Nvidia, SK Telecom and Korean chip startup Rebellions signed a three-way memorandum of understanding on Monday to build a heterogeneous AI computing system that links Nvidia GPUs with Rebellions-designed NPUs inside SK Telecom's AI data centers. The deal was signed at Nvidia AI Day Seoul 2026, held at COEX in Seoul's Gangnam District, and marks the first time Nvidia has formalized cooperation with a Korean NPU company.
Under the agreement, Rebellions will connect its next-generation R200 NPU to Nvidia's NVLink Fusion and jointly develop an integrated GPU-NPU system with Nvidia. Rebellions aims to deploy R200 systems in 2028.
NVLink Fusion is the Nvidia program that lets hyperscalers and custom chip designers plug their own processors, whether CPUs, GPUs or NPUs, into Nvidia's NVLink-based rack-scale infrastructure. Nvidia unveiled it in May of last year, and Rebellions follows Marvell and MediaTek as a partner.
Key facts
- 01Nvidia, SK Telecom and Rebellions signed a three-way NVLink Fusion MOU at Nvidia AI Day Seoul 2026 on Nov. 9.
- 02Rebellions will connect its next-generation R200 NPU to Nvidia's NVLink Fusion and co-develop a GPU-NPU system.
- 03Nvidia GPUs will handle the prefill stage of inference while Rebellions NPUs handle token-by-token decode.
- 04Rebellions aims to deploy R200 systems in 2028, starting in SK Telecom's AI data centers.
- 05Nvidia also said it will invest in four Korean startups: Point2 Technology, Upstage, Motif Technologies and RealWorld.
The technical pitch is disaggregated inference. Nvidia GPUs will take the compute-heavy prefill stage, where a model processes the incoming prompt, while Rebellions' more power-efficient NPUs take the token-by-token decode stage. The partners say the split is meant to raise throughput and power efficiency and lower total cost of ownership. The R200 is tuned for low-latency inference on frontier models, including mixture-of-experts models with trillions of parameters.
Rebellions will build its systems on Nvidia's MGX standard server rack design, so GPU and NPU hardware can share rack space, power supply and cooling equipment. NVLink Fusion will wire the NPUs in each rack directly into Nvidia's high-speed scale-up network.
SK Telecom will stand up the combined Nvidia-Rebellions system at its AI data centers to verify performance and stability, then put it to work in live environments. The carrier plans to connect its own A.X K2 model and public AI services built under the government's AI for All project, with the stated goal of a full-stack sovereign AI framework. The partners describe the end product as a Sovereign AI Token Factory serving the Korean language, local data and domestic industry demand.
South Korea's Ministry of Science and ICT is backing the partnership and said it could broaden applications for domestic advanced semiconductors. The ministry plans to extend the same model to Korean NPU makers beyond Rebellions, linking technology development, pilot verification, commercialization and exports. Ryu Je-myung, second vice minister of science and ICT, called the arrangement an important opportunity for domestically developed NPUs to enter the computing infrastructure market alongside Nvidia.
Rebellions CEO Park Sung-hyun framed the deal in national terms. He said the company would help Korean-made technology grow into a core pillar that drives the world's AI.
“This announcement is an important turning point in expanding South Korea's sovereign AI ecosystem”— Park Sung-hyun, CEO of Rebellions
Nvidia paired the MOU with investment news. The company said it will put money into four Korean startups: Point2 Technology, which makes RF wireless interconnect components and already counts NVentures as an investor; large language model developers Upstage and Motif Technologies; and RealWorld, which builds robot foundation models. The investments in Upstage, Motif and RealWorld will come directly from Nvidia rather than through NVentures, and those procedures have not yet been completed.
Deepu Talla, Nvidia's vice president of robotics and edge AI, opened the event by casting Korea as a central player in physical AI over the next decade, pointing to Hyundai Motor Group's virtual factories and a Doosan robot being trained in Nvidia simulation software to polish car door surfaces. The ministry and Nvidia also hosted an AI startup forum attended by 22 Korean companies, including Upstage, TwelveLabs, FuriosaAI, DeepX and Rebellions.
No financial terms were disclosed for the MOU. Actual cost and power savings from the GPU-NPU split will depend on system configuration and operations, and Rebellions expects to begin supplying the integrated systems in 2028.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
