Nvidia News: Korean AI Chip Deal Builds Sovereign AI
Nvidia news from Seoul on Monday showed the chipmaker opening its platform to a partner it once would have seen as a rival. Nvidia, SK Telecom and the South Korean artificial intelligence chip startup Rebellions agreed to combine their strengths to establish a heterogeneous computing system, a deal that will link Korean-designed AI processors to Nvidia’s infrastructure and anchor a sovereign AI push backed by the South Korean government. The agreement, signed on the margins of AI Day Seoul, is a sign of how the AI chip business is changing: the company that dominates the market is now inviting others to build on its platform, and the countries that buy its chips are insisting on a role in making them.
Under the trilateral memorandum of understanding, the companies will cooperate in the establishment of a heterogeneous AI computing system connecting Nvidia’s graphic processing units with neural processing units designed by Rebellions, which will be used in the building of SK Telecom AI data centers. The tie-up will be based on Nvidia’s NVLink Fusion program, which allows hyperscalers and custom chip designers to integrate their own AI processors directly into Nvidia’s rack-scale infrastructure. The MOU was signed on the margins of AI Day Seoul, a tech event hosted by Nvidia that kicked off its two-day run in southern Seoul, and the deal was announced by the companies on Monday, November 9.
Rebellions plans to adopt NVLink Fusion to connect its next-generation R200 neural processing units to Nvidia’s open infrastructure platform for rack-scale inference. The company will harness Nvidia’s MGX rack architecture for rack-scale system development, and SK Telecom plans to deploy these systems in its AI data centers, drawing on its AI data center operations expertise, proprietary models and services. The collaboration is supported by Korea’s Ministry of Science and ICT, and it aims to support AI services tailored to Korea’s language, industries and priorities, connecting Korean semiconductor development with the systems and operational capabilities needed to deliver AI at scale.
What NVLink Fusion Means
The technical core of the deal is NVLink Fusion, a program that extends Nvidia’s vertically integrated, horizontally open AI platform to custom processors. The program allows hyperscalers and custom chip designers to integrate their own AI processors, known as XPUs, directly into Nvidia’s rack-scale infrastructure, rather than building separate systems or competing head-on with Nvidia’s own chips. For Rebellions, the deal means that its R200 neural processing units can work alongside Nvidia’s GPUs in the same rack, combining the strengths of both: Nvidia’s general-purpose computing power and Rebellions’ specialized inference acceleration.
Rebellions is developing its next-generation R200 NPU, optimized for low-latency and energy-efficient inference using frontier models, including multitrillion-parameter mixture-of-experts models. The R200 is designed for inference, the stage of AI computing where trained models generate answers, rather than training, where models are built. Inference is the fastest-growing part of the AI computing market, because every chatbot answer, every image generation and every voice assistant response runs on inference chips, and the demand for efficient inference hardware is rising as AI services scale. Rebellions’ R200 systems are expected to be ready for deployment in 2028, a timeline that gives the company two years to move from agreement to production.
The heterogeneous approach, combining GPUs and NPUs in a single system, reflects a broader shift in how AI data centers are being built. The early AI boom was dominated by Nvidia’s GPUs, which could handle both training and inference. As the market matures, specialized chips designed for specific tasks are becoming more important, and data center operators are looking for systems that combine different kinds of processors for different jobs. The deal in Seoul is a bet that the future of AI computing is heterogeneous, with Nvidia’s platform as the connective tissue and specialized chips from partners like Rebellions handling the workloads they do best.
Sovereign AI and Korea’s Ambitions
The deal is also a sovereign AI play, and the language the companies used makes that clear. SK Telecom plans to deploy these systems to support its vision for full-stack sovereign AI spanning infrastructure, proprietary models and public-facing services, including those under Korea’s AI for All project. Sovereign AI, the idea that nations should build their own AI infrastructure, models and services rather than relying entirely on American or Chinese technology, has become a priority for governments around the world, and South Korea is one of the most active.
The collaboration is supported by Korea’s Ministry of Science and ICT, and the ministry’s involvement signals that the government sees the deal as part of a national strategy. Korea is one of the world’s leading semiconductor nations, home to Samsung and SK Hynix, the two companies that dominate the memory chip market, and the government has been pushing to extend that strength into AI chips. Rebellions, a startup founded to build inference accelerators for enterprises and nations, is the kind of company the government wants to champion: a domestic designer with the technical capability to compete globally, and a partner that can bring Korean silicon into the AI infrastructure that the country is building.
