Rebellions has expanded the commercial use of its neural processing units (NPUs) within SK Telecom’s artificial intelligence services. The South Korean company says its chips, including Atom Max, are now processing models used across four areas: voice synthesis for A., AI-based customer service centers, call summarization, and Scam Vanguard.
Rebellions’ NPUs at SK Telecom in 20 seconds
- Rebellions is expanding Atom Max across SK Telecom’s AI services.
- NPUs are used alongside GPUs to run AI models.
- The infrastructure processes 14 million requests a day, roughly 4 billion tokens.
- Commercial deployment began in December 2025 with call summarization.
- Rebellions has built infrastructure with hundreds of NPUs to serve real-time workloads.
The expansion marks a new step in the large-scale commercial deployment of Rebellions’ own AI chips. SK Telecom combines its graphics processing units (GPUs) with Atom Max to run different models depending on each service’s needs.
Rebellions announced on September 21 that its NPUs are now present in services that run SK Telecom’s AI models. The company doesn’t detail the exact split of workloads between GPUs and Atom Max, but it does note that both technologies are part of the infrastructure the carrier uses.
Atom Max moves from testing to commercial services
Atom Max’s first commercial deployment at SK Telecom came in December 2025, initially to generate call summaries. Since then, Rebellions has expanded the number of services that use its accelerators.
These now include A., SK Telecom’s AI platform, with voice-synthesis features. AI-based customer service centers and Scam Vanguard, a service aimed at fighting scams, are also on the list. The company places these cases within a strategy of using NPUs for inference workloads — that is, running already-developed AI models and generating responses to user requests.
Rebellions says Atom Max is designed specifically for data centers and inference workloads. Unlike GPUs, which are used for both training and inference, NPUs are specialized for AI operations. According to the company, that specialization can reduce energy consumption and operating costs for certain workloads.
The company also highlights its chips’ ability to work with low latency in services that need to respond in real time. In SK Telecom’s case, Rebellions says it has tailored its compute resources to the size and performance requirements of the carrier’s different AI models.
Infrastructure that processes 4 billion tokens a day
The scale Rebellions reports is one of the most notable figures in this deployment. The AI infrastructure SK Telecom uses processes 14 million requests every day, equivalent, according to the company, to roughly 4 billion tokens daily.
Rebellions compares that volume to the AI operations of large international companies, though the published information doesn’t provide a consistent metric that would allow the two kinds of infrastructure to be measured directly against each other.
To handle the workloads, the company says it has built compute farms made up of hundreds of NPUs and cards. The setup lets it distribute tasks across numerous accelerators and adapt available resources to each model’s needs.
This point matters because deploying an AI accelerator in production doesn’t depend on chip capacity alone. It also requires infrastructure capable of keeping workloads stable and responding within timeframes compatible with the services customers use.
Rebellions says its NPUs operate within AI development and deployment environments considered standard across the industry. The company also notes it has worked on adapting its resources to SK Telecom’s specific models, rather than using an identical configuration for every application.
SK Telecom plans to keep expanding Atom Max use
The expansion of services falls within a process that started in late 2025. In April 2026, Park Byung-kwan, head of Core Platform at SK Telecom, told government officials the company planned to keep expanding its Atom Max-based commercial services.
Rebellions is now presenting the four use cases as a demonstration of its NPUs moving into production operations. The figure of 14 million daily requests reflects the load the company currently attributes to SK Telecom’s AI infrastructure, though no breakdown by service or accelerator type is provided.
The company also ties the deployment to its goal of expanding chip use in AI services that require continuous availability and low latency. Rebellions CEO Park Sung-hyun has said that running daily-use services on this infrastructure serves as a reference point for continuing to expand its adoption.
The development also points to a broader trend in AI infrastructure: combining different types of accelerators depending on workload. In this case, SK Telecom keeps GPUs alongside Rebellions’ NPUs, while Atom Max is progressively rolled into commercial services that require AI model inference.
The move follows Rebellions’ acquisition of SqueezeBits earlier this year to round out its AI inference stack, and adds to a broader industry conversation about which processor AI workloads actually need — a question CPUs, GPUs, TPUs, and NPUs are all competing to answer.
Frequently asked questions
What is Rebellions’ Atom Max?
Atom Max is a neural processing unit (NPU) designed by Rebellions for data centers, aimed specifically at AI inference workloads.
What does SK Telecom use Rebellions’ NPUs for?
SK Telecom uses Atom Max in services related to voice synthesis for A., AI-based customer service centers, call summarization, and Scam Vanguard.
How many requests does SK Telecom’s AI infrastructure process?
Rebellions says the infrastructure processes 14 million requests a day, equivalent to roughly 4 billion tokens.
When did SK Telecom start using Atom Max?
Atom Max’s first commercial deployment at SK Telecom began in December 2025, initially for call-summarization services.
via: thelec.net
