Application Notes
AI Data Centers Are Going Liquid: Why Temperature and Pressure Monitoring Moved Into the Rack
AI racks draw hundreds of kilowatts each and air cooling can't keep up. Direct-to-chip liquid cooling is going mainstream — and every loop needs coolant temperature, pressure, flow and leak monitoring.
An AI training rack in 2026 can draw hundreds of kilowatts — more than an entire row of ordinary servers used to. Air cooling has hit its limit at that density, so the industry is going liquid: coolant runs through cold plates bolted directly to the processors, carrying heat away to a coolant distribution unit (CDU) and out to the building. Direct-to-chip liquid cooling has gone from a niche for supercomputers to the default design for new AI halls in the space of a few years. And every liquid loop is, at its core, a measurement problem: coolant temperature, pressure, flow and leaks. The sensors in the loop decide whether it protects the servers — or becomes the failure.
The numbers behind the shift
Industry analyses project direct-to-chip cooling at roughly 44% of the data-center cooling market in 2026, with the 251–500 kW rack-density class leading — the power band where a single cooling fault can take a large slice of expensive computing offline. Hyperscale AI facilities are driving the change: at these densities, a localized cooling fault is no longer a facilities nuisance, it is a direct hit on computing capacity. The CDU has become the heart of the hall — pumps, heat exchangers, filters, sensors and controls in one cabinet, feeding cold plates across a row of racks.
What a liquid loop actually measures
A modern CDU watches at least six things continuously: supply temperature, return temperature, pressure, flow rate, humidity and coolant level — plus leak detection, pump vibration, filter restriction, coolant conductivity and valve position. Each measurement answers a different question:
Supply and return temperature — and the difference between them — tell you whether heat is actually being removed. A narrowing differential means the cold plates are not pulling heat, or the heat exchanger is losing capacity.
Pressure and pressure differential across filters and pumps tell you about clogging and pump health. A rising differential across a filter means it is loading up; a falling pump differential means the pump is wearing.
Flow rate confirms the loop is moving at all — a pump can be running while a valve is shut or a line is blocked.
Leak detection is the non-negotiable. Water and servers share the same rack now, so a micro-leak caught early is a maintenance ticket, and one caught late is an outage.
Cold plates expose temperature and pressure changes before room sensors register anything, which is why the sensing has moved into the loop rather than staying in the room.
From alarms to prediction
The bigger change is what happens to the data. Every liquid loop now carries dozens of sensors, and that stream is being fed into data-center infrastructure management (DCIM) and AIOps platforms. Instead of a scheduled inspection every 30 or 90 days, operators watch temperature differentials, pressure changes, pump speed, coolant quality and flow imbalance continuously — and the software flags pump wear or micro-leaks weeks in advance, activates a standby pump within seconds, isolates a leaking branch, or shifts workload away from a rack that is running hot. In March 2026, ZutaCore introduced HyperCool Cloud, linking direct-to-chip hardware with near-real-time CDU telemetry and fleet-level alarm workflows — the announcement framed the operating data as the thing that separates a normal load-driven thermal swing from a real fault.
The same loops are also turning waste heat into an asset. Warm-water loops running at 45–60 °C can feed municipal district heating — data centers in Finland, Sweden and the Netherlands already do — and the EU Energy Efficiency Directive now treats thermal reuse as a compliance metric. At those temperatures, the supply and return measurements are not just control signals; they are the meter for a heat product.
What this means for sensor selection
For the people who build and operate these loops, the sensing requirements are specific:
Temperature points multiply. Supply and return at the CDU, at manifolds, and at cold-plate inlets and outlets — all of them Pt100-class measurements, because CDU control works on small temperature differentials, and a sensor that drifts a degree is a control loop that hunts. Stability and interchangeability matter more than exotic accuracy.
Pressure is a diagnostic, not just a reading. Loop pressure and differential pressure across filters and pumps are the earliest indicators of clogging and wear. A pressure transmitter that holds its zero over years is worth more than one with a slightly better datasheet number.
The protocol is the product. CDUs and DCIM platforms talk Redfish, SNMP, TCP/IP, Modbus and BACnet. A temperature transmitter module that outputs 4–20 mA or RS485/Modbus drops straight into that world; a sensor with no standard output does not.
Retrofits are a real market. Existing halls are adding liquid loops around air-cooled equipment, and sensor gateways and monitoring retrofits can add temperature, pressure and leak visibility without replacing the primary cooling equipment.
The trend was on show at Data Centre World Asia 2026 in Singapore (September 29–30), where vendors presented modular liquid-cooled data centers with CDUs featuring redundant pumps, leak detection and real-time temperature and pressure monitoring — the monitoring no longer an add-on, but part of the cooling product itself.
Where the sensing goes
Liquid cooling turns the data center's thermal system into a measured, controlled process — and measured processes are what industrial sensors are for. Universal Ruike supplies the measurement points: Pt100 probes in custom lengths and fittings for coolant supply, return and manifold points, industrial pressure transmitters for loop and differential pressure, and temperature transmitter modules with 4–20 mA and RS485 outputs for DCIM integration. If you are building CDUs, manifolds or retrofit monitoring for liquid-cooled halls, tell us the loop media, temperature and pressure range, and the protocol the controls expect — and we will match the sensor to it.
Related products: Platinum Resistance Sensors · Industrial Pressure Transmitter · Temperature Transmitter Modules
中文
