SK Telecom now processes 4 billion tokens per day on neural processing units supplied by Rebellions, the Seoul-based chip designer, according to a report by Chosunbiz. The figure marks one of the most concrete production-scale deployments of a Korean-designed NPU inside a domestic telecom operator's AI infrastructure.

The 4 billion-token daily throughput covers the AI services SK Telecom operates for its subscriber base. A token is the basic unit of text that language models consume and generate; 4 billion tokens per day translates into sustained, high-volume inference work across customer-facing applications rather than batch research runs.

The deployment carries weight for both companies. For Rebellions, a startup competing against established accelerator vendors, a live carrier-scale workload demonstrates that its silicon can hold up under continuous, real-time demand. Inference at telecom scale requires the hardware to serve many concurrent users with low latency, around the clock, without the bursty tolerance that offline processing allows.

For SK Telecom, the arrangement supports its strategy of building AI capability on infrastructure it can specify and control. Running inference on domestically designed NPUs gives the operator an alternative supply line for accelerator hardware, a category that has been constrained by limited availability and long lead times across the industry.

The two companies have positioned the partnership as a test of whether Korean-designed chips can carry national AI infrastructure ambitions. SK Telecom has invested in AI as a core business line, and the operator has repeatedly framed its infrastructure choices as strategic rather than purely transactional. A production metric like daily token volume offers a harder measure than benchmark demonstrations or pilot programs.

Rebellions designs NPUs specifically for inference workloads rather than general-purpose computing. That focus matters for operators: inference, not training, dominates the operating cost of deployed AI services. Every query a subscriber sends to an AI service triggers compute that the operator pays for in power and silicon utilization. Hardware tuned for that workload can change the economics of large-scale AI service delivery.

The reported 4 billion tokens per day also serves as a baseline the partnership can be measured against. If SK Telecom expands its AI subscriber services, throughput requirements will grow accordingly, and the NPU fleet will need to scale with them. Operators watching the deployment will track whether Rebellions hardware maintains performance and cost efficiency as volume rises.

The news lands amid intensified competition in the AI accelerator market, where buyers weigh performance per watt, software maturity, and supply reliability against incumbent platforms. A carrier endorsement with production numbers gives Rebellions reference-case material few startups in the segment can claim.

Chosunbiz reported the deployment figures. Neither company has publicly broken down the figures by individual service, and SK Telecom has not disclosed how the token volume splits across its AI product lines.

For Korea's semiconductor sector, the milestone signals that domestic NPU design has moved past the demonstration phase into revenue-bearing, carrier-grade operation. The next indicators to watch are capacity expansion at SK Telecom, further operator customers for Rebellions, and any published efficiency data from the deployment.