Rising GPU and CPU thermal loads are forcing hyperscale data-centre operators to reconsider cooling architecture, with CETA System Co., Limited positioning direct-to-chip liquid cooling as an infrastructure response to density, energy and compliance pressures rather than an optional technical upgrade.
The shift is being driven by processor heat loads that have moved beyond conventional air-cooling design assumptions. Current GPU thermal loads exceed 1.5 kilowatts, while CPUs consistently operate above 300 to 500 watts, pushing total thermal design power beyond 700W. Rack densities that historically sat within 5 to 10 kW now routinely reach 30 to 100 kW in GPU-dense environments, with high-performance computing racks surpassing 100 kW under operating conditions. At 200+ kW density levels, air transport capacity becomes insufficient, exposing operators to stranded power capacity, thermal inefficiency and higher outage risk.
“Liquid cooling is becoming a planning question for capacity, resilience and cost control, not simply a mechanical-engineering preference,” says Lee Tsz-Hin, Chief Executive Officer of CETA System Co., Limited. “When thermal density outruns air movement, operators need a cooling model that protects both compute performance and facility economics.”
Forecasts in the sector place global data-centre capacity near 165 GW by 2030, with artificial intelligence workloads driving a 16% compound annual growth rate over the forecast period. Traditional power distribution and cooling systems designed for 10 to 40 kW rack densities cannot accommodate 200+ kW AI-optimised configurations entering production deployment. Waste heat volumes also weaken the effectiveness of containment strategies and optimised airflow designs, increasing the importance of cooling approaches that remove heat at the source.
Direct-to-chip systems use cold plates mounted on processors, creating a thermal pathway between silicon and circulating coolant loops connected to facility heat-rejection systems. Coolant options include deionised water, water-glycol mixtures and dielectric fluids in closed-loop systems integrated within server chassis. Single-phase deployments maintain chip-to-coolant temperature differentials of 17 to 20°C under 500W processor loads, compared with air-cooled systems exceeding 60°C under equivalent conditions.
For operators, the commercial case is closely tied to energy and utilisation. Cooling is the largest energy consumer after IT equipment, accounting for up to 40% of total facility energy consumption. As density rises, cooling systems run at higher capacity utilisation and take a larger share of facility power. Without efficiency gains, thermal-management energy costs can approach parity with server power consumption.
Capital costs remain material. A 1MW direct-to-chip cooling installation requires approximately $174,200, compared with $80,400 for an air-based system. Annual power costs, however, decline to $1,165,800 for liquid cooling compared with $1.7 million for equivalent air-cooling capacity. Five-year total cost of ownership calculations indicate $6 million for direct-to-chip cooling versus $8.7 million for air-cooling equivalents, representing a 31.5% reduction with a 2.2-month payback period.
CETA System also identifies reliability and performance as central to the investment case. Temperature-controlled environments show sustained AI training performance improvements of 12% through reduced thermal throttling and more consistent model development. Component reliability data shows heat-related failure reductions of 50%, server reliability improvements of 23% and mean time between failures extending by 20% under controlled thermal conditions.
The transition also brings integration and operating challenges. Direct-to-chip cooling links IT systems more tightly with facilities infrastructure, making coolant flow rates, fluid quality and Coolant Distribution Unit performance critical to rack-level thermal stability. Coordination failures, including unannounced coolant-flow changes or rack additions without facilities notification, can create thermal overload conditions and unplanned downtime. Inadequate filtration can introduce contaminants that obstruct cold plate channels and degrade heat transfer.
The operating model therefore requires stronger collaboration between facilities teams, IT operations and mechanical and electrical specialists. Teams need competence in coolant chemistry, pressure management and manifold-system design, while practical training programmes using hands-on workshops and system simulations can support safer commissioning and maintenance.
Direct-to-chip cooling also differs from immersion cooling in its operational trade-offs. Direct-to-chip designs preserve conventional server architecture while targeting CPUs and GPUs, although memory, power delivery, storage and networking components still depend on airflow. Immersion cooling absorbs heat across complete servers submerged in dielectric fluid, but requires specialised enclosures, structural reinforcement and more complex maintenance procedures. For many existing facilities, direct-to-chip offers stronger compatibility with established infrastructure and simpler hardware servicing.
Regulatory and reporting obligations add another layer to the business case. Under the EU Energy Efficiency Directive, data-centre operators with total rated power of at least 500 kW must submit annual energy-performance data, including energy consumption, Power Usage Effectiveness, temperature set points, waste-heat utilisation, water usage and renewable-energy deployment. Facilities above 1 MW must implement waste-heat recovery for heating applications or alternative energy-recovery systems unless technically or economically unfeasible.
AI-driven optimisation is becoming part of the operational discussion, but deployment remains advisory-first and operator-supervised. Operators continue to have reservations about autonomous AI control, particularly where reinforcement-learning systems make non-intuitive adjustments to optimise long-term performance. Contemporary frameworks incorporate guard-mode functionality that engages additional cooling capacity when environmental parameters exceed defined tolerances, supporting fail-safe behaviour while keeping human oversight central.
“Trust is as important as optimisation,” Lee says. “AI can support better coolant-temperature monitoring and energy-consumption decisions, but critical facilities still need operator oversight, clear constraints and predictable fallback behaviour.”
The broader adoption of direct-to-chip liquid cooling will depend on disciplined procurement, vendor-agnostic integration and practical commissioning. Open API frameworks supporting bidirectional connectivity with DCIM systems, building-management platforms and third-party monitoring tools are essential to avoiding proprietary constraints. Flow, temperature and pressure mapping during commissioning also allows operators to calibrate thermal models using real-time telemetry.
As hyperscale infrastructure moves towards denser AI workloads, the cooling decision is becoming a capital-efficiency decision as much as a technical one. CETA System views successful migration as dependent on integration discipline, operator competency and advisory-first AI thermal intelligence that supports higher-density computing without weakening resilience, compliance or cost control.
About CETA System
CETA System Co., Limited is a Hong Kong-incorporated technology company founded in 2017, delivering artificial-intelligence solutions for data-centre infrastructure. Its vendor-agnostic platform combines HVAC and chiller-plant energy optimisation with predictive maintenance for critical assets including UPS systems, generators and chillers, integrating with existing building-management and DCIM environments through an advisory-first deployment model for colocation, enterprise and hyperscale operators across Asia-Pacific and beyond.
• Website: https://cetasystem.com
• Registered business: CETA System Co., Limited (Hong Kong BRN 67731517; CRN 2533166)