Two large infrastructure agreements and rising rack power densities are pushing liquid cooling, energy spend and uptime risk to the centre of AI data-centre planning, as operators face expanding reporting and audit duties in Europe and Hong Kong.
CETA System Co., Limited is highlighting two CoreWeave infrastructure agreements concluded within 48 hours as a sign that thermal stability is fast becoming a binding constraint for AI data centres. An expanded $20.04 billion agreement with Meta Platforms extends AI cloud capacity through the end of 2032, while a separate multi-year deployment partnership with Anthropic is set to support Claude AI models.
For operators, these deals translate into practical questions about how quickly facilities can scale without breaching cooling limits or eroding resilience margins. “Multi-year capacity commitments sharpen the focus on whether a site can hold stable temperatures as rack densities move beyond air cooling,” Lee Tsz-Hin (李子軒), Chief Executive Officer of CETA System Co., Limited, observed, tying thermal headroom to outage exposure and operating cost.
GPU-intensive workloads are pushing heat loads beyond conventional cooling thresholds. NVIDIA’s GB200 NVL72 rack consumes about 130 kW per rack, and air-based cooling hits performance limits once racks exceed 40 kW. The Vera Rubin platform requires 250 kW or more per rack. Beyond about 200 kW, direct-to-chip liquid cooling becomes mandatory because air cannot transport heat efficiently at extreme rack-level concentrations.
The investment envelope around this scaling is substantial. CoreWeave has guided for revenue of $11.45 billion to $12.40 billion in the next annual period against capital expenditure of $28.63 billion to $33.40 billion, and reported $20.04 billion in debt at the end of the latest reported year. It has also recently secured an additional $8.11 billion in financing for contract-related infrastructure. Against that backdrop, any constraint that slows capacity delivery or raises downtime risk can affect how quickly contracted work converts into revenue.
Energy cost and thermal efficiency sit at the centre of that delivery discipline. A 100 MW data centre operating at full capacity consumes 876 million kWh annually, translating to an annual power bill of about $35 million at $0.04 per kWh and exceeding $122 million at $0.14 per kWh. Building Management Systems can function independently from IT workload patterns, a separation that can drive 20% to 40% overcooling and waste energy while localised hotspots emerge around high-density cabinets.
Reliability exposure rises at the same time. HVAC system failures constitute the second leading cause of data-centre outages after power system failures, while GPU downtime costs range from $562.9 to $2,360 per hour per node. Data-centre completion delays are estimated at $14.2 million in lost revenue, cost overruns and contractual penalties for each month of delay.
CETA System frames advisory-first optimisation as a practical response, where AI analyses sensor data and produces recommendations while operators retain oversight of critical cooling decisions. “The value is in surfacing the adjustment, but the operator still approves the change when conditions shift,” Lee noted, describing a model intended to balance efficiency gains with accountability. Within that approach, AI-driven analysis supports recommendations for fan speeds, coolant flow and temperature set points, alongside alerts when maintenance becomes necessary or failure probability increases. Vendor-agnostic integration is highlighted as central to adoption, with open-source integration tools interfacing with existing site equipment to standardise telemetry across legacy or new-build environments. Predictive maintenance and condition-based monitoring can help identify equipment degradation before failure and support planned maintenance windows instead of reactive response.
Regulatory obligations are adding a parallel driver for consistent measurement and reporting. The European Union’s Energy Efficiency Directive requires data-centre operators with installed IT power demand of at least 500 kW to report key performance indicators annually to a European database, including energy consumption, PUE (power usage effectiveness), temperature set points, waste heat utilisation, water usage and renewable energy share. In Hong Kong, a Buildings Energy Efficiency Amendment Ordinance extends statutory requirements to data centres for the first time, with mandatory compliance scheduled to commence on 20/09/2026 and energy audit intervals shortened from ten years to five years.
For accelerated-compute facilities, cooling has shifted from an engineering line item into a capacity, compliance and financial variable. CETA System argues that operator-supervised, advisory-first deployment models, backed by vendor-agnostic integration and condition-based monitoring, can help data-centre teams maintain uptime, control energy cost and evidence compliance as rack densities rise.
About CETA System
CETA System Co., Limited is a Hong Kong-incorporated technology company founded in 2017 that develops artificial-intelligence solutions for data-centre infrastructure, combining HVAC and chiller-plant energy optimisation with predictive maintenance for critical assets such as UPS systems, generators and chillers. Its vendor-agnostic platform integrates with existing building-management and DCIM environments and follows an advisory-first deployment model serving 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)
