NVIDIA opens NVLink Fusion to Rebellions NPUs in SK Telecom sovereign AI push

Rebellions to connect R200 NPUs to NVLink Fusion

NVIDIA, Korean chip designer Rebellions and carrier SK Telecom announced a collaboration on 8 November 2026 to build AI infrastructure that combines NVIDIA's platform with Korean accelerators. According to dev.to, which examined NVIDIA's newsroom release, Rebellions will attach its next-generation R200 NPU to NVLink Fusion — NVIDIA's scheme for letting other chipmakers plug their processors into its high-bandwidth scale-up networking — and will build systems on the MGX rack architecture. SK Telecom intends to run the hardware in its AI data centre, with support from Korea's Ministry of Science and ICT.

The interconnect is the notable part. Scale-up networking, the layer that shuttles data between accelerators inside a rack, has long been among the most closely guarded pieces of NVIDIA's stack, and licensing it to a third-party accelerator maker effectively turns a would-be rival into a component of an NVIDIA-defined system. The chipmaker gains a route into deployments already on customers' roadmaps plus an established manufacturing supply chain; NVIDIA keeps ownership of the interface and its roadmap.

Who supplies which layer

The announcement divides the stack neatly. Rebellions provides the R200, an inference NPU built for fast, energy-efficient work on frontier-scale models — among them mixture-of-experts designs with trillions of parameters — alongside a software stack the company says rests on open source technology and open standards. SK Telecom provides the facility, its own foundation models such as A.X K2, and public services created under Korea's AI for All project. NVIDIA provides the fabric, the rack reference designs and its ecosystem of manufacturers and component suppliers.

Splitting inference across two chip types

The core technical plan is disaggregated inference. As dev.to describes it, the companies want to assign each stage of response generation to the silicon that fits it: NVIDIA GPUs would run prefill, which processes an entire prompt at once and behaves like very large matrix arithmetic, while Rebellions NPUs would run decode, which emits tokens one at a time and is usually constrained by memory bandwidth. Matching stages to hardware should lift throughput and power efficiency, but dev.to stresses that the release describes intended work rather than shipped product.

Hardware in 2028, decisions now

Rebellions expects the R200 systems to be ready for deployment in 2028, and that gap shapes how the deal reads. Data centres are planned around power, cooling and floor space long before equipment arrives, and under MGX the GPU and NPU systems share a physical envelope and draw on the same power and cooling infrastructure. Picking a rack specification early therefore decides which suppliers can bid for the building — a procurement commitment made years before anyone can benchmark the chip.

Sovereign AI on NVIDIA's terms

The ministry's backing turns a vendor deal into a national programme. dev.to reports that a deputy prime minister's statement framed the effort as Korean-developed NPUs entering the computing infrastructure market alongside NVIDIA, with government support spanning physical AI and GPU supply. The tension dev.to's analysis highlights is structural: the chips and the models are domestic, but the interconnect and the rack specification are NVIDIA's, so the "sovereign" stack stays compatible only at the layers NVIDIA controls — and follows whatever roadmap NVIDIA pursues.

Why it matters

The deal shows NVIDIA choosing to monetise its interconnect lead by licensing it outward rather than holding it exclusively. Folding national accelerator programmes into MGX racks extends NVIDIA's platform into procurements that might otherwise have gone entirely elsewhere, while giving Rebellions a credible path to deployment at the price of building on a competitor's interface. For other countries pursuing sovereign AI, it sketches the template on offer: domestic silicon and domestic models, hosted in racks whose critical specification NVIDIA still writes.