NVIDIA opens NVLink Fusion to Korean NPUs, with GPUs for prefill and Rebellions chips for decode

NVIDIA, Korean chip startup Rebellions and SK Telecom announced on Monday that Rebellions will connect its next-generation R200 neural processing units (NPUs) to NVIDIA's platform through NVLink Fusion, the program that lets other companies' custom processors sit on NVIDIA's scale-up interconnect. SK Telecom plans to deploy the resulting rack-scale systems in its own AI data centers. Korea's Ministry of Science and ICT (MSIT) is backing the effort. The announcement came with several smaller NVIDIA posts tied to the same push in Korea (NVIDIA release).

The most concrete technical detail is how the three companies want to split the work. They say they plan to build disaggregated inference in which NVIDIA GPUs handle prefill, the compute-heavy pass over the prompt, and Rebellions NPUs handle decode, the token-by-token generation that is usually limited by memory bandwidth. The stated aims are higher inference throughput and better power efficiency. These are plans, though. The release includes no benchmark, no cost figure and no performance target.

What each company brings

• Rebellions will use NVLink Fusion to link R200 NPUs inside a rack over NVLink, and will build systems on NVIDIA's MGX rack architecture. It says the R200 targets low-latency, energy-efficient inference for frontier models, including multitrillion-parameter mixture-of-experts models.
• SK Telecom will deploy the systems and connect them to its own models, including A.X K2, and to public services built under the government's *AI for All* project.
• NVIDIA supplies the interconnect, the reference rack designs and access to the MGX supplier ecosystem. The company says MGX lets GPU and NPU systems share rack footprint, power delivery and cooling designs.

The timeline matters more than the headline: NVIDIA's release says R200 systems are expected to be ready for deployment in 2028. Nothing here ships soon. It is a two-year integration plan, not a product launch.

The rest of NVIDIA's Korea package

In a companion post, NVIDIA says its expanded work with MSIT lays groundwork for other Korean NPU and XPU makers to explore joining the NVLink Fusion ecosystem. The post also covers training and certification programs, and work with three consortia and four Korean foundation-model builders under *AI for All*, using NVIDIA's Dynamo inference software (NVIDIA blog).

NVIDIA also published Dynamo-optimized deployment recipes for four Korean open models: SK Telecom's A.X K2, LG AI Research's K-EXAONE 2.0, Upstage's Solar Open 2 and Motif Technologies' Motif-3. NVIDIA claims the recipes deliver "up to 2x" performance out of the box on B200 GPUs. It does not say what that is measured against, and the figure comes only from NVIDIA (NVIDIA blog).

At an MSIT-NVIDIA startup forum at COEX in Seoul, NVIDIA announced planned investments in Motif Technologies, Upstage and robotics startup RLWRLD. It also listed an investment in interconnect startup Point2 Technology that it had announced earlier. It did not disclose any amounts (NVIDIA blog).

Why this is worth watching

NVLink Fusion is NVIDIA's answer to a market that increasingly wants custom silicon. Instead of losing those sockets, NVIDIA offers to be the fabric they plug into. A Korean NPU startup putting its flagship inference chip on NVIDIA's interconnect, rather than competing on a separate stack, shows that strategy working. It also keeps NVIDIA GPUs in the rack for the prefill stage.

For developers, the prefill/decode split is the part to watch. Long prompts, large contexts and agent loops that resend history all lengthen prefill, while chat-style output is bound by decode. If dedicated decode chips do lower the cost per generated token, providers could eventually price input and output tokens even more differently than they do today. That is a possible consequence, not something the companies have claimed, and it depends on hardware that is two years out.

Context: the announcements landed the day after NVIDIA's board said it was opening an independent investigation following reports that CEO Jensen Huang had been the target of a foreign intelligence effort (NVIDIA board statement). The board said it had no reason to believe national security or NVIDIA secrets were compromised. Monday's Korea releases do not mention it.