Nvidia, Rebellions and SK Telecom Team Up on Korea's Sovereign AI

NVIDIA, Korean chip designer Rebellions and SK Telecom announced on November 8, 2026 that they are collaborating to build out South Korea's sovereign AI infrastructure, using NVIDIA's open NVLink Fusion interconnect platform to link Rebellions' Korean-designed AI processors into NVIDIA-based racks run by SK Telecom.

What NVIDIA, Rebellions and SK Telecom announced

The three companies said they are combining NVIDIA's open infrastructure platform, Rebellions' domestically designed AI processors, and SK Telecom's data center, model and service operations into a single sovereign AI stack for Korea. The goal, according to the companies, is to let Korea run its own large language models and public AI services on infrastructure that is partly built on Korean-designed silicon rather than relying solely on imported accelerators.

NVIDIA founder and CEO Jensen Huang framed the effort around national AI independence: "Sovereign AI gives nations the power to turn their language, culture and expertise into intelligence," he said in the announcement.

What NVLink Fusion actually is

NVLink Fusion is NVIDIA's interconnect technology for extending its rack-scale platform to non-NVIDIA chips, letting custom XPUs such as ASICs and NPUs plug into NVIDIA's high-bandwidth, low-latency NVLink scale-up networking alongside its own GPUs. It is paired with NVIDIA's MGX rack architecture and reference designs, which give partners a standardized system blueprint plus access to an established MGX ecosystem of original design manufacturers (ODMs) and component suppliers, rather than requiring a custom rack built from scratch.

For this deal, that means Rebellions' chips do not have to replace NVIDIA GPUs outright — they can sit in the same rack footprint, sharing power delivery and cooling design with NVIDIA systems.

NVLink Fusion's ecosystem so far

NVIDIA originally unveiled NVLink Fusion in May 2025 at Computex in Taipei, pitching it as a way for the rest of the chip industry to build semi-custom AI infrastructure around NVIDIA's networking rather than against it. At launch, NVIDIA named Fujitsu and Qualcomm Technologies as CPU partners planning to pair their own custom CPUs with NVIDIA GPUs, and MediaTek, Marvell, Alchip Technologies, Astera Labs, Synopsys and Cadence as partners building custom AI silicon around the platform. Rebellions' R200 NPU is a newer, Korea-specific addition to that same ecosystem — notable less for the chip itself and more for the fact that a national telecom operator and a government ministry are the ones putting it into production.

Who is Rebellions

Rebellions is a South Korea-based company, also operating in the United States, that designs inference accelerators for enterprises and governments. Its current flagship chip is the Rebel100, which uses a chiplet architecture and ships with an open-source software stack. The company is now developing a next-generation chip, the R200 neural processing unit (NPU), which it says is optimized for low-latency, energy-efficient inference of frontier models — including mixture-of-experts models with multiple trillions of parameters. Under the new collaboration, R200 NPUs will connect to each other within a rack using NVLink Fusion. NVIDIA's announcement lists Rebellions' backers as Aramco, Arm, Kindred Ventures, KT, Mirae Asset Group, Samsung, SK Hynix and SK Telecom — a roster that mixes Korean conglomerates with global semiconductor and energy investors. That SK Telecom is both a Rebellions backer and the deployment partner in this announcement underlines how concentrated Korea's sovereign AI push is around a small set of repeat players: SK Hynix supplies memory, SK Telecom supplies data centers and models, and Rebellions supplies the processor design, with NVIDIA's interconnect tying the pieces together.

Today's production chip, the Rebel100, is pitched at inference workloads rather than training, which lines up with the broader industry split between NVIDIA's training-focused GPU lineup and a growing field of inference-specific accelerators from multiple vendors. The R200 is positioned as the chip that will actually ship into NVLink Fusion-based racks, rather than the Rebel100, since it is being built specifically with that interconnect in mind.

SK Telecom's role: data centers, models and "AI for All"

SK Telecom — a telecommunications operator that traces its history to 1984 and has a long-running collaboration with NVIDIA — plans to deploy the combined GPU-and-NPU systems in its own AI data centers. The company says it will connect that infrastructure to its proprietary foundation models, including A.X K2, and to public-facing AI services delivered under Korea's "AI for All" initiative. SK Telecom CEO Jung Jai-hun described the plan as linking the company's AI data centers with its proprietary models and "AI for All" services to strengthen a full-stack approach to sovereign AI.

NVIDIA's newsroom separately details SK Telecom's broader AI infrastructure partnership with NVIDIA, a relationship this announcement builds on rather than starts from scratch. The "AI for All" branding matters here: it's the public-facing label SK Telecom and the Korean government have used for consumer and citizen-facing AI services, distinct from the enterprise and government inference work Rebellions' chips are aimed at. By connecting A.X K2 and other proprietary models to infrastructure that includes Korean-designed silicon, SK Telecom is positioning itself as the operator that runs Korea's AI stack end to end — from the chip layer up through the model layer to the public service layer.

