NVIDIA, Korean chip startup Rebellions and telecom giant SK Telecom announced a collaboration on November 9 to build what they call sovereign AI infrastructure inside Korea, using NVIDIA's NVLink Fusion technology to connect Rebellions' Rebel R200 neural processing units to NVIDIA's platform S¹. The move lets a Korean-designed chip plug into NVIDIA's rack architecture rather than compete against it. NVLink Fusion is NVIDIA's framework for linking third-party processors into its systems, broadening the platform beyond GPUs the company itself manufactures P⁵S¹. Korea's Ministry of Science and ICT is backing the collaboration S¹.
My read: The framing here is openness, and NVLink Fusion is genuinely a door NVIDIA has opened. But the door leads into NVIDIA's own house. Every partner that adopts MGX rack architecture and NVLink Fusion is buying into NVIDIA's interconnect standards, which makes switching costs enormous once deployed. The claim I would watch is whether Rebellions' R200 actually ships in volume inside SK Telecom data centres, because right now there is no deployment timeline, no benchmark, and no independent test of the energy-efficiency claims. This is an announcement of intent, not a product launch.
Jensen Huang said Korea has the talent and industrial ambition to lead in AI, and that NVLink Fusion gives Korean chipmakers a way to attach their own ASICs to NVIDIA's AI factory platform S¹. Sunghyun Park of Rebellions called the announcement a turning point, saying the integration would bring low-latency, energy-efficient inference for frontier models at scale S¹ — though these energy-efficiency figures are Rebellions' own, with no released code, no error bars, and no independent verification. Neither company has published benchmarks for the R200 inside an NVLink Fusion system, and no one outside the three parties has tested the claims.
NVLink Fusion lets outside chips slot into NVIDIA racks
NVIDIA markets NVLink Fusion as a way for hyperscalers and AI-focused firms to bring their own custom XPUs and CPUs into NVIDIA's infrastructure P⁵. In practice, a chipmaker like Rebellions can design its own silicon for inference — the work of running a trained model, and still fit it into NVIDIA's MGX rack architecture alongside NVIDIA's own GPUs S¹. The Rebel R200 is a Korean-designed AI processor S¹. Rebellions plans to use MGX for rack-scale system development S¹. SK Telecom plans to deploy the resulting systems in its AI data centres, linking them to its proprietary A.X K2 models and public-facing services under Korea's AI for All project S¹.
Korean telecoms gain a domestic-chip path to AI inference
For a Korean telecom running inference on foreign GPUs, the pitch is straightforward: domestic chips, domestic data centres, domestic models. Jung Jai-hun of SK Telecom said the collaboration validates next-generation sovereign AI infrastructure and that SKT will link its data centres, A.X K2 models, and public AI services S¹.
A rack engineer at SK Telecom would now need to configure MGX chassis to accept R200 inference boards alongside NVIDIA GPUs, rather than sourcing and integrating foreign-only silicon — a shift in procurement and system design, even before any benchmark exists.
NVIDIA's own blog, published November 8, frames the Korea collaboration as spanning from AI semiconductors and data centre infrastructure to talent P², but the press release discloses no timeline for commercial deployment.
The next checkpoint is whether Rebellions demonstrates a working R200 inside an MGX rack at a public event. Until then, this is architecture on paper.
Sources: S1 — NVIDIA, Rebellions and SK Telecom Announce Collaboration to Advance So · P2 — From Chips to Skills, NVIDIA and Korea Expand Sovereign AI Collaborati · P3 — NVIDIA Hardware Innovations and Open Source Contributions Are Shaping · P4 — Latest News | NVIDIA Newsroom · P5 — Build Semi-Custom AI Infrastructure | NVIDIA NVLink Fusion
Written from 5 sourced items, 5 of them primary.
