- SK Telecom is moving beyond AI models and agents to build its own end-to-end AI stack
- The operator is hoping the custom stack it is building with Arm and Rebellions will help it reap efficiency gains
- SKT pointed to sovereignty as a driving factor in its push
SK Telecom (SKT) has going hard in the AI realm. Between the development of its own large language model A.X and a series of AI agents, its partnership with OpenAI and the launch of its new subsidiary designed to fast track its AI innovation, the operator has had its hands full. Now, it’s tackling an even more ambitious project: building its own AI stack.
The company has teamed up with semiconductor giant Arm and accelerator startup Rebellions to design custom data center hardware geared toward serving inference workloads. The trio will also co-develop the entire software stack to go along with the hardware.
But why build and not just buy?
Lee Jaeshin, head of AI Business Development at SK Telecom, told Fierce the combination of Arm’s AGI CPU and Rebellions’ RebelCard accelerator forms “a heterogeneous architecture that offers superior power efficiency and lower operating costs compared to conventional GPU-based servers.”
Put another way, “this means more inference throughput within the same power envelope — a critical advantage as data centers become increasingly constrained by power capacity, not just compute budget,” he explained.
The trio are planning to do technical trials in SKT’s data centers to validate the stack’s performance and stability. Ultimately, the operator is hoping to run its AI foundation model, A.X K1, on its custom stack as part of a quest to take greater ownership of its AI infrastructure, Lee said.
Sovereign impetus
Beyond its focus on achieving “substantial” efficiency gains by designing its own infrastructure, an SKT spokesperson pointed to rising demand for sovereign AI as a driver for its efforts.
Gartner has forecast that worldwide sovereign cloud infrastructure-as-a-service spending will top $80 billion in 2026 and grow to more than $100 billion in 2027 as regulatory frameworks adapt to AI and geopolitical tensions rise. Sovereign AI is a deeper subset of this trend, encompassing both infrastructure and the models themselves.
The South Korean government last year launched the “Artificial Intelligence Foundation Model” initiative, tapping five teams to build a new sovereign foundational model for the country. SKT was among those chosen to participate and recently one of three companies to advance to Phase 2 of the project following an initial evaluation of progress. SKT’s Haein GPU cluster – one of Korea’s largest – is notably being used as the compute layer for the government’s foundation model project.
SKT has also committed to building 1 GW of AI data center capacity across Korea.
The SKT spokesperson told Fierce the operator believes its work “can serve as a compelling reference case for how telecom operators can find new growth engines through AI.”
“Telcos bring a genuinely differentiated set of assets to sovereign AI. We own and operate the network, the data centers and increasingly the AI models that run on top of them, end-to-end. That integration is difficult to replicate,” the spokesperson concluded. “We believe the business case is real and broad — and our experience with Haein and A.X K1 is starting to demonstrate that.”