Why NVIDIA GPUs and Rebellions NPUs are being split by task

A specific technical detail in the announcement: the companies plan to use disaggregated inference, where NVIDIA GPUs handle the prefill stage of a model's work and Rebellions NPUs handle decode. Splitting inference this way is intended to improve throughput and power efficiency compared with running the entire inference pipeline on one chip type. The MGX rack architecture is what allows GPU-based and NPU-based systems to share the same rack footprint, power delivery and cooling design, so the disaggregated setup doesn't require two separate physical infrastructures.

Prefill and decode are the two distinct phases of running a large language model: prefill processes the prompt a user sends in, which is compute-heavy and benefits from raw GPU throughput, while decode generates the response token by token, which is more sensitive to memory bandwidth and latency per step. By routing prefill to NVIDIA GPUs and decode to Rebellions NPUs, the companies are betting that each chip type can specialize in the phase it's best suited for, rather than using one chip architecture to do both less efficiently. This disaggregated approach has been gaining traction across the inference industry more broadly, and this announcement effectively commits Rebellions' NPU roadmap to supporting it within NVLink Fusion racks.

Korean government backing

Korea's Ministry of Science and ICT (MSIT) is supporting the collaboration, starting with Rebellions and SK Telecom specifically. Deputy Prime Minister and Minister of Science and ICT Bae Kyung-hoon called Korean-developed NPUs entering the computing infrastructure market alongside NVIDIA "a significant step forward," and said the government will support the work going forward, including on physical AI and GPU supply. Rebellions co-founder and CEO Sunghyun Park said Korean technology will become "a core pillar powering AI worldwide."

Rebellions' R200-based systems are expected to be ready for deployment in 2028, meaning this week's announcement is a roadmap commitment rather than hardware shipping today. That multi-year runway is typical for custom-silicon partnerships of this scale: designing and validating a new NPU, qualifying it for NVLink Fusion compatibility, and building out MGX-based racks around it all take time, and NVIDIA's own launch partners from 2025 have likewise rolled out their NVLink Fusion-compatible products gradually rather than all at once.

Part of a broader Korea AI push this week

The NVLink Fusion collaboration arrived alongside several other Korea-focused posts from NVIDIA in the same week, pointing to a coordinated push around the country's AI buildout:

• On November 9, NVIDIA announced new planned investments in Korean AI startups at the MSIT-NVIDIA AI Startup Partnership Forum, where MSIT Vice Minister Ryu Je-myung met NVIDIA leaders alongside more than 20 Korean startups. NVIDIA said it had already invested in AI-interconnect company Point2, with planned investments in Motif Technologies, RLWRLD and Upstage.
• On November 8, NVIDIA published a post on advancing physical AI across Korean industries, covering manufacturing and semiconductor applications.
• Also on November 8, NVIDIA detailed work on securing agentic AI in Korea involving Hyundai Motor Group and AIM Intelligence.
• NVIDIA additionally described an expanded sovereign AI collaboration with MSIT covering everything "from chips to skills."
• And NVIDIA released Dynamo-optimized deployment recipes for Korea's leading open models, including SK Telecom's A.X K2 alongside LG AI Research's K-EXAONE 2.0, Upstage's Solar Open 2 and Motif Technologies' Motif-3, tuned for NVIDIA B200 GPUs.

Taken together, the week's announcements span chips, data center infrastructure, government partnership, startup funding and model deployment tooling — all aimed at Korea's AI sector specifically. The startup forum alone drew more than 20 Korean AI companies to COEX to meet with MSIT Vice Minister Ryu Je-myung and NVIDIA leadership, and the Dynamo recipes cover four distinct Korean-built 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 — tuned to run on NVIDIA B200 GPUs with up to 2x performance out of the box, according to NVIDIA. That breadth suggests NVIDIA is treating Korea less as a single customer win and more as a template it can point to when courting other countries pursuing sovereign AI strategies.

Why this matters beyond Korea

NVLink Fusion is NVIDIA's answer to a broader industry trend: large AI buyers and chip designers outside NVIDIA want custom silicon that can still plug into NVIDIA's networking and rack ecosystem, rather than building fully separate infrastructure stacks. Rebellions becomes one of the public examples of a non-NVIDIA chip designer building toward NVLink Fusion compatibility, and the fact that a national telecom operator and a national science ministry are co-signing the effort gives it political as well as technical weight. For readers tracking the broader custom-silicon trend, NVIDIA's NVLink Fusion play sits alongside other recent moves in the accelerator space, including AMD's MI450 push against NVIDIA's Vera Rubin roadmap and NVIDIA's own consumer-facing RTX Spark platform.

What happens next

The companies did not publish a near-term shipping date for combined GPU-NPU racks; the R200 NPU itself is targeted for deployment readiness in 2028, so near-term work will likely focus on software integration, NVLink Fusion compatibility testing, and SK Telecom's data center planning. Watch for further detail at future NVIDIA events — the company has previously used its GTC conference and quarterly earnings calls to update NVLink Fusion partner lists — and for any follow-up from MSIT on how the chip collaboration fits into Korea's wider sovereign AI funding plans announced the same week.

Written byMara Lindqvist

Mara Lindqvist edits the hardware desk. She covers graphics cards, processors, memory and storage, the foundries and chip designers behind them, and what the numbers on a spec sheet mean for people choosing a PC. Specifications in her stories come from manufacturer spec pages and datasheets.

More from Mara LindqvistAll Hardware coverage