Nvidia, Rebellions, SK Telecom target 2028 AI-chip deployment

By Priya Raman

Nvidia, Rebellions and SK Telecom target commercial deployment of an integrated data-center accelerator system in 2028. The three companies signed a memorandum of understanding at AI Day Seoul 2026, held on the 9th at COEX in Seoul’s Gangnam district.

The planned system combines Nvidia graphics processing units (GPUs) with Rebellions neural processing units (NPUs) in one data center. Rebellions, a Korean AI-chip design company, develops the accelerators used for artificial intelligence (AI) inference. The two chip types would divide AI computing workloads.

Nvidia is building the framework. It contributes the rack architecture and chip-interconnect technology. SK Telecom plans to support system validation with its own infrastructure. Rebellions plans to deploy the integrated NPUs in operating data centers.

Rebellions CEO Park Sung-hyun said at the signing ceremony: “By combining Nvidia’s interconnect technology, Rebellions’ chips and SK Telecom’s operating capabilities, we will create a standard for cost-competitive AI infrastructure.” The statement sets a standard as the objective. It provides no cost figure against which that claim can be tested.

What changes inside the data center?

The core plan links multiple units of R200, Rebellions’ next-generation NPU, through NVLink Fusion. Nvidia unveiled NVLink Fusion at Computex in Taiwan in May 2025. The interconnect allows AI accelerators designed by outside companies to join systems built around Nvidia’s NVLink technology.

NVLink itself enables multiple GPUs to exchange data required for AI computation rapidly. Shorter transfer times let many GPUs operate as one system. For the partnership, Rebellions plans to apply NVLink Fusion to its NPUs and integrate them into server racks built to MGX, Nvidia’s server-system design standard.

MGX provides the rack-level integration, while NVLink Fusion provides the chip-level pathway that links Rebellions accelerators to the Nvidia ecosystem for shared AI workloads inside one data center.

The MGX standard is intended to make the installation work with Nvidia’s GPU ecosystem. Rebellions plans to operate the integrated units in actual data centers. SK Telecom plans to support validation using its infrastructure and cooperate toward commercial deployment in 2028.

What could change for memory suppliers?

The framework is expected to open new supply opportunities for Korean memory suppliers Samsung Electronics (005930.KS) and SK hynix (000660.KS), although no R200 supply contract or component volume has been disclosed.

Rebellions equipped its current-generation R100 with HBM3E. That technology is the fifth generation of high-bandwidth memory (HBM). Samsung Electronics supplied the HBM3E used in the chip, according to sources.

• Memory configuration: four HBM3E units.
• Total memory capacity: 144 gigabytes.
• Memory bandwidth: 4.8 terabytes per second.

Detailed specifications for the R200, which will incorporate Nvidia technology, remain undisclosed. Industry expectations point to either HBM3E or HBM4, the next generation of high-bandwidth memory.

Rebellions may also consider SK hynix as an R200 HBM supplier, although discussions remain preliminary. Compared with the R100’s disclosed memory figures, the R200 has no published capacity or bandwidth. The memory opportunity therefore remains tied to an unconfirmed supplier and configuration.

What does the 2028 target establish?

Rebellions was formed through a December 2024 merger with Sapeon Korea, then SK Telecom’s AI-chip affiliate. It also holds a strategic equity relationship with SK Telecom (017670.KS). Those corporate links give the project an identified chip developer, interconnect supplier and infrastructure partner, but the announcement provides no independent performance, cost or deployment results.

Park added: “What we ultimately want is to become an exporter of tokens, the units in which AI generates text, and I see this as a first step toward creating a standard for producing cost-effective tokens.” That goal extends the project beyond data-center operation. It supplies no export schedule or volume.

Commercial deployment remains a goal rather than a completed deployment. The memorandum does not specify launch sites, customer commitments or unit volumes. The announcement leaves three tests: disclosure of R200 specifications, successful validation on SK Telecom infrastructure and execution against the 2028 deployment goal.