Nvidia, Rebellions and SK Telecom Plan Mixed GPU-NPU Racks for Korea's Sovereign AI

Under the collaboration, which Nvidia announced on November 8, Rebellions will connect its next-generation Rebel R200 NPUs to Nvidia's platform through NVLink Fusion. That is Nvidia's programme for letting other companies' chips use its high-speed interconnect. Rebellions will build rack-scale systems on Nvidia's MGX reference rack design, and SK Telecom plans to install them in its AI data centre. The companies expect the R200 systems to be ready for deployment in 2028. Korea's Ministry of Science and ICT is backing the project.

The deal gives a rival accelerator a role inside an Nvidia system instead of replacing it. For Nvidia, NVLink Fusion is a way to keep its interconnect and rack designs at the centre of AI data centres even when customers want other chips. For Korea, it offers a way to put a domestic chip into national AI infrastructure without leaving the Nvidia-based rack ecosystem that suppliers are already building for.

The Division of Labour

The technical idea is disaggregated inference. A language model answers a request in two stages. First it reads the whole prompt in one pass, a stage called prefill. Then it writes the response one token at a time, a stage called decode. The two stages load hardware differently. Prefill is mostly raw computation, while decode depends more on how fast model data can be read from memory. The companies said Nvidia GPUs would handle prefill and Rebellions NPUs would handle decode, with the aim of improving throughput and power efficiency.

No performance figures have been published. Rebellions has not released specifications for the R200, and its current listed chip is the Rebel100, according to StorageReview. StorageReview also reported that no sampling schedule, early-access programme or pricing has been announced.

SK Telecom will run its own foundation models on the systems, including A.X K2, according to StorageReview. It will also run public-facing services developed under the government's "AI for All" project, Nvidia said. The work builds on SK Telecom's wider partnership with Nvidia.

Who Backs Rebellions

According to Nvidia's announcement, Rebellions' investors include Aramco, Arm, Kindred Ventures, KT, Mirae Asset Group, Samsung, SK Hynix and SK Telecom. Aramco's stake links the Korean company to the Gulf, where governments are spending heavily on national AI computing.

The four parties each gave a statement. "Sovereign AI gives nations the power to turn their language, culture and expertise into intelligence," said Jensen Huang, Nvidia's founder and chief executive. Sunghyun Park, Rebellions' cofounder and chief executive, said: "Today's announcement marks a turning point, opening a new chapter for sovereign AI." Jung Jai-hun, SK Telecom's president and chief executive, said the work "validates the potential of next-generation sovereign AI infrastructure". Bae Kyung-hoon, Korea's Deputy Prime Minister and Minister of Science and ICT, said: "The government will actively support this collaboration to deliver tangible results."

It is one of several Korean moves by Nvidia this month. The chipmaker has also backed the Korean startup Point2 and plans to take stakes in Upstage, Motif and RLWRLD.

The mixed racks are a plan for now. The first test of the split design will come when Rebellions shows working R200 hardware and data on how the systems perform against all-GPU racks before the 2028 deployment.