Powering AI data centers: the role of power supply

From saying "Hey Siri" to your phone automatically organizing faces in photos, these seemingly simple AI features rely on the rapid computations of thousands of servers behind the scenes. These servers are concentrated in one place: data centers. You can think of an AI data center as a "smart factory" that continuously processes data from around the world. And to keep this factory running, what it needs most is a stable and robust power supply.

AI models are becoming increasingly complex, their computational power is growing, and so is their demand for electricity. The power consumption of a single AI data center may equal that of an entire neighborhood! Therefore, stable, intelligent, and energy-efficient power systems have become critically important.

But even when power is available, it can't be used directly. Just as household appliances can't be connected to high-voltage electricity, servers also require conversion and voltage regulation to operate safely. This is where the power supply unit (PSU) acts like a transformer, converting electricity from the grid into a form suitable for servers. Additionally, AI data centers cannot afford sudden power outages - a blackout could lead to data loss or system crashes. This necessitates uninterruptible power supply systems (UPS) and backup power to ensure seamless operation even during external power failures.

On the other hand, energy efficiency is also a priority. Waste must be minimized. High-efficiency power conversion technology can reduce energy waste and heat generation, which in turn improves overall cooling efficiency. For AI data centers, this is a win-win for both energy savings and environmental sustainability.

Moreover, power can become "smart." Modern data centers don't just need "power on demand"; they also require real-time monitoring of power status, automatic load distribution adjustments to reduce single-point stress, and integration with renewable energy sources like solar and battery storage to achieve green power supply. All of this is made possible by intelligent power management systems (PMS). It's like having a meticulous butler that ensures the entire AI center operates powerfully and efficiently.

AI can perform rapid computations, understand language, recognize images, and even predict the future. But behind all of this lies a foundation of stable power supply. Electricity is the "lifeline" of AI data centers, providing not just energy but also bearing the responsibility of sustainability, safety, and intelligence.

In the world of AI, data is the fuel, and electricity is the lifeline. To keep artificial intelligence running at high speed, the data centers behind it require powerful, stable, and efficient power supply systems. Two key technologies in this regard are power conversion and smart interconnects.

Data centers receive high-voltage alternating current (AC) from the grid, but servers, storage devices, and networking equipment actually require low-voltage direct current (DC). This is similar to how our phone chargers work - they can't be plugged directly into high-voltage electricity but must first convert it into a safe and usable form.

In AI data centers, the task of power conversion includes transforming the incoming grid voltage into the voltages and formats required by different devices. It also involves improving conversion efficiency to reduce energy loss and excess heat generation, as well as integrating with battery storage systems and UPS to ensure uninterrupted power during outages. As AI computing scales up and server density increases, power conversion is no longer just about "supplying power" but about "supplying power with precision" - every watt must be used just right, safely, and intelligently.

On the other hand, traditional power systems are one-directional - "deliver power and you're done." But AI data centers can't operate so simply. These servers run around the clock, with power demands fluctuating constantly. Therefore, the power supply system must also have real-time sensing and response capabilities. Through smart interconnects, the power system can monitor and report power consumption across different zones, automatically adjust loads to avoid local overloads or waste, and collaborate with other systems (such as cooling, battery storage, and energy management platforms) to provide visual data that helps engineering teams optimize operations and energy efficiency.

In short, smart interconnects make the power supply system "connected" and "intelligent." It doesn't just deliver power but also adjusts and predicts in sync with the data center's rhythm, enabling truly smart energy infrastructure. Whether training large language models, performing real-time image recognition, or processing vast amounts of global data, every task in AI relies on a reliable power supply system. Power conversion and smart interconnects are the two key technologies that make this power stable and intelligent. Building the next generation of AI data centers isn't just a race for speed - it's also a challenge of energy efficiency and smart operations. Starting with electricity, we can make AI faster, more stable, and more sustainable.

Arrow Electronics and its partners are committed to making AI data center operations more stable and improving energy efficiency. They have introduced numerous power supply solutions. Below, we will introduce these solutions by application type, including AC-DC power conversion, DC-DC power conversion, power conversion for AI processing units, and smart interconnects.

In AI data centers, the electricity from the grid is alternating current (AC), but servers and computing equipment actually require direct current (DC). This is where AC-DC power conversion comes in, acting as a "power bridge." Its functions include converting high-voltage AC into safe and stable DC for equipment use, ensuring the conversion process is efficient and low-energy to reduce heat generation, and supporting diverse output voltage requirements to flexibly supply various server architectures. With efficient AC-DC solutions, AI data centers can achieve stable power supply and energy-efficient operation, making it an indispensable cornerstone of the overall energy infrastructure.

