
South Korea’s ministry of science and ICT backed the signing, and it expects the partnership to widen the use of domestic advanced semiconductors and open global opportunities for local companies. At a media roundtable after signing, Jesse Clayton, Nvidia’s principal manager for AI infrastructure, said Korea is a technology leader and the collaboration could accelerate innovation there.
Nvidia also used the event to deepen its startup bets in Korea, holding a forum with more than 20 Korean startups working on foundation models, physical AI, AI infrastructure and AI chips. It announced plans to invest in Point2 Technology, alongside further investments in Motif Technologies, RLWRLD and Upstage.
The push follows Nvidia’s $1 billion stake in Naver, another big Korean bet from the same week, and it shows Seoul becoming a second front for the company’s expansion beyond Silicon Valley, as Nvidia courts the carriers, chip startups and government backing behind its sovereign-AI sales pitch abroad.
Our Read
This is Nvidia opening its walled garden just enough to keep it. NVLink Fusion lets a domestic chipmaker like Rebellions ride on Nvidia’s interconnect instead of building a rival one, which keeps Nvidia’s rack as the default platform while giving Seoul a local champion to point to.
For Rebellions, the prize is distribution: its NPUs get a path into SK Telecom’s data centers without displacing Nvidia’s GPUs. For SK Telecom, the payoff is a system it can call Korean-built, with government blessing, at a moment when every carrier wants its own AI infrastructure story.
The test will be whether the memorandum turns into shipping hardware, because a signed MOU is not a purchase order. Rebellions still has to prove its NPUs earn their place beside Nvidia’s GPUs at scale, and customers will judge the system on performance per dollar, not on ceremony.
The competitive pressure is real. Google’s new TPU 8tx and 8i Max chips show rivals racing to offer the same kind of heterogeneous compute, and the winner will be decided by which system lets customers mix and match processors most cheaply. Nvidia’s answer is to make its own rack the meeting point.
Why It Matters
Korea is the test case for Nvidia’s sovereign-AI sales pitch: a US chipmaker, a national carrier and a domestic chip startup building infrastructure a government can bless. If the heterogeneous system works in SK Telecom’s data centers, Nvidia can sell the same template to other countries that want local silicon inside an Nvidia rack.
If it stalls at the MOU stage, it joins a long list of AI Day announcements that never became products. The difference here is the hardware: Rebellions’ NPUs either run well on NVLink Fusion or they do not, and that verdict will come from engineers, not from signing ceremonies.
Background
Nvidia’s AI Day Seoul ran for two days at COEX, with workshops, a startup forum and the sovereign-AI showcase that opened the event. The Seoul push sits alongside Nvidia’s partnerships with Samsung and SK hynix, and its scheduled deliveries of GPUs to South Korea.
Source: Yonhap News Agency
0 Comments