NVIDIA, Rebellions and SK Telecom build Korea's sovereign AI racks — R200 NPUs plug into NVLink Fusion

On Monday morning, inside the COEX convention center in Seoul's Gangnam district, three companies put their names to the same piece of paper. NVIDIA was holding its AI Day Seoul event in the building. Rebellions, the Korean AI chip startup, and SK Telecom, the country's largest carrier, signed a memorandum of understanding with NVIDIA to build sovereign AI infrastructure on Korean-designed silicon — the kind of deal that, until recently, would have sounded like a contradiction: a national AI stack built partly on NVIDIA's own interconnect.

The announcement, published Monday morning, is the hardware half of a much larger Korean AI push. The same day, NVIDIA also expanded its collaboration with Korea's Ministry of Science and ICT across chips, data centers, models, skills and agent security. But the Rebellions deal is the concrete part: a named chip, a named interconnect, a named buyer, and a date.

What happened

Rebellions will connect its next-generation Rebel R200 neural processing units to NVIDIA's platform using NVLink Fusion, and build rack-scale inference systems on NVIDIA's MGX rack architecture. SK Telecom plans to deploy those systems in its AI data center, linking them to its proprietary foundation models — including A.X K2 — and to public-facing services developed under the Korean government's AI for All project. The collaboration is supported by the Ministry of Science and ICT, which framed it as the first of several partnerships between NVIDIA and Korean NPU companies.

Roles as described in the companies' joint announcement of November 9, 2026.

The technical idea is disaggregated inference: split the two stages of generating an AI response across the processor best suited to each. NVIDIA's GPUs take prefill — the compute-heavy first pass over a prompt — while Rebellions' NPUs take decode, generating tokens one by one. The companies say the split should improve inference throughput and power efficiency. The R200 itself is designed for frontier models, including multitrillion-parameter mixture-of-experts models, and the first R200 systems are expected to be ready for deployment in 2028.

NVIDIA's chief executive framed the deal in national terms. "Sovereign AI gives nations the power to turn their language, culture and expertise into intelligence," Jensen Huang said in the release. "With NVLink Fusion, Korea's chipmakers can connect their custom ASICs to NVIDIA's AI factory platform, tightly coupling their processors with NVIDIA compute to run AI workloads." Rebellions cofounder and CEO Sunghyun Park called the announcement "a turning point, opening a new chapter for sovereign AI by connecting Korea's AI silicon and models with global systems technology."

Why it matters

The striking part is the direction of the door. NVLink Fusion is NVIDIA's way of letting outside chips into its ecosystem — and until now, the assumption in Seoul and elsewhere was that sovereign AI meant building away from NVIDIA. This deal runs the other way: Korea's homegrown NPU plugs into NVIDIA's interconnect, shares its racks, and rides its supply chain of original design manufacturers and component suppliers. Sovereignty, in this version, is about owning the silicon and the models, not about owning every layer of the stack.

It is also a pattern that is spreading. Cohere's sovereign AI deal in Singapore put models inside a national-security network; here the sovereign layer is the chip itself. And the timing is no accident: Korea's AI chip exports nearly doubled in October, and the government is backing domestic NPU makers as a strategic industry. For Rebellions, whose current flagship is the Rebel100 inference chip, the deal is a shortcut past the hardest problem a chip startup faces — not building the silicon, but getting it into real data centers. SK Telecom's data centers and MGX's established supply chain do that work.

The honest caveat is the calendar. Nothing in the announcement is a shipped product: the R200 is still under development, the systems are two years from deployment, and the agreement is a memorandum of understanding rather than a binding supply contract. There are no performance measurements, no deployment sizes and no investment figures in the release. The 2028 date is the number to hold the companies to.

What's next

The nearer-term evidence arrives in software. Under the wider MSIT collaboration announced the same day, NVIDIA will work with Korean foundation-model builders on inference optimization using its Dynamo serving software, with deployment recipes covering SK Telecom's A.X K2, LG AI Research's K-EXAONE 2.0, Upstage's Solar Open 2 and Motif Technologies' Motif-3. Those recipes are the part developers can actually use now, while the mixed GPU-NPU racks are still being built.

Watch three things. First, whether other Korean NPU companies follow Rebellions into NVLink Fusion — MSIT said this is the beginning of a broader program, not a one-off. Second, whether the 2028 deployment date holds once the R200 moves from design to silicon. And third, whether the model-serving recipes turn into real deployments on Korean infrastructure, or stay as documentation. NVIDIA announced plenty more in Seoul on Monday — including four open Cosmos 3 world models — but this is the deal that decides whether Korea's AI runs on Korean chips.