Liquid Cooling for AI Data Centers: How Engineers Keep Chips from Overheating

AI computes in silicon — and heat needs a way out

In a data center, artificial intelligence relies on calculations taking place in microscopic semiconductors and on a thermal system that removes the heat they produce. Cold plates, pumps, manifolds and heat exchangers form part of this physical infrastructure. A project announced in Mumbai on October 12, 2026 provides a starting point for following that path.

Direct-to-chip liquid cooling removes heat from electronic components through cold plates in thermal contact with them and through liquid circulating internally in a dedicated circuit. The system carries energy to another circuit, the environment or a compatible thermal use. We will follow that route and solve a hypothetical 100 kW example, independent of the Indian project.

What was announced in Mumbai — and what 30 MW does not measure

According to a Vertiv announcement reproduced by Express Computer and Webnewswire on October 12, 2026, the company deployed a high-density direct-to-chip solution with CapitaLand India Trust at CLDC Mumbai 01 in Airoli, Navi Mumbai. The arrangement is described as a Secondary Fluid Network, or SFN, intended to support 30 MW of AI computing capacity for a hyperscale customer.

The description of it as India’s first deployment is a supplier claim. The announcement also describes manufactured and tested skids, contamination control, balancing and Level 4 commissioning. The pages reviewed do not provide an independent audit or measured efficiency, flow rate, chip temperature or water consumption results for this solution.

The announcement date should not be confused with the start of operation. In a release dated July 29, 2026, CapitaLand had already reported handing over its first liquid-cooled data center in Navi Mumbai to a hyperscale customer. Its factsheet presents infrastructure and cooling possibilities; it does not measure the heat load removed by the network announced in October.

The 30 MW figure is announced capacity for AI workloads. It is not a measurement of 30 MW removed by the liquid circuit or the rating of a single CDU. Nor does it allow a design flow rate to be calculated without knowing actual loads, the fraction captured by liquid, temperatures, fluids and architecture. The following exercise uses its own fictional inputs.

Vertiv deployment announcement reproduced by Express Computer →

Vertiv disclosure reproduced by Webnewswire →

CapitaLand India Trust — 1H2026 release; July handover →

CLDC Mumbai 01 — Factsheet →

Why does concentrating computing concentrate heat?

The electrical energy supplied to CPUs, GPUs and other components is converted mainly into heat across the servers. That heat must be removed continuously to keep components within their permitted operating conditions. Concentrating power in small areas creates challenges involving heat flux, uniformity, contact and the distribution of heat removal.

The problem goes beyond the room’s air temperature. Thermal resistances lie between the heat-generating region, the package, the interface and the cooling device. A poorly integrated cold plate can receive liquid under apparently appropriate conditions and still fail to keep the chip temperature within the required limit.

There is no universal rack power ceiling here that makes air cooling unviable in every project. The architecture depends on the equipment and application. Even in direct-to-chip systems, components without cold plates may continue to require airflow. The fraction of heat captured by liquid must be determined for the actual equipment set.

From processor to facility circuit

The path starts at the chip, crosses the thermal interface and reaches the cold plate. Its internal channels bring fluid into contact with a heat-transfer surface. The warmed liquid returns through manifolds to the coolant distribution unit, or CDU. Distribution among branches must serve the servers without assuming equal flow rates in every configuration.

The Technology Cooling System, or TCS, is the technology circuit that serves the equipment and connects it to the CDU. The Facility Water System, or FWS, is the facility circuit. In the illustrated liquid-to-liquid architecture, the heat exchanger transfers heat between these circuits without normal mixing of their fluids. Quality and operating requirements on one side should not automatically be applied to the other.

Hoses, pipes and connectors form part of the route: sealing, materials, contamination and maintenance conditions matter. Depending on its construction, a CDU may combine temperature, flow and pressure control with pumping and heat exchange. Liquid-to-air solutions and other architectures also exist; the name CDU does not define one single arrangement.

