What changed
NVIDIA, Rebellions and SK Telecom announced a planned Korean sovereign-AI stack: Rebellions will connect its next-generation R200 inference NPUs with NVLink Fusion and NVIDIA MGX racks, while SK Telecom plans deployment in its AI data center. The companies also plan split inference with NVIDIA GPUs for prefill and R200 NPUs for decode.
Why it matters
The design is a concrete test of mixing national accelerator silicon with NVIDIA rack infrastructure and assigning different inference phases to different processors. If delivered, procurement shifts from a single-chip choice to interconnect, scheduler, software and failure-domain evidence.
What to do
Wait for hardware and service evidence. Ask for R200 silicon status, an orderable system, supported model and precision matrix, interconnect topology, prefill-decode scheduler, failover, measured latency and energy, price, data residency and an independent acceptance benchmark before planning capacity.
Caveat
This is a vendor collaboration and 2028 deployment plan, not available infrastructure. No orderable configuration, price, independent benchmark, production reliability result or completed heterogeneous prefill-decode system was published.