Which parts are Korean
Start with what the announcement actually assigns to each side, because the headline word is "sovereign" and the detail points in two directions at once.
Rebellions, a South Korean chip startup, is developing the R200. SK Telecom, a Korean operator, plans to run the systems, operate the data center and supply the models and services on top, including its A.X K2 foundation model. The Ministry of Science and ICT is supporting the collaboration.
That is a real domestic layer: silicon design, operation, models, services, and government support.
Which parts are Nvidia's
Nvidia supplies several other proposed platform elements. NVLink Fusion is Nvidia's interconnect, the link that ties processors together inside a rack. The MGX architecture is Nvidia's foundation for shared rack-design elements, including rack footprint, power delivery and cooling. And under the inference split the partners plan to develop, Nvidia GPUs would run the prefill stage of generating a response, which processes the prompt, while the R200 takes the decode stage, which generates the response token by token.
So the Korean chip does not replace the Nvidia system. It joins it, on Nvidia's connector, in systems built on Nvidia's rack architecture, performing one of the two stages. NVLink Fusion exists for exactly this purpose: it extends Nvidia's platform to specialized processors built by other chipmakers, including the proposed R200 systems.
The date and the missing numbers
The release has an availability section, and it is one line long: R200 systems are expected to be ready for deployment in 2028. The chip itself is still in development. Its description, optimized for low-latency and energy-efficient inference on frontier models including multitrillion-parameter mixture-of-experts models, is a design target stated by the companies, not a result from a running system. The efficiency case for the prefill-and-decode split is likewise presented as an intention, to improve throughput and power efficiency, with no figure attached to either.
No financial terms are disclosed. The release discloses no order commitment, investment amount, unit volume or pricing. BigGo Finance, covering the signing at Nvidia AI Day in Seoul, characterizes the arrangement as a memorandum of understanding rather than a binding supply contract. The joint release uses the words "collaboration" and "agreement", and does not name a contract value either.
How to read it
None of this makes the deal empty. A domestic inference chip designed into rack-scale systems at a national operator, with government support, is a concrete starting point. It does mean the test is specific. The useful evidence will be deployment status in 2028, production workloads running on R200 silicon, and measured cost and energy consumption against a clearly defined alternative.
ALFW covered a second Korean Nvidia announcement from the same Seoul event today, [Kakao Mobility's Level 4 driving agreement](https://alfw.ai/article/kakao-mobility-nvidia-mou-level-4-no-timeline/). In both announcements covered here, Korean companies plan to combine their own technology with Nvidia platform components.
Sources
- [1] NVIDIA, Rebellions and SK Telecom, joint announcement via GlobeNewswire, 9 November 2026 — read in full by the desk (TradingView mirror of the release).Read source
- [2] Aju Press, "Rebellions to pair Korean AI chip with Nvidia GPUs", 9 November 2026.Read source
- [3] BigGo Finance, "Nvidia Taps Rebellions, SK Telecom to Build Korean Sovereign AI Racks", 9 November 2026 — source for the memorandum-of-understanding characterization.Read source