NVIDIA, SK Telecom and Rebellions plan GPU-plus-NPU racks with NVLink Fusion

NVIDIA, SK Telecom and Korean AI chipmaker Rebellions announced a collaboration to connect Rebellions NPUs into NVIDIA GPU rack systems using NVLink Fusion. The arrangement was presented in Seoul alongside a memorandum of understanding.

The technical core is a split of labor across architectures. According to reports, NVIDIA GPUs handle prompt processing in the prefill stage while the Korean NPUs generate replies in decode. The two stages stress compute and memory bandwidth differently, so assigning them to different chip designs is one way to trade cost against throughput.

NVLink Fusion is what makes the combination possible, since it lets non-NVIDIA accelerators join NVIDIA's rack-scale interconnect and system designs. For a domestic NPU vendor like Rebellions, that means access to mainstream AI data-center power, networking and operations without having to build a full system ecosystem alone.

The three companies frame the work as sovereign AI infrastructure. Korea wants a loop connecting local models, local silicon and local data centers to reduce dependence on a single overseas supply chain, and the practical reality is that NPUs reach production by coexisting with GPU ecosystems rather than replacing them.

On timing, the partners expect systems to be ready in 2028. An MOU is not a binding commercial contract, though, and silicon volumes, software-stack maturity and actual customer demand all remain open questions. Cross-architecture systems are also far more complex to schedule and program than homogeneous clusters.

If the plan holds, the effect runs both ways: Rebellions gets a route into global AI infrastructure supply chains, while NVIDIA keeps heterogeneous partners inside its interconnect standard. What to watch is whether the software stack becomes transparent to developers, and whether named customers and deployment volumes follow.