MediaTek announced the opening of a new research and development data center at the Tongluo Science Park in Miaoli, Taiwan, aimed at supporting growing demand for edge AI and cloud AI development.
Taiwan's first DGX SuperPOD cluster
The company said the facility houses Taiwan's first Nvidia DGX SuperPOD cluster powered by the Nvidia DGX B200 platform, creating a large-scale AI computing environment for model training and inference workloads. The site also marks Taiwan's first large-scale R&D data center to adopt immersion cooling technology as part of its energy efficiency strategy.
Built to the highest green standards
According to MediaTek, construction of the Tongluo R&D data center began in 2023 under Taiwan's highest "diamond-grade" green building standards. The project follows a modular three-phase development plan, with the first phase completed and operational in 2026. Future expansion will proceed based on business growth and actual AI computing requirements.
The first phase currently operates at approximately 15 MW of computing-related power consumption. Each future phase is expected to carry a similar scale, bringing the total projected capacity to around 45 MW once fully completed.
Capex undisclosed as AI demand evolves
MediaTek did not disclose the total capital expenditure required for all three phases, stating that the pace of future investment would depend on evolving AI technology and internal R&D demand. The company noted that rapid changes in AI computing architectures make it difficult to provide long-term deployment timelines or spending forecasts.
Rising AI computing demand may continue to tighten the supply of high-performance computing infrastructure globally. Based on the estimated three-year construction timeline for the first phase, analysts suggested subsequent phases could also require multiple years to complete, depending on equipment availability and demand growth.
138 billion tokens a month — and counting
MediaTek said the AI training and inference platform can process up to 138 billion tokens per month and support more than 24,000 model training iterations. The company added that software optimization tools — including Nvidia NIM inference microservices and TensorRT-LLM frameworks — can improve inference speed by 40% and increase token throughput by 60%.
Immersion cooling sets a new efficiency benchmark
The facility combines traditional air cooling for Nvidia DGX B200 systems with single-phase immersion cooling technology for CPU server racks. In immersion cooling systems, servers are submerged in non-conductive dielectric fluid designed to transfer heat efficiently. MediaTek said the design improves cooling efficiency by 2.6 times and achieves a power usage effectiveness (PUE) rating of 1.1.
The company said immersion cooling also reduces exposure to airborne dust while eliminating fan noise and vibration, thereby improving system stability.
Fab-grade power infrastructure for uninterrupted uptime
To support continuous operation, the Tongluo data center was designed with semiconductor fab-level power infrastructure standards. MediaTek said the site includes a 2N dual-redundancy power architecture and full-load backup generators, allowing the facility to maintain operations during regional power disruptions.
Article translated by Jingyue Hsiao and edited by Jerry Chen
