Power, connectivity and security for data centers

Modern AI data centers rely on a complex grid-to-chip power architecture designed to deliver electrical energy efficiently, reliably and with minimal losses from the utility grid all the way to the processor.

The typical power chain begins with medium-voltage grid connections and transformers, followed by UPS systems that ensure continuous operation. Power is then distributed via PDUs and increasingly through rack-level power architectures, before being converted by server power supplies and point-of-load (PoL) converters into the low voltages required by CPUs, GPUs and AI accelerators.

Driven by rapidly increasing AI workloads, this architecture is evolving towards higher power density, fewer conversion stages and improved efficiency. Emerging technologies such as wide-bandgap semiconductors, 48 V distribution, liquid cooling, high-voltage direct current systems (HVDC) and solid-state transformers are reshaping the future of data center power delivery and energy infrastructure.

Silicon carbide (SiC) and gallium nitride (GaN) are increasingly replacing legacy silicon semiconductors because they have lower switching and conduction losses. This enables reduced energy consumption and heat generation. In AI data centers, these wide bandgap semiconductors (WBG) help reduce operating costs and meet sustainability requirements while supporting increasingly demanding power requirements.

Both SiC and GaN offer efficiency gains, but they are used in different parts of the power chain:

• SiC is suitable for high-power applications such as grid interfaces, uninterruptible power supplies (UPS and solid-state transformers. Their high breakdown voltage, low switching losses and excellent thermal performance enable efficient operation at high power levels.
• GaN is suitable for high-frequency applications such as server power supplies and DC-DC converters. Its fast switching enables smaller passive components and higher power density.

AI workloads fundamentally change the way data centers are designed. This is because they require large clusters of GPUs and accelerators that consume significantly more power (even up to 1.2 kW per processor) than traditional server architectures.

As a result, rack power levels are increasing from tens of kilowatts to well above 100 kW. Some next-generation AI racks are expected to exceed 500 kW or even 1 MW+. These developments place new demands on power distribution, cooling and power conversion efficiency. Dense power systems are crucial to reduce the overall server farm space.

Increasing rack power levels are reshaping nearly every aspect of data center power electronics design. While enterprise data centers operate at around 5-15 kW per rack, hyperscalers already operate at up to 60kW per rack. The next-generation AI data centers go for 100-300 kW per rack with some going up to 100 kW-1 MW+.

Higher currents and tighter space constraints require more efficient power conversion, lower losses and improved thermal management. Engineers are therefore moving towards 48 V or even direct 800 V rack architectures, centralized power conversion, busbar-based distribution systems and advanced cooling technologies (e.g., direct liquid cooling). At the component level, wide-bandgap semiconductors (such as SiC and GaN) enable higher switching frequencies and greater power density, while new converter topologies help reduce losses and improve scalability.

As AI workloads continue to grow, optimizing efficiency, transient performance and thermal behavior will become increasingly important for future rack-level power systems.

Power conversion is one of the most important factors that influence overall data center efficiency. Electrical power is converted multiple times before it reaches the servers, storage systems, networking equipment, etc. Each conversion stage introduces energy losses that are dissipated as each, and these losses accumulate across the entire data center power chain. As AI workloads continue to increase, efficient power conversion becomes even more critical. Even small improvements in conversion efficiency can translate into significant energy savings at the data center scale.

As power densities continue to rise, optimizing efficiency is no longer limited to individual components. Instead, engineers increasingly focus on the performance of the complete grid-to-chip power chain, balancing efficiency, thermal behavior, power quality and system reliability to maximize overall data center performance.

Most of the energy in modern AI data centers is consumed by IT equipment such as the server, which includes CPUs, GPUs and AI accelerators. However, the supporting power infrastructure also has a major influence on overall energy consumption. Cooling systems, UPS installations, power distribution equipment and server power supply units all contribute to the total energy footprint.

In particular, server power supply units account for the largest share of up to 45% of total data center losses. However, remaining individual infrastructure losses may appear comparably small but accumulate across thousands of servers and many megawatts of installed power. Improving the efficiency of server power conversion and cooling systems therefore offers significant savings potential.

Open Rack architectures and advanced redundancy concepts are transforming the way power is distributed and managed in modern data centers.

Traditionally, data centers distribute power through a chain of isolated systems. This leads to inefficiencies, duplicated components and limited scalability. Open Rack architectures centralize and standardize power shelves, often using shared rectifiers and DC bus architectures. In practice, this means that Open Rack architectures rely on centralized power conversion, 48 V distribution and busbar-based power delivery. This reduces the number of conversion stages, simplifies maintenance and improves overall efficiency.

At the same time, redundancy models (such as N+1, 2N and distributed redundancy) ensure high availability and fault tolerance for mission-critical applications. The challenge lies in balancing efficiency, scalability and resilience while coping with rapidly growing AI workloads.

