A reinforcement learning model has cut overcooling in controlled testing without pushing server inlet temperatures beyond safe limits, keeping operators in control of every setpoint while seasonal validation continues before wider automation.
CETA System Co., Limited has confirmed that its reinforcement learning cooling model met an internal efficiency target in controlled testing. The Hong Kong technology company reported the result on 18 March 2025 as evidence that the system can strip out overcooling without pushing server inlet temperatures beyond safe operating conditions. The work validated setpoint optimisation across varying weather and load patterns, addressing data-centre energy efficiency through the systematic removal of waste rather than the compression of the thermal margins that protect critical hardware. Industry-wide, cooling accounts for up to 40% of total energy consumption in operational data-centre facilities, which makes it the largest controllable load available for disciplined efficiency improvement.
The milestone rests on the model’s ability to learn thermal behaviour specific to a single site, where dynamic interactions between equipment, operating practices and environmental conditions create complexity that formula-based engineering frequently fails to capture. Setpoint adjustments responded to shifting IT loads and external weather during the test, with the model building thermal response patterns unique to each facility’s equipment configuration rather than applying a fixed schedule. That distinction matters to Lee Tsz-Hin, Chief Executive Officer of CETA System Co., Limited, who frames the result as proof that the platform “removes cooling that never protected equipment, rather than running anything closer to its limits.”
Most operators maintain overcooling as a precaution against equipment failure, setting conservative temperatures that prioritise uptime and, in doing so, carry an energy cost with no matching gain in thermal safety. The model targets that redundancy directly, adjusting supply air temperature, chilled water setpoints and air-handling-unit static pressure against zone demand and weather forecast, and predicting thermal loads so that cooling is matched to demand before deviations occur rather than corrected after them. Pumps, chillers and fans are managed dynamically as IT loads change, with zone temperature feedback sustaining stability even as total cooling demand falls.
Safety sets a hard boundary that the optimisation cannot cross, and here CETA System treats server inlet temperature as the governing thermal metric, since equipment control systems read it as their reference point. ASHRAE environmental classifications place the recommended envelope for standard data-centre classes at 18°C to 27°C, while the tighter Class H1 band, created for high-density air-cooled systems such as AI accelerators, narrows the recommended range to 18°C to 22°C with an allowable upper limit of 25°C. For facilities running GPU-dense artificial-intelligence workloads, that narrower envelope is a constraint the model must respect without exception, which is why efficiency is pursued only inside approved thermal limits and never by encroaching on the margins that guard critical assets.
Deployment begins in advisory mode, with the platform proposing each setpoint change alongside its operating context while operators approve or reject it and every action is logged for audit. Lee describes the deployment as deliberately cautious, a design in which “the operator approves every setpoint change, and the system steps aside the moment on-site rules are chosen instead.” Control stays with the operator throughout, who can leave AI-assisted operation at any point, with a clean handover to established on-site rules and existing automation heuristics.
Integration follows the same principle of working with installed infrastructure rather than replacing it, connecting to building-management and DCIM systems through standard protocols including BACnet, Modbus, SNMP and OPC-UA, and feeding recommendations into existing controls without displacing established safety systems. Sensor placement and calibration carry real weight, since a baseline error of a single degree can accumulate and distort the thermal model the software depends upon. Machine-learning cooling control is by now an established category, and earlier published deployments elsewhere provide only general backdrop, distinct from the controlled and still-qualitative result that the company reports as its own.
What the controlled result does not yet establish is performance across a full year. CETA System is explicit that seasonal validation and extended load-profile testing must come before automation widens, since heat load tracks both outdoor temperature and operating schedules and year-round data collection remains a prerequisite for any move towards autonomous setpoint control. Lee keeps the claim measured, noting that “a controlled test settles the principle, and a full cycle of seasonal load has to follow before automation is allowed to widen.” The advisory-first model keeps operators in command of setpoint decisions throughout that programme, with every recommendation logged and subject to approval, and it is that measured sequence of validation before any expanded automation scope that the company presents as the discipline behind the work.
About CETA System
CETA System Co., Limited is a Hong Kong-incorporated technology company that builds artificial-intelligence software for data-centre infrastructure. Its platform pairs HVAC and chiller-plant energy optimisation with predictive maintenance for critical assets such as UPS systems, generators and chillers, running vendor-agnostic on a single layer that integrates with existing building-management and DCIM environments and follows an advisory-first deployment model for colocation, enterprise and hyperscale operators across the Asia-Pacific region and beyond.
- Website: https://cetasystem.com
- Registered business: CETA System Co., Limited (Hong Kong BRN 67731517; CRN 2533166)
