Nvidia, SK Telecom and South Korean chip designer Rebellions signed an agreement in Seoul on Monday to develop and test an artificial intelligence system that uses two kinds of chips to answer a single request. Nvidia’s graphics processors would handle the initial work of processing a prompt, while a chip under development by Rebellions would generate the answer. SK Telecom plans to test the combined system in one of its AI data centers.
The project would give Rebellions a way to test its hardware inside an Nvidia rack, while giving SK Telecom a chance to see whether assigning different jobs to different chips can make AI services cheaper or more efficient to run. The telecom operator described a development and validation agreement. The companies have yet to publish performance results for the combined system.
What the companies agreed to build
The agreement was signed at Nvidia’s AI Day event in Seoul. According to the company announcement, the design would connect the Rebel R200, a neural processing unit still in development, to Nvidia hardware using NVLink Fusion. That technology lets chip designers integrate their own processors with Nvidia’s rack infrastructure. The proposed system would use Nvidia’s MGX rack design to bring the hardware together.
SK Telecom’s role is to install the system in an AI data center and evaluate its performance and stability before deciding whether it is suitable for commercial services. The operator’s plan calls for testing a whole rack, where the chips and the links between them must work together. Nvidia says it expects Rebellions’ R200 systems to be ready for deployment in 2028. That is a target for the hardware, rather than a confirmed launch date for an SK Telecom service.
Why a single answer takes two kinds of work
Inference is the work an AI model does after it has been trained. For a language model, the first step is to process the user’s input and any context that comes with it. This step, known as prefill, builds a working record of the input. The model then uses that record during decoding, when it generates one token at a time. Tokens can be words or pieces of words. The technical documentation for Nvidia’s Dynamo software describes why those steps put different demands on a machine.
A long prompt can take substantial computation to process at once. During decoding, the model repeatedly reads from memory as it generates the response. DistServe researchers, who studied separating those steps on GPU systems, found that putting them on different workers can keep the work of reading new prompts from interrupting answers already being generated. Their experiments did not test the Rebellions chip. The Seoul project proposes taking that division a step further by assigning each phase to a different kind of processor.
What a data center test could establish
The handoff matters as much as the speed of either chip. In a split system, the part that generates the answer needs access to the working record created by the part that read the prompt. As Nvidia’s documentation explains, how efficiently the system moves that record affects how well the split works. If the transfer takes too long or the software queues work unevenly, the benefit of using a separate processor can shrink.
A useful test would compare the time users wait for the first part of an answer, the pace at which the rest appears, and how many requests a rack can handle while maintaining those response times. The DistServe study measured first-token delay and the time between later tokens separately because they capture different parts of the experience. A system might do well on short questions and behave differently when many long requests arrive together. Power measurements for the whole rack would also help show whether any savings on one chip survive the added work of moving data and keeping both sets of processors busy.
Why South Korea is backing the project
South Korea’s science ministry sees the collaboration as a way for domestic chip developers to demonstrate their hardware on a platform used internationally, Yonhap reported. The ministry said it wants to support similar opportunities for other South Korean developers. SK Telecom has linked the work to a broader effort to combine locally developed models and chips with computing capacity it operates, including for potential public and private sector services.
The design would bring a Korean processor into the system while retaining Nvidia’s GPUs, connections and rack design. That makes the deal a test of how much room domestic chip designers can gain within an established platform. The next evidence will come from whether SK Telecom can run the planned system reliably, and whether it can deliver the same quality of service at a lower cost or with less power. For now, the companies have announced how they intend to divide the work; the joint announcement provides no measured savings.
