NVIDIA, South Korean chip designer Rebellions and SK Telecom announced a collaboration on Monday, Nov. 9, Seoul time, to link Rebellions' next-generation R200 neural processing unit (NPU) to NVIDIA's NVLink Fusion fabric, with NVIDIA GPUs handling prompt processing and Rebellions NPUs generating tokens.

That split depends on attention state computed on one vendor's chip reaching another vendor's chip fast enough, and the release names neither the transfer method nor R200's memory capacity, the spec that decides whether a 688-billion-parameter model such as SK Telecom's A.X K2 fits.

Rebellions says R200 systems are expected to be ready for deployment in 2028. South Korea's Ministry of Science and ICT (MSIT) is supporting the effort, and the release gives no numeric specifications for R200.

NVIDIA, Rebellions and SK Telecom Plan NVLink Fusion Racks Built on MGX

NVLink Fusion is NVIDIA's program for connecting non-NVIDIA processors, which NVIDIA calls XPUs, to its scale-up network. Under the plan, R200 chips will connect within a rack over NVLink, and Rebellions will build its rack-scale systems on NVIDIA's MGX architecture.

NVIDIA's release says MGX gives Rebellions access to an established supply chain of server makers and component suppliers, and gives SK Telecom a rack footprint, power delivery and cooling design shared across systems.

The release describes R200 as an inference chip for low-latency, energy-efficient serving of frontier models, including multitrillion-parameter mixture-of-experts (MoE) models. NVIDIA says NVLink Fusion improves performance and cuts deployment risk, but those are vendor claims, and no independent test results accompany them.

Why Does GPU Prefill Plus NPU Decode Need KV-Cache Handoff?

Inference has two phases with opposite bottlenecks. Prefill, which processes the prompt, is compute-heavy, while decode, which emits tokens one at a time, is limited by memory bandwidth. NVIDIA's own answer is its Rubin CPX prefill part, which uses GDDR7 memory, with HBM-equipped GPUs reserved for decode. What is new here is that the decode half would run on a third-party chip.

Prefill writes a key-value (KV) cache, the stored attention state for every token in the prompt and decode reads it for every token it generates. When two different chips split the phases, that cache has to cross between them before the first token appears, so handoff speed sets how quickly users see a response. That is an inference from how disaggregated serving works, not a figure from the companies.

NVIDIA's Dynamo serving framework already moves KV cache between GPU workers using its NIXL transfer library over InfiniBand or RDMA, according to its recipe for A.X K2. That recipe targets 8 or 12 B200 GPUs, or 32 H200 GPUs, not NPUs. The announcement does not say whether R200 will work with NIXL or Dynamo.

On the hardware side, NVIDIA says custom XPUs reach NVLink through a UCIe-to-NVLink bridge chiplet it supplies. NVIDIA says NVLink 6 connects 72 XPUs at 3.6 TB/s per XPU, with 260 TB/s across a 72-accelerator domain, and claims latency 3x lower than off-the-shelf Ethernet alternatives. Rebellions' current Rebel100 already joins four chiplets over UCIe-Advanced, though whether R200 takes the same route is undisclosed.

A.X K2 Weights Alone Would Fill Five Rebel100-Class Packages at FP8

A.X K2 has 688 billion total parameters and 33 billion active per token, and SK Telecom's released checkpoint uses block-scaled FP8, an 8-bit number format. Only 33 billion parameters run per token, but the router chooses among 256 routed experts, and because any token can be routed to any expert, serving in practice means keeping every expert loaded. That makes memory capacity, not compute, the likely binding constraint, an engineering inference rather than a company statement.

Because R200 memory is unknown, Rebel100 is the nearest proxy. The model card puts the FP8 weights at about 647 GiB, roughly 695GB, and Rebellions lists 144GB of HBM3E per Rebel100 package, so the weights alone need at least five packages (720GB) before any KV cache or activations. NVIDIA's Dynamo recipe also uses an NVFP4 checkpoint, a 4-bit format, on B200 GPUs, and whether Rebellions silicon supports that format is not stated.

Rebellions' efficiency claims remain vendor-reported. Jon Peddie Research says independent benchmarks remain thin for Rebel100 and presents the 3.2x tokens-per-second-per-watt edge over H100 as a vendor claim.

NVLink Fusion Remains Either-Or: Every Node Still Needs NVIDIA Silicon

NVLink Fusion is only semi-open. ServeTheHome described NVLink Fusion at its launch as an either-or technology that still puts NVIDIA silicon in every node, with NVIDIA keeping NVLink itself proprietary. UALink has released version 2.0, a competing open standard for linking accelerators.

South Korea's science minister and deputy prime minister, Bae Kyung-hoon, said in the release that Korean-developed NPUs can now enter the computing infrastructure market through NVLink Fusion. A route through NVLink Fusion is not independence from NVIDIA. Sovereign AI has no agreed definition, and one survey of the concept notes that sovereignty does not require domestic production of every component, though dependence on global supply chains remains a concern.

SK Telecom Already Runs Rebellions ATOM and Plans 2027 AI Factory on Vera Rubin

SK Telecom plans to deploy the systems in its AI data center and connect them to its proprietary models, including A.X K2, and to public services under Korea's "AI for All" project. Rebellions' ATOM-Max chips already run four SK Telecom AI services, including call summarization and fraud detection, according to RCR Wireless, which cited company-reported figures.

Separately, NVIDIA said on July 24 that SK Group and NVIDIA had signed letters of intent to formalize an AI factory of up to 2 gigawatts in Korea, built on NVIDIA's DSX platform with Vera Rubin GPUs and a first facility planned for 2027. That announcement does not mention NPUs.

Bloomberg Reported Acquisition Talks in August; Announcement Mentions No Investment

Bloomberg reported early-stage talks on Aug. 21 after NVIDIA CEO Jensen Huang met Rebellions CEO Sunghyun Park at NVIDIA's Santa Clara headquarters, with a technical partnership, an investment or a possible acquisition under discussion. EE Times Japan noted that neither company commented on the report. Monday's announcement describes a collaboration and mentions no investment or acquisition, which leaves open whether it is the whole arrangement or a first step.

Rebellions raised $400 million in March in a pre-IPO round at a reported $2.34 billion valuation. Seoul's posture has also shifted: an earlier plan reserving up to 50% of a national AI compute center's chips for domestic NPUs was dropped in a 2025 tender rewrite, leaving Korean NPUs to win placement voluntarily. NVIDIA said in October 2025 that Korea's government and industry would add over 260,000 NVIDIA GPUs.

R200 Specs, Handoff Benchmarks and Delivery Dates Remain Undisclosed

Readers tracking this plan should watch four things: R200 memory per package, measured KV-cache handoff latency between GPU and NPU racks, any announced Dynamo or NIXL support for R200, and independent benchmark results. The release also gives no SK Telecom delivery timeline, capacity figures or commercial terms. The announcement is a vendor-issued press release with forward-looking statements, and the companies warn actual results may differ materially.

Frequently Asked Questions

What is NVLink Fusion?

NVLink Fusion is NVIDIA's program for connecting custom CPUs and XPUs to its NVLink scale-up network and MGX rack architecture.

What are prefill and decode in AI inference?

Prefill processes the user's prompt and is compute-heavy, while decode generates the reply one token at a time and is limited by memory bandwidth.

Does NVLink Fusion make Korea's NPUs independent of NVIDIA?

Not at the interconnect layer. ServeTheHome reported that adopters must still include NVIDIA silicon in each node, and NVIDIA keeps NVLink proprietary.

When will R200 systems be available?

Rebellions says R200 systems are expected to be ready for deployment in 2028.