The heat still needs to leave the infrastructure. Dry coolers, chillers, cooling towers or heat recovery can participate in rejection or reuse, depending on the design, climate and demand. The liquid circuit does not destroy energy or eliminate this stage. The diagram is generic and does not depict the equipment used at the Mumbai development.

CoolChip CDU — liquid-to-liquid and liquid-to-air architecture →

What can a liquid carry?

For a single-phase fluid at steady state, treating properties as constant, we use Q̇ = ṁ c_p ΔT. Q̇ is the heat-transfer rate in watts, ṁ is mass flow in kg/s, c_p is specific heat capacity in J/(kg·K), and ΔT is the temperature rise in kelvin. A rise of 5 K has the same magnitude as a rise of 5 °C; this does not make absolute temperatures interchangeable.

Volumetric flow is V̇ = ṁ/ρ, with density ρ in kg/m³. Checking units helps: (J/s)/(J/(kg·K) × K) gives kg/s. Dividing by density yields m³/s. The balance estimates how much fluid carries a defined load under these assumptions; it does not calculate the processor’s junction temperature.

The fluid may be treated water or a mixture compatible with the system. Pure water should not be assumed in every direct-to-chip application. Composition and temperature change properties; viscosity also influences flow. Data should come from the relevant documentation. In this exercise, c_p = 4,180 J/(kg·K) and ρ = 997 kg/m³ are fixed approximations adopted for teaching.

A higher ΔT reduces theoretical flow — within limits

For the same load and properties, calculated flow is inversely proportional to ΔT. Allowing the fluid to warm up more reduces the quantity that needs to circulate in this balance. The table and chart show the relationship for 100 kW, and the chart also covers loads of 50 and 200 kW.

This result does not demonstrate that the chip tolerates any ΔT. Supply temperature, thermal contact, distribution among plates and equipment limits remain decisive. A correct total flow can conceal an underserved branch. Fluid temperature should not be raised without assessing materials, operating conditions and reliability margin.

Pumping power also enters the calculation

The fluid encounters losses in pipes, filters, valves, fittings, manifolds, heat exchangers and plate channels. Operating flow depends on the system curve and pump curve. Pressure loss cannot be obtained from the thermal balance alone and may change with flow, fluid and configuration.

For a defined segment, idealized hydraulic power is P_hyd = Δp × V̇. With Δp in Pa and V̇ in m³/s, the result is W. To estimate the electrical portion, P_el = P_hyd/η_overall, where the aggregate efficiency must match the pumping boundary considered. It is not automatically just the pump’s hydraulic efficiency.

We add a hypothetical total pressure loss of 100 kPa and an overall efficiency of 60% to the example. Thus, P_hyd = 100,000 × 0.0047991 ≈ 479.9 W. Dividing by 0.60 gives P_el ≈ 799.8 W, or 0.800 kW. This is estimated pumping power, not chiller consumption or total cooling consumption.

The chart varies Δp while holding the example’s flow constant and compares three fictional aggregate efficiencies. This teaches the expression’s sensitivity; it is not a pump or network curve. In a real facility, these variables cannot be chosen independently without analyzing operation and component characteristics.

Pressure loss and pipe diameter →

The CDU creates a boundary; the exchanger carries heat across it

In a liquid-to-liquid arrangement, the CDU brings two circuits together thermally without mixing their fluids during normal operation. The technology side must serve the equipment; the facility side connects to heat-rejection infrastructure. This separation helps organize fluid control and quality but does not remove the need to consider materials, pressure and integrity.

The expression Q̇ = UA ΔT_lm relates heat-transfer rate, overall heat-exchanger conductance and the logarithmic mean temperature difference under the applicable model. U depends on heat transfer and resistances; A is the relevant area. Stating a kW rating alone, without temperatures, flow rates and other conditions, is insufficient for comparing capacities.

In the arrangement described by OCP, approach can be expressed as the difference between TCS supply and FWS supply temperatures, measured at the specified boundaries. This difference is not the 5 K ΔT through the exercise’s cold plates. A real exchanger requires compatible thermal conditions; zero approach should not be stipulated as universal performance.