These architectural changes are driving innovation across the entire power chain, from UPS systems and rack-level power distribution to server power supplies and point-of-load converters.

High-voltage DC (HVDC) distribution and solid-state transformers (SSTs) are among the most promising technologies for future AI data centers.

HVDC architectures can reduce the number of power conversion stages, lower distribution losses and simplify the integration of renewable energy sources and battery storage systems. The higher the voltage level, the lower the required current. As such, higher voltages result in less heat and more power dense architectures.

Solid-state transformers (enabled by SiC and GaN semiconductors) offer fast power control, bidirectional power flow and significantly higher power density than conventional transformers. They can be up to a factor of 100 denser for an equal power rating.

While both technologies are still emerging in data center environments, interest is growing rapidly as rack power levels and energy demands continue to increase. First projections and deployments are already officially reported. Future AI data centers will likely combine HVDC distribution, wide-bandgap semiconductors and intelligent power management to achieve higher efficiency, flexibility and scalability.

Data center power systems must comply with a wide range of standards, covering power quality, electromagnetic compatibility (EMC), harmonics, efficiency and grid interconnection.

Important standards include:

• IEC 61000 series for EMC and power quality
• IEEE 519 for harmonic control
• Regional grid codes and efficiency regulations

Depending on the application, additional requirements may apply to UPS systems, server power supplies, renewable energy integration or energy storage systems. Specifically for server power supplies, 80 PLUS certification and included tests are legally mandated or required for market compliance in several regions such as EU and USA to meet corresponding performance standards.

Compliance is essential not only to ensure reliable operation and minimize disturbances to the electrical grid but also to improve efficiency and reduce operational risks. Accurate measurements of harmonics, transients, efficiency and power quality are therefore key in the development and validation of modern data center power electronics.

There is a reason why data centers are considered critical infrastructure: they are designated as Critical National Infrastructure (CNI) in the EU, and they have similar designations in other countries around the globe.

Modern society and businesses no longer operate “alongside” the digital world; they operate “within” it. When data centers fail or are compromised, they impact cascades from a single company to entire cities, states and nations.

Data center security is essential for the stability and resilience of societies:

1. Protection of essential public services

2. Stability of the global financial system

3. Resilience of critical infrastructure

4. Preservation of truth and democratic stability

Data center security is no longer just about protecting “company secrets” or “server uptime”. It is the digital armor of modern civilization. Any vulnerability, disruption of operation, data corruption, sabotage, espionage or any other incidents endangering operation continuity and data integrity directly impact the business model of data center operators. Regardless of the final impact, any kind of incident (and especially insider threats) has the potential to harm the foundation of trust that the data center business is based on.

Critical infrastructure and data center security is always a combination of cybersecurity and physical security: one does not work without the other. People screening is one element in a physical security chain. It ensures that no prohibited or potentially harmful items enter protected zones, and it prevents valuable assets from leaving those protected zones.

EN 50600-2-5:2021 (published internationally as ISO/IEC 22237-6:2024) defines the physical security requirements for data center facilities. Rather than focusing solely on access control, it establishes a comprehensive framework for protecting critical infrastructure through layered security measures. The standard covers:

• Designation of protection classes
• Prevention of unauthorized access
• Protection against fire and environmental hazards
• Deployment of security systems, such as video surveillance and alarms

A layered security architecture is the best-practice solution to data center physical security. This means that personnel are screened first at the building entrance for general threats and again at high-security data floor areas for data-bearing devices. AI-based millimeter wave screening is the ideal complement to conventional security measure. It enables fast, contactless screening while integrating with perimeter protection, access control and security operation centers.

Next-generation data centers are being shaped by challenges with computational performance, power efficiency and scalable connectivity. This evolution is driven primarily by rapidly increasing AI workloads.

Key infrastructure challenges include:

• Power density: AI servers require significantly more power per rack. This is driving the adoption of more efficient power conversion technologies and advanced power distribution architectures.
• Thermal management: With increasing rack power, liquid cooling is becoming essential to dissipate heat and maintain reliable operation.
• High-speed interconnects: Technologies such as PCIe 6.0/7.0, 1.6T Ethernet, and Ultra Ethernet are needed to provide the bandwidth and low latency required for distributed AI training.
• Memory and resource disaggregation: Technologies such as Compute Express Link (CXL) enable memory pooling and composable infrastructure, improving resource utilization.
• Signal integrity: As data rates continue to rise, accurate characterization of PAM4 signals, de-embedding, and high-speed compliance testing become increasingly critical to ensure reliable communication.
• Synchronization and timing: Precise clock synchronization is essential for coordinating distributed compute nodes, networking equipment and system monitoring across large AI clusters.