As the enterprise rush towards artificial intelligence deployment continues, the staggering cost of compute hardware and energy remains a formidable barrier. On 2 October 2026, South Korean telecommunications giant LG Uplus and AI optimisation specialist OptAI announced a significant expansion of their strategic partnership, unveiling token optimisation techniques designed to drastically reduce the resource burden of running advanced AI models.

The collaboration, which now extends from edge devices to server-side GPU environments, focuses on streamlining the fundamental computational processes required for AI response generation. According to LG Uplus, the joint research has yielded a staggering improvement in server environments, enabling a single GPU to process up to four times as many tokens as standard configurations. This breakthrough effectively quadruples throughput without necessitating any additional hardware investment.

These server-side gains build upon the duo's recent success in the on-device AI space. By applying sophisticated lightweighting techniques to LG's EXAONE-based small language model (sLM), the companies have developed a highly efficient model capable of running directly on a smartphone's neural processing unit (NPU). This optimised architecture, which is currently being integrated into LG's ixi-O AI calling application, reportedly slashes power consumption by 78 per cent and shrinks the overall model footprint by 82 per cent compared to traditional CPU-based execution—crucially, without compromising on performance.

To further drive the conversation around operational sustainability, LG Uplus and OptAI co-hosted the "Efficient AI Tech Seminar" on 1 October. The gathering brought together key players from across the Asia-Pacific technology ecosystem, including AI semiconductor developer FuriosaAI, machine learning platform VESSL AI, the Korea Electronics Technology Institute (KETI), and researchers from Hanyang University. Discussions focused on the entire lifecycle of model optimisation, exploring practical strategies for transitioning highly efficient AI architectures from the laboratory into live, large-scale commercial services.

For readers and businesses in Aotearoa, these developments represent a highly practical opportunity. As local enterprises look to integrate generative capabilities into their daily operations, a heavy reliance on expensive, power-hungry offshore cloud infrastructure often limits scalability and raises persistent data sovereignty concerns. The ability to deploy highly efficient, lightweight models on local edge devices, or to run them on optimised domestic servers, provides a clear blueprint for building sovereign AI solutions. By adopting similar token optimisation strategies, New Zealand developers can reduce operational latency, lower their cloud computing expenditure, and actively minimise the environmental footprint of their growing digital infrastructure.

Sources


AI disclosure: Automated systems assisted with research and drafting. Gemini independently peer-reviewed this article before publication. Sources are linked for verification.