Korea’s Rebellions gets an NVLink Fusion slot, but its NPU racks arrive in 2028

Nvidia, Korean chip startup Rebellions and carrier SK Telecom announced a collaboration to build Korean sovereign AI infrastructure on NVLink Fusion. Rebellions’ Rebel R200 NPUs will plug into Nvidia’s platform, with systems expected to be ready for deployment in 2028.

Nvidia GPUs for prefill, Korean NPUs for decode

According to Nvidia, Rebellions will use NVLink Fusion to connect its R200 neural processing units to the Nvidia platform. It will also use the Nvidia MGX rack architecture to develop rack-scale systems. Inside a rack, the NPUs will be linked through Nvidia’s NVLink scale-up networking.

Nvidia describes the R200 as built for low-latency, energy-efficient inference on frontier models, including mixture-of-experts models with several trillion parameters. The announcement gives no performance figures, memory size or power numbers for the chip.

The three companies also plan a split approach to inference called disaggregated inference. Nvidia GPUs would handle prefill, the stage where the model reads the prompt, while Rebellions NPUs would handle decode, the stage that generates the answer token by token. The stated goal is better throughput and power efficiency. No measured gains were published.

SK Telecom intends to deploy the resulting systems in its AI data center. They would run alongside its own models, including A.X K2, and public-facing services under Korea’s “AI for All” project. The collaboration is supported by Korea’s Ministry of Science and ICT (MSIT) and begins with Rebellions and SK Telecom.

A companion Nvidia blog post puts the deal in a wider program with MSIT covering chips, data centers, talent, startups, models, agents and security. It says the groundwork is laid for other Korean NPU and XPU companies to explore joining the NVLink Fusion ecosystem. For AI for All, Nvidia says it will work with three selected consortia and four Korean foundation model builders, using its Dynamo software to optimize and scale inference.

Who gets anything today, and who does not

Nobody can buy or test this yet. The only date in the announcement is 2028 for deployment-ready R200 systems, and no pricing, volumes or regional availability were given. The 2028 date is also a stated expectation rather than a commitment.

For customers, the practical change is architectural. NVLink Fusion lets a non-Nvidia accelerator sit inside Nvidia’s rack design instead of forcing an operator to choose between separate stacks. Korean buyers get a domestic chip option inside an Nvidia-shaped system. Korea’s Deputy Prime Minister Bae Kyung-hoon said the collaboration “enables Korean-developed NPUs to enter the computing infrastructure market alongside NVIDIA”.

Nvidia’s own framing is about control over national AI capability. Jensen Huang said “Sovereign AI gives nations the power to turn their language, culture and expertise into intelligence.” The claims about efficiency, latency and trillion-parameter readiness are the companies’ own and have not been independently verified.

Signals that will show if the 2028 plan holds

The first thing to watch is whether the R200 turns into hardware with published specifications. Rebellions’ current flagship, the Rebel100, is described by trade press as a chiplet design with UCIe interconnects and HBM3E memory, according to The Next Web, which also gave no performance specifications for it. The R200 would be the successor to prove itself in rack-scale form.

The second is funding and customers. The Next Web reported that Rebellions closed a $400 million pre-IPO round on March 30, 2026 at a valuation of about $2.34 billion, bringing total capital raised to $850 million, with the Korea National Growth Fund contributing 250 billion won. That is a sizable cushion for a company that must deliver rack-scale systems, but it is small beside the scale of the GPU racks it would sit next to. Those racks bring their own power and cooling demands, which we covered in our look at why GPU racks need liquid cooling and so much power.

Third, watch who else joins. Nvidia says other Korean NPU and XPU makers may explore the ecosystem, but names none. If a second vendor commits, NVLink Fusion starts to look like a platform rather than a one-off partnership. Rebellions’ backers already include SK Telecom, Samsung, SK Hynix, Arm and KT, according to Nvidia’s description of the company, so the Korean supply chain is closely tied to the project.

What the announcement does not establish is any commercial term between the three companies, how many R200 systems SK Telecom will buy, or how workloads will be split in production. Until those appear, the deal is a roadmap with government backing and a 2028 target.