Supporting technologies for AI data centres – power electronics and digitalisation for resilient systems

Thermal management plays a central role in AI data centres: it reliably removes the waste heat produced by compute-intensive applications and keeps processors and accelerators within a suitable temperature range. This is decisive because high thermal loads can impair the performance, service life and operational safety of the hardware. At the same time, the growing packing density in AI racks increases the requirements for energy-efficient and precise cooling.

The most important technology trends include direct-to-chip liquid cooling, in which cold plates absorb the heat directly at the CPU or GPU, and rear-door heat exchangers for cooling the exhaust air at the back of the rack. In addition, immersion cooling, hybrid air-liquid systems and continuous temperature measurement are gaining importance. Sensors, intelligently controlled pumps and fans, and DCIM- and AI-based systems make it possible to adapt cooling performance dynamically to utilisation and thermal load.

The semiconductor industry also contributes to thermal management with solutions at chip, package and system level. These include thermally conductive housing materials, optimised TIM layers (thermal interface materials), micro-structured cold plates and integrated microchannels that reduce the thermal resistance between chip and cooling system. At the same time, power semiconductors made from silicon carbide and gallium nitride improve the efficiency of the power supply and thus reduce conversion losses and additional heat. Research approaches such as porous membranes for evaporative cooling also show the potential that new materials offer for even higher heat flux densities in future.

Particularly relevant at present is the combination of these technologies with predictive thermal monitoring. Operators increasingly rely on integrated, data-based concepts that capture thermal conditions in real time, detect maintenance needs early and reuse waste heat as efficiently as possible. Liquid cooling, improved chip packages, efficient wide-bandgap power semiconductors and intelligent pump and fan controls are becoming central building blocks of modern AI data centres.

Renewable, locally used power supply architectures suitable for microgrids are becoming the strategic basis in AI data centres: they couple local generation, storage, UPS and intelligent control in order to supply high loads reliably and with as little strain on the grid as possible. With AI workloads in particular, this is important because power density, availability requirements and grid connection times are all rising at once. Current trends are hybrid solutions with solar, wind, fuel cells and BESS, grid-interactive UPS concepts, microgrids for island operation, and software-supported energy management extending to load shifting and direct on-site use.

Particularly relevant at present are architectures that combine several sources into a flexible energy system and thus bring together security of supply, sustainability and scalability. On the semiconductor side, this development is supported above all by wide-bandgap devices such as SiC and GaN, along with HVDC and 800 V approaches, more compact converters and more intelligent inverter functions.

Solid-state transformers (SSTs) take on a key role in AI data centres: they are intended to convert medium voltage directly into a compact 800-volt DC supply for power-dense AI racks. This shortens the previously multi-stage conversion, supports bidirectional energy flows to battery storage and improves power quality through active control. At the same time, they can balance rapid load steps during the training of large models within a short time. Current trends are therefore modular SST architectures for medium voltages up to 36 kV, higher power densities, digital condition monitoring, and close coupling with HVDC distribution, UPS systems and energy storage.

At the centre are fast power semiconductors: SiC MOSFETs and SiC diodes enable high blocking voltages, low switching losses and high switching frequencies. IGBTs remain relevant for certain power ranges; GaN devices open up prospects for particularly fast, compact converter stages. These trends are supported by advances in wide-bandgap semiconductors, integrated gate drivers, high-frequency transformers, cooling and digital control. Advanced packaging concepts reduce parasitic effects and improve reliability, decisive prerequisites for SSTs to work efficiently, space-savingly and in a grid-friendly way in AI data centres.

In AI data centres, solid-state circuit breakers take on the role of an ultra-fast protective switch for high-density DC infrastructures: they isolate faults within microseconds, limit damage and increase availability in networks with strongly fluctuating loads. Relevant here are power semiconductors such as SiC MOSFETs, IGBTs and increasingly GaN devices, complemented by gate drivers, sensors, control logic and components for energy absorption. 

Technologically, the market is shifting clearly towards direct DC distribution, selective protection concepts and digital monitoring. SSCBs are regarded as a building block for intelligent protective devices that not only disconnect but also supply data for diagnostics, remote monitoring and predictive maintenance. In parallel, coupling with solid-state transformers is gaining importance, because AI workloads force high power density and rapid load changes.

Particularly relevant at present are advances in SiC: higher dielectric strength, lower switching losses and more compact designs make the technology more attractive for DC distribution in the data centre. Added to this are improved packaging concepts, faster drivers and more robust protection algorithms that turn a pure switch into an intelligent protection system. That is exactly where the trend lies: away from the mechanical interrupter, towards a connected semiconductor solution for fast fault isolation and more system stability.

