为何人工智能数据中心冷却需要变得更智能

The job of a data center cooling system is simple in concept: absorb heat where it isn’t wanted and move it where it can be released safely. AI is making that job much harder, pushing rack power past 100 kW, with some AI clusters expected to exceed 1 megawatt within the next few years. Operators racing to add AI capacity must deploy quickly while protecting uptime, managing energy use, and safeguarding valuable hardware.

There is no single way to cool an AI data center. The right architecture depends on rack power, workloads, and facility requirements. Engineers typically reserve the most advanced liquid cooling for AI racks, while other approaches serve standard workloads and facilities without existing water infrastructure. Liquid cooling is becoming increasingly important, but this mix of approaches adds complexity: mitigating leak risk, supporting multiple architectures at once, and ensuring that cooling responds precisely as thermal demands change.

Moving heat more efficiently is no longer enough. The next step is to make cooling systems more intelligent through real-time motor control, precision flow sensing, and high-efficiency power management. Together, these technologies can make cooling more responsive, reduce energy consumption, and maximize uptime while laying the foundation for AI-enabled predictive maintenance.

Why Smarter Control Matters

Consider what happens when a liquid-cooled rack develops a leak. Many of today’s rack installations use on-off valves that stop the coolant when a sensor detects a leak, shutting down the entire rack until technicians can locate and fix the leak source. With a whole rack at stake, detection must be reliable. Yet the most common leak sensors, moisture-sensitive resistive ropes, can be limited in both accuracy and response time.

A more intelligent approach could detect leaks faster and control coolant flow at the server-tray level, helping keep unaffected parts of the rack running. That points to three capabilities working together: real-time motor control, precision flow sensing, and high-efficiency power management.

Real-Time Motor Control Keeps Coolant Moving

Inside the coolant distribution unit (CDU), pumps and fans form the mechanical heart of the cooling loop. Reliability is the top priority. Depending on the CDU placement, a single pump failure can take an entire rack – or worse, an entire row – of AI computing equipment offline.

Real-time motor control helps keep that system stable. Microcontrollers process sensor data and adjust performance accordingly, while motor drivers run pumps and fans efficiently.

Motor control also extends to the server tray. Valves at the CDU typically open or close, but linear valves at the tray level can regulate coolant flow much more precisely. As power density inside the tray continues to rise, that level of precision becomes more important.

Precision Flow Sensing Tracks Coolant at the Tray

Precise flow control requires precise flow measurement, so linear valves at the tray work alongside accurate flow sensors. Together, they enable more granular control of GPU and CPU die temperatures.

How the sensor is installed matters, too. Conventional flow sensors are inserted into the coolant line, creating an additional potential leak point. Non-invasive ultrasonic sensing can avoid that risk while also making installation easier and faster.

Moisture-sensitive resistive ropes are still commonly used for leak detection. Still, concerns about accuracy and response time are pushing engineers to evaluate alternatives such as humidity sensors, capacitive sensors, and negative-pressure CDU designs.

Flow readings also contribute to a broader operating picture. Combined with temperature, humidity, and pressure sensors, they can help detect changes that any single sensor might miss.

High-Efficiency Power Management Keeps Cooling Lean

The pumps and fans cooling AI racks consume power themselves, so their efficiency matters. Power management and motor control meet in the pump drive, where gallium nitride (GaN) power devices can improve overall pump-system efficiency by up to 2% compared with traditional silicon while enabling a smaller form factor. In server trays that still combine air and liquid cooling, fans remain critical, and many are shifting from 12 V to 48 V to reduce power consumption and improve efficiency.

Power delivery is evolving as well. CDU pumps already operate on input voltages ranging from 48 V DC to 380 V AC. As 800 V high-voltage DC (HVDC) distribution gains traction in parts of the industry, operators will need pumps designed to support it.

Edge AI Lays the Groundwork for Predictive Maintenance

Sensing can also become predictive. Edge AI can process data from a cooling system’s full sensor array, an approach known as “sensor fusion,” enabling faster response times and lower latency.

That matters for leak detection, which remains a bottleneck for reliable liquid cooling. Sensor fusion can improve detection speed and accuracy.

It also opens the door to predictive maintenance for pumps and other critical mechanical components. Combined with tray-level flow control, edge AI and sensor fusion point toward cooling systems that sense changing conditions, anticipate problems, and adjust automatically.

Building Intelligence into Every Architecture

No single cooling design fits every AI data center, and mixed architectures are likely to serve the industry for years. But every architecture, from rear-door heat exchangers to immersion tanks, shares many of the same building blocks: sensors to measure conditions; motor control to drive pumps, fans, and valves; power management to supply them efficiently; and processors to coordinate the system as a whole.

The gains come when those pieces work as one system. Precise flow sensing guides linear valves at each tray. Real-time motor control adjusts pumps and fans as conditions change, while more efficient power devices reduce the energy needed to move coolant. Sensor fusion can help flag leaks and failing components early enough for operators to act.

AI will continue driving computing densities higher. The data centers that keep pace will be the ones equipped with cooling systems intelligent enough to respond in real time and resilient enough to see trouble coming before it affects uptime.

About the Authors

Texas Instruments

Patrick Zeng is the sector general manager for Texas Instruments’ systems and marketing team, where he is responsible for data center thermal management. Over the past 11 years in the semiconductor industry, Patrick has held several leadership roles in TI’s sales and applications departments. In his current role, Patrick focuses on shaping strategy and driving technical innovation in data center thermal management, helping advance TI’s system-level capabilities in a fast-evolving market.