NEWS → · NVIDIA · Sovereign AI
NVIDIA AI Day Seoul opens: Rebellions' next NPU joins NVLink Fusion for SK Telecom's sovereign racks
Nov 9, 2026
At AI Day Seoul, which opened on Nov 9, 2026 at COEX in Seoul, NVIDIA, Rebellions and SK Telecom announced a collaboration to build sovereign AI racks in which NVIDIA GPUs handle prefill and Rebellions' next-generation Rebel R200 NPUs handle decode — connected through NVIDIA's NVLink Fusion and built on its MGX rack architecture. The R200 systems are expected to be ready for deployment in 2028.
Image: NVIDIA via blogs.nvidia.com
What was announced, from the primary sources read on Nov 9, 2026: the three companies' release (GlobeNewswire, Nov 9, 2026) describes a memorandum of understanding, backed by Korea's Ministry of Science and ICT, under which Rebellions connects its R200 to NVIDIA's platform via NVLink Fusion and SK Telecom deploys the racks in its AI data centres for its sovereign AI services. No R200 specifications, volumes or prices were disclosed, and the 2028 readiness date makes this a roadmap, not a product, an order or a contract. NVIDIA's own live blog from the event (blogs.nvidia.com, Nov 8–9, 2026) carries the wider Seoul package: a deeper sovereign-AI collaboration with the ministry, physical-AI work with Samsung, SK hynix, Hyundai and LG, and planned investments in Korean startups.
The part renters can use now: NVIDIA is publishing Dynamo-optimised deployment recipes for four Korean open models — SK Telecom's A.X K2, LG AI Research's K-EXAONE 2.0, Upstage's Solar Open 2 and Motif Technologies' Motif-3 — optimised for B200, with NVIDIA claiming up to 2× out-of-the-box performance (NVIDIA's claim; the post publishes no workload or GPU count to check it against). Also at the event, LG AI Research opened its K-EXAONE 2.5 API free of charge: a 750-billion-parameter model with 31 billion active, text-only, with a 512K context — API access only, no weights, and no end date or quota published for the free tier.
What it means if you rent GPUs
NVLink Fusion opening to a third-party NPU inside MGX racks is the clearest sign yet that future rental fleets may be heterogeneous — GPU prefill plus cheaper NPU decode — which would change inference price-per-token maths. With readiness set for 2028 it changes nothing bookable today. What it does tell a renter in 2026 is where Korean sovereign demand is heading, and that NVIDIA is willing to let other silicon into its rack standard to keep that demand inside its ecosystem.