For SK Telecom, the deal fits a broader transformation. The carrier has been investing heavily in AI data centers, models and services, positioning itself as an AI company rather than just a telecom operator. The deployment of the Nvidia-Rebellions systems in its data centers will give SK Telecom the infrastructure to run its own models and services, and the full-stack approach, spanning infrastructure, proprietary models and public-facing services, is designed to make the company a complete AI provider rather than a customer of someone else’s platform.
The Context: Nvidia’s Week
The deal comes at a moment when Nvidia is facing questions at home even as it expands abroad. The chipmaker’s board of directors has opened an investigation into its chief executive, Jensen Huang, and the news weighed on the market on Monday, with Nvidia’s Frankfurt-traded shares more than 6 percent lower in early European trading. The investigation is a governance matter, not a business one, and there is no evidence of operational disruption, but it adds a new kind of uncertainty to a company that has been the bellwether of the AI trade for two years.
The Seoul deal is a reminder of why Nvidia’s position remains strong despite the turbulence. The company’s platform, from its GPUs to its NVLink interconnect to its MGX rack architecture, is the foundation on which the AI industry is built, and the decision to open that platform to partners like Rebellions strengthens it rather than weakening it. By inviting custom chip designers to build on NVLink Fusion, Nvidia is making its infrastructure the standard for AI computing, and the more companies that build on it, the harder it becomes to displace.
The event itself underlined Nvidia’s reach in Korea. AI Day Seoul, hosted by Nvidia at COEX in southern Seoul, drew participants from across the Korean tech industry, and Deepu Talla, vice president of robotics and edge AI at Nvidia, spoke at the event. Rebellions’ ATOM-Max chip was on display at the Physical AI Tech Week held by the Ministry of Science and ICT at COEX on October 13, a sign of how closely the startup is working with the government and the industry. The MOU was the culmination of that engagement, and the two-day event provided the stage for the announcement.
What Happens Next
The timeline for the deal is measured in years, not months. Rebellions’ R200 systems are expected to be ready for deployment in 2028, which means the work of integrating the chips, building the systems and deploying them in SK Telecom’s data centers will take place over the next two years. The MOU is a memorandum of understanding, not a binding contract, and the details of the commercial arrangement, the volumes involved and the specific data centers that will host the systems have not been disclosed. But the direction is clear: Korea intends to build sovereign AI infrastructure on a combination of Nvidia’s platform and domestic silicon, and the three companies have agreed to build it together.
The deal also has implications beyond Korea. Nvidia’s NVLink Fusion program is open to hyperscalers and custom chip designers around the world, and the Seoul agreement is a template for how the program can work in practice. If the model succeeds in Korea, other countries pursuing sovereign AI may follow the same path, combining Nvidia’s infrastructure with their own chip designs. For Rebellions, the deal is a validation of its technology and a path to scale: a startup’s chip, deployed in a major carrier’s data centers, running on the industry’s dominant platform.
For the AI chip market, the agreement is a sign of a maturing industry. The early years of the AI boom were defined by a single company’s dominance. The next phase is likely to be defined by partnerships, heterogeneous systems and national strategies, and the deal in Seoul brings all three together. The chips will not be ready until 2028, but the direction of travel is already clear, and the companies that signed the MOU on Monday intend to be there when it arrives.
Rebellions itself is a product of Korea’s push into AI hardware. The company develops inference accelerators for enterprises and nations, bringing together computing performance, energy efficiency and supply chain resilience, and its focus on inference rather than training sets it apart from the chipmakers chasing the training market. Inference is the workhorse of deployed AI: every time a user asks a chatbot a question, generates an image or dictates a message, an inference chip does the work, and the efficiency of that chip determines how much the service costs to run. As AI services scale to millions of users, the demand for chips that can deliver answers quickly and cheaply is growing faster than the demand for training hardware, and Rebellions has built its business around that shift. The R200, optimized for low-latency and energy-efficient inference on frontier models, is the company’s bet that specialized inference silicon will be as important to the next phase of AI as general-purpose GPUs were to the last one.
*Source: Yonhap*