Nvidia Puts Korean Silicon Inside Its Own Racks: Rebellions R200 on NVLink Fusion by 2028

Nvidia used its AI Day event in Seoul on Monday to do something it has avoided in most markets: put another company's silicon inside its own rack. Under a collaboration with Rebellions and SK Telecom, backed by Korea's Ministry of Science and ICT, the Korean chip designer will connect its next-generation R200 neural processing units to Nvidia's NVLink Fusion programme and build rack-scale inference systems on Nvidia's MGX architecture.

The split is explicit. Nvidia GPUs would handle prefill, the pass that reads a prompt and builds its internal state; Rebellions NPUs would handle decode, the token-by-token generation that dominates cost at scale. SK Telecom plans to deploy the systems in its AI data centres and tie them to A.X K2, its in-house foundation model, and to the government's AI for All programme. Nvidia says the arrangement targets better throughput and power efficiency. Rebellions expects R200 systems to be ready for deployment in 2028; the chip itself is still in development, and no prices, volumes or contract values were disclosed.

The concrete releases were softer. Nvidia published Dynamo-optimised deployment recipes for four Korean open models -- A.X K2 from SK Telecom, K-EXAONE 2.0 from LG AI Research, Solar Open 2 from Upstage and Motif-3 from Motif Technologies -- tuned for its B200 GPUs and claimed to deliver up to 2x performance out of the box. Those figures are the company's own and have not been independently benchmarked. It also released Nemotron-Personas-Korea, seven million fully synthetic Korean personas grounded in official national statistics, for builders training and testing models on Korean data.

The factory claims were the most measurable. Samsung is expanding its work with Nvidia on accelerated computing and AI physics across materials research, process development and device engineering; in one materials case Nvidia says a GPU-accelerated workflow ran roughly 1,600 times faster than the CPU version, cutting a simulation from about 18 hours to 41 seconds, and the company's sMOL platform uses its cuEST quantum chemistry library for speedups of up to 5x. Kakao Mobility plans a Level 4 autonomous driving system on Nvidia's Hyperion platform and DRIVE AGX Thor, validated through the Halos safety stack and starting with Kia purpose-built vehicles. LG Electronics demonstrated Isaac GR00T-driven manipulation on its CLOiD robot in a real manufacturing line, and Hyundai Motor Group said it is exploring Nvidia's NemoClaw blueprints and OpenShell runtime.

Security got a mention too. AIM Intelligence, a Korean AI safety company, is integrating its Starfort agent-control platform with OpenShell to combine runtime access control with detection of malicious instructions and data exfiltration, with a first proof of concept under discussion in the public energy sector.

Read the fine print and the event is less a product launch than a framework agreement. The R200 racks arrive in 2028, Hyundai is only exploring, and several items re-present October announcements such as Switchyard and OpenShell in a Korean setting. Memoranda are not orders, and the ministry's role is to convene rather than to buy.

That is still the point. Korea has its own memory makers, its own foundation models and its own factories, which makes it the cleanest test of Nvidia's sovereign AI pitch: let national champions plug into Nvidia's rack-scale platform, and the country's AI ends up routed through Nvidia hardware even where the silicon is local. For Rebellions the trade is access to a supply chain and a customer it could not reach alone. For Seoul the bet is that compute sovereignty is worth more than silicon purity. The test is whether the 2028 hardware lands, and whether recipes shipping today are enough to keep Korean developers on the stack until it does.