NVIDIA, Rebellions and SK Telecom have announced a collaboration to advance Korea’s sovereign AI infrastructure with NVIDIA NVLink Fusion. In an 8 November Newsroom release, the companies said Rebellions plans to adopt NVLink Fusion to connect its next-generation R200 neural processing units, and that SK Telecom plans to deploy these systems in its AI data center.
Nothing in the announcement is described as already live. The same release says Rebellions’ R200 systems are expected to be ready for deployment in 2028. The deal is an MoU-backed plan to couple Korean-designed NPUs with NVIDIA’s rack-scale platform, supported by Korea’s Ministry of Science and ICT, rather than a completed build-out.
NVLink Fusion is meant to join custom ASICs to NVIDIA’s rack stack
NVLink Fusion, on NVIDIA’s account, extends its open infrastructure platform so chipmakers can connect custom XPUs to NVIDIA’s high-bandwidth scale-up networking and MGX rack architecture. Rebellions plans to use that path to link Rebel R200 NPUs within a rack through NVIDIA NVLink scale-up networking, while drawing on MGX reference designs and the MGX supplier ecosystem.
The R200 is framed as a next-generation inference NPU aimed at low-latency, energy-efficient serving of frontier models, including multitrillion-parameter mixture-of-experts models. Those product aims are NVIDIA’s and Rebellions’ forward-looking description; the release does not claim commercial R200 racks are shipping today.
The Korean stack angle sits beside other TSN coverage of national AI build-outs, from Microsoft’s Japan sovereign-infrastructure push to SK Telecom’s own agent-operations work and Oracle’s NVIDIA Vera Rubin cloud systems. The common thread is countries and carriers trying to own more of the model-and-infrastructure stack without writing every interconnect from scratch.
Jensen Huang said: “With NVLink Fusion, Korea’s chipmakers can connect their custom ASICs to NVIDIA’s AI factory platform, tightly coupling their processors with NVIDIA compute to run AI workloads.” Sunghyun Park, Rebellions’ cofounder and CEO, called the announcement “a turning point, opening a new chapter for sovereign AI by connecting Korea’s AI silicon and models with global systems technology.” Jung Jai-hun, SK Telecom’s president and CEO, said the collaboration “validates the potential of next-generation sovereign AI infrastructure by combining global AI infrastructure technology with Korean AI semiconductors.”
Prefill on GPUs, decode on NPUs
The companies plan to develop disaggregated inference: splitting the stages of generating an AI response across different processors. In plain terms, the planned work assigns NVIDIA GPUs to prefill—the heavy first pass that reads the prompt and fills the model’s working state—and Rebellions NPUs to decode—the step-by-step generation of the reply tokens. NVIDIA says that split is intended to improve inference throughput and power efficiency; it remains a planned approach, not a measured production result in this release.
SK Telecom, in NVIDIA’s telling, plans to connect the resulting infrastructure with its proprietary foundation models and public-facing services, including work under the government’s AI for All project, using the standardised MGX rack footprint for power and cooling planning. Bae Kyung-hoon, Korea’s deputy prime minister and science and ICT minister, said the collaboration “enables Korean-developed NPUs to enter the computing infrastructure market alongside NVIDIA… through NVLink Fusion.”
Separately, a 9 November NVIDIA blog said NVIDIA also announced an investment in Korean startup Point2, with additional investments planned in Motif Technologies, RLWRLD and Upstage. No amounts were disclosed. Those startup lines sit alongside the NVLink Fusion MoU in the same Korea programme week; they are not proof that the R200 racks exist yet.
An interconnect MoU is not a rack on the floor
Sovereign AI announcements often collapse three different things: a political commitment, a silicon roadmap and a data-centre deployment. This one is clearest as the first two. MSIT support, CEO quotes and an NVLink Fusion plan tell operators that Korean NPUs are meant to sit inside NVIDIA’s MGX world. The 2028 readiness line for R200 systems is the calendar that keeps the story honest.
Disaggregated inference is the technical reason the MoU matters if it works. Prefill and decode stress memory bandwidth and compute differently; giving each stage a specialist chip is a recognised industry direction. Calling it planned work, with GPUs on prefill and NPUs on decode, is more useful to readers than treating “fusion” as a finished product name. Until SK Telecom reports what it actually validated in its AI data centre—performance, stability and any commercial review—the efficiency claims stay aspirational.
The useful next evidence is concrete: MGX reference systems with R200 silicon, public validation numbers from SK Telecom’s tests, and whether Point2 and the planned Motif, RLWRLD and Upstage investments show up as disclosed financings. Until then, Korea has a documented path to couple domestic NPUs to NVIDIA’s rack fabric—and a date, 2028, before those R200 systems are expected to be ready for deployment.
Sources
- NVIDIA Newsroom: NVIDIA, Rebellions and SK Telecom NVLink Fusion collaboration, 8 November 2026
- NVIDIA Blog: Korea sovereign AI collaboration context, 8 November 2026
- NVIDIA Blog: planned Korean startup investments, 9 November 2026
- SK Telecom newsroom release, 9 November 2026
- MSIT press release via korea.kr, 9 November 2026
Related stories
SK Telecom’s AI-agent employee-ID programme is the nearest operator-side counterpart: the same carrier trying to industrialise Korean AI services while this MoU plans the rack layer underneath.
Follow @tsnmediahub on X for breaking AI news as it happens.
Follow on X
