NVIDIA and Korean partners plan a GPU–NPU inference system for 2028 NVIDIA and Korean partners plan a GPU–NPU inference system for 2028

Different chips for different stages

SK Telecom describes two distinct demands inside language-model inference. Prefill processes the incoming prompt and needs substantial computation. Decode generates the response incrementally and puts pressure on memory access. The proposed arrangement assigns the first stage to NVIDIA GPUs and the second to Rebellions' neural processing units. Separating those jobs lets the partners design around their different requirements. For an application, the first stage affects how long a user waits before an answer starts. The second affects how quickly that answer continues to arrive. A long document followed by a short response creates a different balance from a short prompt asking for extensive output.

The connection is part of the product

NVIDIA says Rebellions will use NVLink Fusion to join its developing R200 processors to NVIDIA's scale-up network. The project also uses MGX, NVIDIA's modular rack architecture. That provides common design elements for power delivery and cooling, alongside access to component makers and system manufacturers. The partnership therefore covers more than a processor attached to a cable. It gives the custom chip a route into an established rack and supply chain. NVIDIA expects R200 systems to be ready for deployment in 2028. The schedule places this announcement in infrastructure planning, beyond the buying cycle of a team needing inference capacity immediately.

The operator supplies the validation setting

SK Telecom's announcement describes an agreement to develop and verify the combined infrastructure. Its data centers would provide a setting to test performance and stability, then assess service use and commercialization. The operator identifies its A.X K2 model and Korea's AI for All initiative among the intended applications. That connects the hardware project to an operator with models and services of its own. It also defines a practical sequence: assemble the system, establish how it behaves, and evaluate where it belongs in service. Public and private Korean deployments are the intended market for that work.

Custom silicon still uses NVIDIA's infrastructure

NVLink Fusion already had a custom-chip partner before this agreement. On May 19, 2025, Marvell announced plans to combine its custom cloud silicon with NVIDIA's rack architecture and connectivity. The Korean project adds a named processor, an operator and a proposed division of inference work to that approach. Its sovereignty proposition combines domestic chip development and local services with NVIDIA infrastructure. Those are separate layers of control. Choosing a Korean inference processor does not remove the role of the interconnect and rack supplier. The agreement illustrates how a custom accelerator can enter the market through an incumbent's surrounding system.