Nominal CDU capacities are tied to catalog conditions. The existence of a product with a particular rating does not demonstrate that it was installed in Mumbai. Nor should the number of units or facility redundancy be inferred from an announced workload capacity.

ACS Liquid Cooling Cold Plate Requirements — Rev. 1.0 →

PUE and WUE: boundaries and periods matter

PUE is the ratio of total facility energy to energy used by IT equipment. It is dimensionless. For the comparison discussed by DOE, the energies must refer to the same boundary and period. Energy used by pumps is one part of that infrastructure, not the entire numerator.

Site WUE relates liters of water consumed on site to IT energy in kWh over the period, usually a year. The result is L/kWh. Other approaches also include water used off site to generate energy; different definitions should not be compared as though they were the same metric.

A closed liquid circuit does not mean zero total water consumption. Heat may be transferred to infrastructure using a cooling tower or another water-consuming process. Climate, heat-rejection technology, temperature limits and operating regime influence results. Direct-to-chip can improve thermal management but does not guarantee lower PUE or WUE in every scenario.

The assessment should explain where energy and water are counted. An isolated index does not show chip temperature, contingency performance or service continuity. None of these metrics has been calculated for Mumbai in this article.

Cooling Water Efficiency Opportunities for Federal Data Centers →

Best Practices Guide for Energy-Efficient Data Center Design — revised July 2024 →

Refrigeration and COP →

Fluid, sensors and contingency: what supports reliability

A circuit can remove 100 kW in the balance yet remain unsuitable because of unbalanced distribution, poor thermal contact, incorrect pressure or incompatible fluid quality. OCP treats the TCS as a system: wetted materials, connectors, pipes, plates, fluids and the interface with the CDU must be assessed together.

Particles, corrosion and chemical compatibility can affect channels and seals. Quality needs to be monitored over time; filtration must suit sensitive components. Microbiology may matter depending on the fluid and conditions. Introducing water or additives without checking system requirements is inappropriate.

Condensation control involves dew point and local conditions. Measuring temperature, flow, pressure and signs of leakage helps identify deviations, but each sensor needs a defined function and response. An alarm alone does not demonstrate containment or recovery. It should not be assumed that every leak automatically shuts down all racks.

Contingency, redundancy, isolation and maintenance must be coordinated with IT continuity. An operating change can affect heat, pressure and availability simultaneously. Analysis should cover prevention, detection, containment and appropriate intervention without treating this study as a procedure for working on an energized rack.

Cold Plate Cooling Loop Requirements — Rev. 2 →

Thermal Guidelines — 5th edition reference card — Table 3.1 →

Pipe thermal insulation →

Explore the balance — without turning assumptions into specifications

The lab allows you to change heat load, ΔT, specific heat capacity, density, pressure loss and overall pumping efficiency. The curve shows calculated flow using the current load and properties. The point indicates the selected ΔT; an equivalent table provides values without relying on the image.

Calculations run locally in the browser. The model does not determine fluid mixture, chip temperature, actual pressure loss, pump curve, CDU capacity, PUE or WUE. Control ranges delimit the experiment and do not represent product limits. The Python code and Portuguese HTML can be downloaded to reproduce the exercise.

Download the educational simulator in Portuguese (HTML) →

Download the reproducible model (Python) →

From heat recovery to system integration

Recovered heat may find another use when temperature, demand, distance and integration make the solution appropriate. Economization and different heat-exchange technologies can also form part of the architecture. Two-phase systems offer other possibilities with their own properties and requirements; they are not the single-phase model calculated here.

These trends do not demonstrate specific results for the Indian project. The transferable lesson is to consider energy balance, hydraulic losses and reliability together. Reducing everything to a flow rate or efficiency index leaves part of the engineering out of the analysis.

Artificial intelligence may be digital, but its limits remain physical. Each chip creates a heat-removal demand that needs a reliable path to another use or the environment. Along that route, mechanical engineering helps keep the servers running.