News · 2026-11-09

NVIDIA, SK Telecom and Rebellions plan a hybrid AI inference rack

NVIDIA, SK Telecom and Rebellions have agreed to develop and validate a hybrid rack system that divides AI inference between NVIDIA GPUs and a Korean-designed R200 processor. NVIDIA expects R200 systems to be ready for deployment in 2028; the announcement is a development commitment, not evidence that the combined system is already operating.

Key facts

• SK Telecom dates the signing of the memorandum to 9 November 2026.
• NVIDIA gives 2028 as the expected readiness date for R200 systems.
• The proposed architecture assigns prompt prefill to GPUs and token decoding to Rebellions NPUs.
• SK Telecom’s announcement describes development and validation, not a completed customer rollout.

Generating an AI response has at least two different phases. First the system processes the input prompt; then it generates output tokens one by one. The proposed split gives the two jobs to different hardware, connected using NVIDIA’s NVLink Fusion and designed around its MGX rack architecture. The partners’ intended payoff is throughput and power efficiency, but they have not published a result proving that this particular design delivers either benefit.

A useful analogy is a kitchen where one station prepares ingredients in bulk and another plates each dish as orders arrive. Different equipment can suit the two stages, but the kitchen must move work between stations efficiently. A promising division of labor does not by itself prove the restaurant serves more customers per hour.

The announcement sits within a wider Korean infrastructure push. SK Telecom previously reported that its Haein service launched with more than 1,000 NVIDIA Blackwell GPUs in one cluster. In a separate AI Cloud announcement, SK Telecom and NVIDIA described a gigawatt-scale plan with a first AI factory targeted for 2027. Neither fact establishes that the planned R200 racks are running or that they will be part of that specific factory.

NVIDIA’s broader sovereign AI account puts domestic models and infrastructure in a national context. The phrase “sovereign AI” describes the participants’ framing, not complete technological independence: the hybrid design explicitly combines Korean processors with NVIDIA technology. The practical question is whether a local chip designer can integrate into a widely used system stack and offer useful capacity on real workloads.

The strongest caution is that an agreement is a starting line. The sources do not specify rack count, contract value, tested speedups or guaranteed customer adoption. A 2028 expectation can be checked later against demonstrated systems and independent measurements. For the underlying distinction between the two inference phases, see prefill and decode.

The technical split also creates coordination costs. Prompt processing and generation do not happen in isolation: the result of the first phase must be made available to the hardware handling the next one. A system’s outcome depends on transfer overhead, workload mix, scheduling and the actual software stack. An architecture can be plausible on paper and still fail to improve the metric that a customer cares about, such as cost per useful response or latency at a given level of demand. The partner announcements do not provide enough measured detail to make that comparison.

The dates are also different types of commitments. A memorandum of understanding sets out cooperation, while a deployment-readiness expectation is a forward-looking target. Neither is equivalent to a completed procurement contract. The 2027 target for the separately announced first AI Cloud factory refers to a different plan; combining the two dates into a single deployment timeline would add a connection the sources do not establish.

For Korea, the broader attraction is control over more of the AI value chain: domestic model development, local services and a role for local chip design. But the reliance on NVIDIA interconnect and rack systems makes the model one of partnership and integration. That can still be strategically useful, while leaving important dependencies in place. Observers should judge it by whether the resulting systems are buildable, maintainable and competitive for the workloads customers actually run.

Future disclosure of an implemented rack, the specific measurement conditions, and independent or customer validation would make the performance claim much more meaningful. Until then, the 2028 date can be tracked as a readiness expectation and the announced technical split as a design objective.

The partners’ next engineering milestones will matter more than the announcement’s ambition.

Primary source, verified: read the paper →

Ground Truth. (2026, November 9). NVIDIA, SK Telecom and Rebellions plan a hybrid AI inference rack. Ground Truth. https://groundtruth.day/news/nvidia-sk-telecom-rebellions-hybrid-inference.html

@misc{groundtruth:nvidia-sk-telecom-rebellions-hybrid-inference,
  title  = {NVIDIA, SK Telecom and Rebellions plan a hybrid AI inference rack},
  author = {{Ground Truth}},
  year   = {2026},
  month  = {nov},
  url    = {https://groundtruth.day/news/nvidia-sk-telecom-rebellions-hybrid-inference.html}
}

Topics: infrastructure · chips · south-korea · inference