NVIDIA NVLink Fusion: Rebellions R200 NPUs join SK Telecom racks, systems due 2028

NVIDIA, the Korean AI-chip designer Rebellions and the telecom group SK Telecom (SKT) announced on 8 November that Rebellions will use NVIDIA's NVLink Fusion technology to connect its next-generation Rebel R200 neural processing units (NPUs) to NVIDIA's rack-scale platform, with SKT planning to run the resulting systems in its own AI data center. It is a notable change of tone for a company that, as recently as August, was reported to be weighing an investment in or even a purchase of Rebellions: for now, the announced deal is a technical partnership, backed by Korea's Ministry of Science and ICT (MSIT), and the hardware is not due until 2028.

Key points

• NVIDIA, Rebellions and SK Telecom announced on 8 November 2026 that Rebellions will use NVIDIA NVLink Fusion to link its Rebel R200 NPUs to NVIDIA's platform and build rack-scale systems on NVIDIA's MGX rack architecture.
• SK Telecom plans to deploy these systems in its AI data center to serve its own models (including A.X K2) and services under Korea's "AI for All" project.
• The companies plan to develop disaggregated inference: NVIDIA GPUs for the prefill stage and Rebellions NPUs for the decode stage. NVIDIA describes this as planned work, not a shipping product.
• Rebellions' R200 systems are expected to be ready for deployment in 2028, according to NVIDIA's release.
• On the same day NVIDIA published deployment recipes for four Korean open models and a set of Korean physical-AI and agent-security tie-ups with Samsung, LG, Kakao Mobility and Hyundai.

What happened

According to NVIDIA's newsroom release, the three companies will combine NVIDIA's open infrastructure platform, Rebellions' Korean-designed processors and SKT's data-center, model and service capabilities. The technical core is NVLink Fusion, which NVIDIA describes as extending its platform "to custom XPUs": a chipmaker keeps its own silicon but attaches it to NVIDIA's high-bandwidth scale-up networking, so that the custom chips and NVIDIA compute can sit in the same rack and run workloads together. Rebellions will use it to connect its R200 NPUs within a rack over NVIDIA NVLink scale-up networking, and will build its systems on the NVIDIA MGX rack architecture and its supplier ecosystem.

Rebellions describes the R200 as optimised for low-latency, energy-efficient inference with frontier models, including multitrillion-parameter mixture-of-experts models. SKT plans to deploy the systems in its AI data center, linking them with its proprietary foundation models, named in the release as including A.X K2, and with public-facing services developed under the government's "AI for All" project. The MGX architecture, NVIDIA says, lets GPU and NPU systems share data-center design elements such as rack footprint, power delivery and cooling, which should simplify infrastructure planning.

The most concrete technical idea is the split of the work in each AI response. The companies plan to develop what the release calls disaggregated inference: NVIDIA GPUs would handle prefill, and Rebellions NPUs would handle decode, to improve throughput and power efficiency. Prefill is the stage where the model reads the whole prompt in one parallel pass; decode is where it generates the answer token by token. The two stress hardware differently (prefill is compute-heavy, decode leans on memory bandwidth and latency), which is why mixing chip types is attractive on paper. No performance numbers were published, so the benefit is a hypothesis, not a result.

On availability, the release says Rebellions' R200 systems are expected to be ready for deployment in 2028. The release carries the usual forward-looking-statement warnings from both NVIDIA and SKT, which note that actual results may differ.

Korea's Deputy Prime Minister and science minister Bae Kyung-hoon said the collaboration "enables Korean-developed NPUs to enter the computing infrastructure market alongside NVIDIA", and that the government will support it, including cooperation on physical AI and on "securing GPU supply". NVIDIA CEO Jensen Huang said NVLink Fusion lets Korea's chipmakers connect their custom ASICs to NVIDIA's AI factory platform.

The wider Korea package

The Rebellions news was one of five NVIDIA posts about Korea on the same day.

• Policy and talent. NVIDIA and MSIT are expanding their collaboration from semiconductors and data centers to talent, startups, models, agents and security. NVIDIA says the work lays the groundwork for Korean NPU and XPU companies to explore joining the NVLink Fusion ecosystem, and includes training and certification programmes, hackathons and technical seminars. Under "AI for All", NVIDIA will work with three selected consortia and four Korean foundation-model builders, using its Dynamo inference software.
• Open-model recipes. NVIDIA released Dynamo-optimised deployment recipes for A.X K2 (SK Telecom), K-EXAONE 2.0 (LG AI Research), Solar Open 2 (Upstage) and Motif-3 (Motif Technologies). They target NVIDIA B200 GPUs, and NVIDIA says they deliver up to 2x performance out of the box. That is a vendor figure for its own software on its own hardware; we have not seen an independent test.
• Agent security. Hyundai Motor Group is exploring NVIDIA's NemoClaw blueprints and the OpenShell runtime, which provides sandboxed execution and policy-based controls for AI agents. The Korean security firm AIM Intelligence is integrating its Starfort platform with OpenShell. LG AI Research is opening API access to K-EXAONE-2.5, built using NVIDIA's Nemotron datasets, including Nemotron-Personas-Korea, an open dataset of 7 million fully synthetic Korean personas grounded in official Korean statistics.
• Physical AI. Samsung says selected quantum-chemistry calculations in its sMOL materials platform ran up to 5x faster end to end with NVIDIA's cuEST library, and that an AI-assisted homogenisation case cut a simulation from about 18 hours (a fine-grid solve on a 32-core CPU) to 41 seconds, roughly a 1,600x speedup. Kakao Mobility plans to build a level 4 autonomous-driving system on NVIDIA's Hyperion platform with DRIVE AGX Thor, starting with Kia purpose-built vehicles. LG Electronics and NVIDIA demonstrated GR00T-powered manipulation with LG's CLOiD robot in a real manufacturing environment.

