Nvidia, Rebellions and SK Telecom partner for Korean sovereign AI

Market context and competition

Rebellions is not the only domestic player attempting to enter this space. Another Korean startup, FuriosaAI, is taking a different technical route by using Ethernet networking with Broadcom. This shows that Korea's domestic chip industry is split on which interconnect standards to adopt for sovereign AI.

Rebellions has already demonstrated some production experience with its ATOM-Max NPUs, which power several SK Telecom services. These services include call summarization and fraud detection, handling billions of tokens daily. However, these results come from internal tests rather than independent third-party benchmarking.

The Rebel R200 will face competition from other decode-specialized chips already integrated with Nvidia. This includes Groq's 3 LPX and d-Matrix's Raptor, which are slated for 2027. Rebellions' 2028 target means it will be entering a crowded market where established alternatives may already have significant installed bases.

}

Technical hurdles and model scale

The project aims to support SK Telecom’s A.X K2 foundation model, which is a massive architecture featuring 688 billion total parameters. Because only 33 billion parameters are active for each token, the model requires significant memory to store its weights. These weights use an 8-bit numerical format called FP8 to reduce the overall memory footprint, yet they still occupy roughly 695GB of space.

To handle this volume, the infrastructure would need to distribute the model across multiple hardware packages. For example, it would require at least five Rebel100-class packages to provide 720GB of combined memory. This calculation does not even account for the KV cache, which is the working memory of the prompt, or other necessary data required for active inference.

Financials and corporate milestones

While the trilateral agreement provides a clear roadmap, the specific financial details of the partnership remain private. The companies have not disclosed any investment commitments, specific order volumes, or pricing for the hardware. This means the current agreement serves as a formal memorandum of understanding rather than a finalized purchase order for a set amount of equipment.

Rebellions is currently preparing for a significant transition in its corporate status. The company is targeting a listing on the Korean stock market during the first half of 2027. This follows a pre-IPO funding round in March 2026 that valued the firm at approximately $2.34 billion, signaling its growth as a key player in the domestic chip landscape.

Missing specifications and unknowns

Despite the high-profile nature of the collaboration, the technical details of the Rebel R200 NPU remain largely under wraps. The companies have published no official specifications regarding its power consumption, memory bandwidth, or process node. This lack of data makes it difficult to independently verify how the new chip will perform compared to existing solutions in the market.

The success of the 2028 deployment depends on how well the different processor architectures can communicate. Specifically, the handoff of data from Nvidia GPUs to Rebellions NPUs must be seamless to maintain efficiency. Because the companies have not explained the specific transfer mechanism or latency for this handoff, the true performance of the system remains a theoretical goal for now.

Corporate distractions and logistics

The signing of the memorandum of understanding occurred during Nvidia AI Day Seoul 2026, but the event faced significant media interference. Coverage of the partnership was largely overshadowed by a personal scandal involving Nvidia CEO Jensen Huang. This situation even impacted the logistics of the meeting, as Madison Huang, a senior director at Nvidia, had to cancel her planned visit to the region. Her trip included a scheduled visit to the Rebellions headquarters and a dinner with their executive team.

The corporate atmosphere was further affected by market reactions to the news surrounding the CEO. Nvidia shares experienced a decline on Monday and continued to fall on Tuesday as the company approached its fiscal third-quarter results. While the board maintained confidence in the leadership, the surrounding controversy served as an unwelcome distraction during a pivotal moment for the Korean sovereign AI initiative.

Pragmatic localization and expansion

Rebellions' leadership has expressed a pragmatic view on the limitations of total technological independence. The company's CEO, Sunghyun Park, noted that attempting to localize every single component, including basic hardware like bolts and screws or complex software like operating systems, is not realistic. He argued that such an approach would hinder the ability to scale quickly or reach the market in a timely manner, justifying the use of Nvidia's established infrastructure.

Beyond the Korean market, Rebellions is seeking to establish a broader international presence for its hardware. The company has announced plans for additional rack projects in Japan and within a research cluster supported by the UK government. While these projects do not yet represent sustained commercial volumes, they indicate an effort to validate the technology in multiple jurisdictions outside of the domestic Korean ecosystem.