Arrow-ST-Bourns 5.5kW AI server power supply — This power supply complies with the ORV3 standard, with a power density of up to 65W/In³. Its output voltage ranges from 48-54V (typical value: 48-50V), and its input voltage range is 180V AC to 305V AC. It achieves a peak efficiency of 97.5%, with a surge power 1.5 times the nominal output power. It supports firmware upgrades in the field (using a bootloader), active current sharing on top of droop, and hot insertion/removal (hot-plug). The rectifier can communicate via PMBus (up to 100kbps) and ModBus (up to 115kbps) and features an internally controlled variable-speed fan. The bill of materials (BOM) primarily uses ST components.

3.3kW 54V Bidirectional DC-DC — This EVAL_3K3W_BIDI_PSFB, introduced by Infineon, is a 3300W 54V bidirectional phase-shift full-bridge (PSFB) evaluation board featuring 600V CoolMOS™ CFD7 and XMC™. The converter achieves up to 98% efficiency in buck mode and up to 97% in boost mode. The EVAL_3K3W_BIDI_PSFB board is a DC-DC stage with telecom-level output, implemented using a phase-shift full-bridge (PSFB) topology module with bidirectional functionality. The board uses CoolMOS™ CFD7 and OptiMOS™ 5 in a fully SMD solution and incorporates an innovative cooling concept.

onsemi AI server power supply — onsemi's integrated solution leverages complementary products, including cutting-edge Si, SiC, and GaN technologies for power switching. It also integrates gate drivers, multi-phase controllers, 48V controllers, smart power stage (SPS) modules, smart fuses, and point-of-load (PoL) buck converters for power management. This provides a comprehensive power solution (ranging from 3kW to 25-30kW high-voltage DC transmission) tailored for data centers, covering everything from the grid to GPUs. This product portfolio enhances system efficiency and power density, reducing the system's footprint.

In AI data centers, servers and various computing modules have diverse voltage requirements. The role of DC-DC power conversion is to transform DC power between different levels, providing more refined and precise power control. The core functions of DC-DC power conversion include converting the main power supply's DC voltage into the voltages required by different modules, delivering stable power to high-performance components like CPUs, GPUs, and memory, and improving conversion efficiency to reduce losses and heat generation. DC-DC solutions act like "precision power distributors," ensuring that every component in the AI computing system receives just the right amount of power. This is a critical element in driving high-density, high-performance computing.

48V to 12V DC 1600W module — This REF_IBC_1600W_GAN, introduced by Infineon, is a scalable 48V to 12V regulated intermediate bus converter featuring 100V CoolGaN™ power transistors. It uses Infineon's CoolGaN™ 100V IGC033S10S1 in a 3mm x 5mm PQFN package with a typical RDS(on) of 2.4 mΩ and dual-side cooling. Additionally, it includes Infineon's EiceDRIVER™ 1EDN7136U, a GaN-optimized 1.8mm x 1.8mm gate driver IC with true differential input, and Infineon's XDPP1148-100B controller, a 4mm x 4mm digital power controller with firmware and a configuration GUI. It is paired with Vishay IHLP7575 inductors, which are 7mm height, and features onboard 3.3V and 5V internal power supplies for easy evaluation.

The core computing power of AI data centers comes from GPUs, TPUs, and other AI processing units. These units demand not only high power but also extreme stability and rapid response capabilities. This requires power conversion solutions specifically designed for AI processors. The primary roles of power conversion for AI processing units include providing low-voltage, high-current precision power, meeting the instantaneous load changes of AI computations, supporting high-speed switching and low-noise design to ensure stable processor operation, and optimizing energy efficiency to reduce heat generation and cooling pressure. Such power solutions are the key enablers for AI processing units to "operate at full power smoothly," ensuring that every deep learning training and inference task is completed quickly and stably.

MPQ1500-12V125-L48 Series 1500W DC-DC Converter — The MPQ1500-12V125-L48 is a high-efficiency, 1.5kW, digitally controlled, isolated intermediate bus converter which converts 40-60Vdc (nominal: 54V) input into an isolated, semi-regulated 12Vdc output, full protected from overcurrent, overtemperature, and overvoltage. It utilizes an industry-standard quarter-brick with optional PMBus™ interface. This DC-DC converter boasts a peak efficiency of 97.2%, supports 500Vdc input/output isolation, supports paralleling of three or more units, and features remote on/off control. It is RoHS-compliant.

In AI data centers, smart interconnects elevate the entire power system from "power supply" to "intelligent management." Their roles include connecting power modules, sensors, and management platforms to enable real-time monitoring and data feedback, assisting in anomaly detection and load changes to prevent outages or overheating, and collaborating with energy management and cooling systems to improve overall efficiency and stability. Through smart interconnects, power supply becomes visible, controllable, and more intelligent, providing a more reliable foundation for AI computing.

As artificial intelligence rapidly advances into the future, every computational instruction relies on stable and efficient power support. From power conversion to smart interconnects, from precision power supply to intelligent monitoring, power systems are not just infrastructure - they are the core driving force behind the reliable operation of AI data centers. Stable electricity is the heartbeat of the AI world's continuous progress. In the future, we will illuminate every leap forward in AI with smarter, more energy-efficient power architectures.