All of these figures come from NVIDIA's own posts and describe results from the companies involved.

Background

Rebellions is based in Bundang, south of Seoul, and was founded in 2020. It designs NPUs for data-center inference, meaning the work of running an AI service once the model exists. According to The Next Web's report of August 2026, its Atom and Atom Max chips entered mass production in 2023, the company merged with Sapeon Korea (a chip business SKT spun out) in 2024, it had raised roughly $850 million from backers including SK Hynix, Samsung Ventures and Arm, and its latest valuation was around $2.3 billion. NVIDIA's own boilerplate lists Aramco, Arm, Kindred Ventures, KT, Mirae Asset Group, Samsung, SK Hynix and SK Telecom among Rebellions' backers, and calls its flagship Rebel100 a chip built specifically for inference.

On 21 August, The Next Web reported, citing Bloomberg, that NVIDIA was in early talks with Rebellions about a technical partnership, an investment or possibly an acquisition, after Huang met Rebellions CEO Sunghyun Park at NVIDIA's Santa Clara headquarters. The talks were described as preliminary, and the sources were unnamed. The announcement of 8 November is the technical-partnership outcome; neither NVIDIA's release nor the other posts mention an investment or acquisition, and we have found no confirmation of one.

The same piece describes a pattern in which NVIDIA licenses technology and hires engineers rather than buying companies, citing its arrangement with Groq and the NVIDIA Groq 3 inference chip. NVLink Fusion is a different mechanism: it keeps the partner's chip in the partner's hands and sells the interconnect and rack design around it. For a company whose GPUs dominate AI training, it is a way to stay at the center of racks even when part of the silicon is someone else's.

For Korea, the announcement fits the national push we covered in South Korea's 2027 budget for AI chips, data centers and physical AI.

How the split could work

The release describes the inference split in one sentence, so it helps to spell out the idea.

CPU, fine grid (18 h)64800

GPU, coarse grid41

Samsung simulation time, fine-grid CPU solve vs GPU-accelerated coarse-grid solve (seconds)

	seconds
CPU, fine grid (18 h)	64800
GPU, coarse grid	41

Source: NVIDIA blog, Korea Advances Physical AI Across Industries With NVIDIA

The chart shows the Samsung simulation result NVIDIA quotes (18 hours is 64,800 seconds), which is a physical-AI example rather than an inference one; there are no published prefill or decode numbers for the Rebellions plan.

For inference, think of a chat request in two halves. In prefill, all of the user's input is processed at once, which keeps GPUs busy. In decode, the model emits one token at a time and each step must read large amounts of model state, so memory bandwidth and latency matter more than raw arithmetic. A design that sends each half to the chip best suited to it can, in principle, raise throughput per watt. The practical obstacles are moving the intermediate state between the two kinds of chips fast enough, and software that schedules across both. NVLink Fusion and MGX are NVIDIA's answer to the hardware side, and its Dynamo software is aimed at the routing and communication side.

What it means

• For Korean chipmakers: a route into data centers that are standardising on NVIDIA racks, with the minister explicitly framing it as letting Korean NPUs enter the market alongside NVIDIA. The price is dependence on NVIDIA's interconnect.
• For SK Telecom and "AI for All" model builders: a path to run domestic models on a mix of imported GPUs and domestic NPUs, and Dynamo recipes that make four Korean open models easier to deploy on B200 GPUs.
• For NVIDIA: it keeps the rack, the network and the software in its orbit even if inference workloads shift toward specialised chips. Whether that outweighs the risk of helping a competitor's silicon mature is a question the release does not address.
• For developers: nothing changes today. R200 systems are years away, and the Dynamo recipes are the only part you could use now.

What we don't know yet

• Any performance, price or power figures for the R200, or for the GPU-prefill and NPU-decode split.
• How many R200 systems SKT will buy, and in which data center.
• Whether NVIDIA will invest in Rebellions, as the August reports of early talks raised; nothing announced says so.
• Which other Korean NPU companies will join the NVLink Fusion ecosystem; NVIDIA says only that they can explore it.
• How the Dynamo recipes' "up to 2x" claim was measured and against what baseline.

What to watch next

Watch for the first technical disclosures on the R200 and on the prefill/decode design, any formal investment announcement, and which additional Korean chip designers are named under the MSIT programme. SKT's data-center plans and the AI for All consortia's use of Dynamo will show whether this moves from press release to deployed racks well before 2028.

Sources

• NVIDIA newsroom: NVIDIA, Rebellions and SK Telecom Announce Collaboration to Advance Sovereign AI Infrastructure With NVIDIA NVLink Fusion (primary)
• NVIDIA blog: From Chips to Skills, NVIDIA and Korea Expand Sovereign AI Collaboration
• NVIDIA blog: NVIDIA Dynamo Recipes Bring Korea's Leading Open Models to Production
• NVIDIA blog: Korea's AI Ecosystem Advances and Secures Agentic AI With NVIDIA
• NVIDIA blog: Korea Advances Physical AI Across Industries With NVIDIA
• The Next Web: Nvidia is in talks with the Korean inference chip designer Rebellions (21 August 2026, citing Bloomberg)
• GlobeNewswire copy of the NVIDIA, Rebellions and SK Telecom release