In AI data centres, an independent power supply secures operability during grid disturbances, dampens load peaks and bridges the time until generators or other backup paths take over. At the same time, UPS, and BESS too, are increasingly understood as an active building block of the energy architecture, for example for grid support, peak shaving and greater resilience.

Technologically relevant in UPS and BESS are above all power semiconductors for AC/DC and DC/AC conversion, bidirectional converters, battery management, and measurement and protection functions at cell and module level. In practice, the requirements are shifting due to high power densities, rapid load changes and the integration of lithium-ion systems, often with a focus on LFP; at the same time, NiZn, sodium-ion and other alternatives are also being discussed.

The most important trends include grid-interactive UPS concepts, greater BESS integration, higher power density, smaller footprints and more software-supported energy management. Particularly relevant for semiconductor manufacturers are SiC and GaN-based power stages, because they reduce switching losses and enable more compact designs; added to this are new approaches in high-voltage battery monitoring, protective circuits and modular, bidirectional platforms.

For UPS in particular, current semiconductor innovations are important that improve efficiency, dynamics and reliability at the same time: faster switches, more integrated drivers, more robust isolation and more precise monitoring of the battery paths. The uninterruptible power supply thus moves up from a classic emergency power system to the intelligent power hub in the AI data centre.

Highly efficient power conversion and distribution is the invisible workhorse in AI data centres: it ensures that high computing loads are supplied with as few losses as possible, high availability and the densest possible construction. For this, power semiconductors such as SiC and GaN MOSFETs, intelligent power modules, digital controllers, gate drivers, galvanic isolation devices and precise current sensors are particularly relevant; in practice, 400 V and 800 V DC architectures, intermediate bus converters and vertical power distribution are also coming into focus.

The most important technology trends include the move to higher DC voltages, shifting conversion closer to the chip, more efficient intermediate stages and greater digitalisation of power management. Added to this is the trend towards modular, scalable supply concepts, because AI clusters are reaching ever higher power densities and classic AC distribution is running into efficiency limits.

Particularly relevant for this development are current innovations from the semiconductor industry: more powerful SiC and GaN solutions for high switching frequencies, improved 400 V SiC MOSFETs, more integrated power modules and new 800 V HVDC approaches for AI data centres. The industry is thus working on fewer conversion stages, lower losses and more compact designs, exactly where the power demand of AI infrastructures is becoming the central bottleneck today.

(Cyber-)secure connectivity and edge intelligenceare increasingly becoming the operational layer between infrastructure, operations and security in AI data centres: they link monitoring, predictive maintenance, energy optimisation and cybersecurity with the DCIM systems (data centre infrastructure management) and enable decisions closer to the plant rather than exclusively in the cloud. Relevant semiconductor devices here are above all secure network and interface ICs, processors for edge AI, memory, sensors, and security elements with hardware root of trust, encryption acceleration and secured update paths.

The most important technology trends include hybrid edge-to-cloud computing, 5G and high-speed networking, zero-trust approaches, end-to-end encryption, and greater integration of AI directly into gateways, cameras, controls and measuring systems. In data centres, this is not only about more computing power, but also about lower latency, higher resilience and more precise control of energy and cooling systems.

Particularly relevant at present are chip innovations such as specialised edge APIs, energy-efficient AI accelerators, integrated security functions and trusted execution environments that process data locally while also protecting it better. The semiconductor industry is thus responding to the need for secure, scalable infrastructure that can be monitored in real time while reducing the growing attack surface of complex AI data centres.

Technical solutions for increasing resilience are becoming ever more important in AI data centres: they secure power quality, withstand load peaks, stabilise interaction with the grid and prevent individual failures from impairing the availability of entire GPU clusters. Particularly relevant for this are power semiconductors such as SiC and GaN devices, intelligent power management ICs, sensors for condition monitoring, and controllers for UPS, battery and distribution architectures. The trend is clearly towards modular, scalable energy and backup concepts that can be expanded quickly while responding to volatile AI loads. At the same time, grid-interactive systems, redundancy concepts such as N+1 or 2N, and digital monitoring solutions are gaining importance, because they bring efficiency and resilience together.

From a semiconductor point of view, higher dielectric strengths, better switching losses and more compact power densities are particularly relevant at present. Infineon points, for example, to new battery backup approaches for AI data centres, while technical articles highlight SiC MOSFETs and modern UPS topologies as building blocks for greater robustness and efficiency. For resilience solutions, this means less energy loss, faster response times and more flexibility in building AI infrastructure that remains reliable even during grid fluctuations and load changes.