Nvidia, Rebellions, SK Telecom Build Korean Sovereign AI

Korea is not trying to build sovereign AI without Nvidia. It is trying to build it with Nvidia, on Korean chips, inside Korean data centres.

Nvidia, Rebellions and SK Telecom announced the collaboration on November 9 in Seoul. Rebellions will plug its next chip, the Rebel R200, into Nvidia platforms. SK Telecom will deploy the finished systems and run its own models and services on them.

The Announcement

A neural processing unit, or NPU, is a chip built only for AI work. Rebellions makes NPUs in Korea, and its current flagship is the Rebel100, built for inference, which is the work of running a trained model to produce an answer.

The R200 is the next chip in that line. The companies say it is being built for models with several trillion parameters, including mixture of experts designs, where only part of a model works on each request. Those are design goals for a chip still in development, not measured results from a shipped product.

NVLink Fusion is the link that makes the deal work. It lets chips from other makers connect to Nvidia scale up networking, the fast wiring that lets many processors in one rack act as one machine. Rebellions will also use the Nvidia MGX rack design, a standard frame that shares the rack footprint, power delivery and cooling across GPU and NPU systems.

“Sovereign AI gives nations the power to turn their language, culture and expertise into intelligence.”

Jensen Huang, founder and CEO of Nvidia, in the November 9 announcement

Important Details

Reporting in Seoul said the three firms signed a memorandum of understanding at the Nvidia AI Day event at COEX. A memorandum states intent and a work plan. It is not the same thing as a purchase order, and the release lists no order quantity and no contract value.

Canada made a different version of this bet last month, backing Cohere with public equity to keep a frontier model company at home. Dirty Static covered that deal in Canada Backs Cohere With Up to C$1.6 Billion. Korea is betting on silicon and racks instead of a model company alone.

How the Work Gets Split

The most interesting part of the plan is not the rack. It is the split inside the rack.

When a model answers a question, it does two kinds of work. First it reads the prompt and builds context, a stage called prefill. Then it writes the answer one token at a time, a stage called decode. The two stages reward different hardware, and one chip does not need to be best at both.

The companies plan to test exactly that split. Nvidia GPUs would handle prefill, and Rebellions NPUs would handle decode. If the split works in practice, SK Telecom could run large models with better throughput and lower power use than a rack built from GPUs alone. The release presents that gain as a goal of the work, and it publishes no benchmark for the combined system.

That caution matters because inference is where AI bills pile up. Training happens once. Inference happens every time a user asks a question, and SK Telecom already runs public services where that cost lands every day. A cheaper decode stage, if it holds up, is real money at that scale.

What’s Next

Confirmed steps are narrow. Rebellions has to finish the R200, connect it through NVLink Fusion, and build rack scale systems on the MGX design. SK Telecom then has to test those systems in its data centres and connect them to A.X K2 and its AI for All services.

The ministry says this is the start of wider work between Nvidia and Korean NPU firms, with Rebellions and SK Telecom first. That framing gives the deal value beyond one vendor. If the Fusion route works for Rebellions, other Korean chip firms have a proven path into Nvidia racks instead of a wall around them.

Open questions are the ones the release does not answer. How many systems will SK Telecom buy, what will they cost, and how will the split design perform against a rack of Nvidia GPUs alone? Until the partners publish tests from real hardware, the honest read is a credible plan with a long clock. Korea has chosen its partners and its wiring. The silicon still has to arrive.

Facts checked November 9, 2026.

Sources

• Nvidia, Rebellions and SK Telecom via Globe Newswire: NVIDIA, Rebellions and SK Telecom Announce Collaboration to Advance Sovereign AI Infrastructure With NVIDIA NVLink Fusion
• Seoul Economic Daily: Rebellions Joins Nvidia Ecosystem to Target Global AI Inference Market
• Hero and body images: generated house style illustrations for Dirty Static. They are illustrations of the rack concept, not photographs of R200 hardware.