With the AI boom in full swing, AI applications have become an inseparable part of human life. In recent years, major global Cloud Service Providers (CSPs) have continuously increased capital investments to build data centers and secure computing resources. According to McKinsey (Note 1), by 2030, global data centers are expected to require $6.7 trillion to meet the demand for computing capacity. Amid such rapid expansion, providing highly efficient power supply to data centers has emerged as a critical challenge.
Unlike traditional data centers, AI data centers that use GPUs as their computing core experience a substantial increase in power demand, exhibiting exponential growth. For example, the current power consumption per rack reaches approximately 120 kW, several times that of a conventional rack, and is projected to reach 600 kW by 2027. At that point, a data center comprising 1,000 such racks would require roughly 1,000 MW of power—equivalent to the output of a nuclear reactor. Reducing power loss during such massive electricity transmission presents a major challenge. To address this, Delta has proposed an “grid-to-chip” solution (Note 2).
From medium-voltage grid facilities (10–33 kVAC) down to the chip-level 0.6 VDC, power undergoes multiple conversion steps, including AC-to-DC transformations, which result in significant energy losses and only about 87.6% end-to-end power conversion efficiency. To address this, Delta introduced the grid-to-chip HVDC solution, which reduces the number of conversion steps and thereby improves efficiency.
Based on the HVDC architecture, the next-generation Power Supply Units (PSUs) can convert three-phase 480 VAC directly into 800 VDC or ±400 VDC to supply computing racks. This new HVDC power supply architecture is expected to increase data center power efficiency by approximately 1.5%. Using the U.S. average electricity price of $0.13 cents/kWh, a 1,000 MW data center could save roughly $17 million per year in electricity costs—a significant amount. However, under the HVDC architecture, a single PSU may reach nearly 30 kW, posing substantial challenges for rack space and thermal management design.
Future AI data center power will grow exponentially, posing major challenges
WBG (Wide Band Gap) Semiconductor Power Modules Pave the Way for AI PSUs
Currently, most semiconductors are produced using silicon (Si) wafers, also known as first-generation semiconductors. However, due to the physical limitations of silicon, further performance improvements are constrained. Second-generation materials such as gallium arsenide (GaAs) have been used for decades, primarily in communication applications. At this stage, third-generation Wide Band Gap (WBG) semiconductors, which offer high energy efficiency and low power loss, become particularly important. The main WBG materials in use today are silicon carbide (SiC) and gallium nitride (GaN).
WBG semiconductors can operate at high frequencies with reduced switching losses, generating less heat, which allows components to be made smaller and achieve higher power density. They are already widely used in electric vehicles, solar power systems, and fast chargers. However, starting in 2023, due to a slowdown in the automotive and industrial markets, combined with aggressive expansion by Chinese silicon carbide (SiC) manufacturers, SiC substrate prices have dropped by more than 50%. As prices approach those of silicon-based materials, the cost-performance ratio has improved significantly, and WBG semiconductors are expected to see increased penetration across various applications (Note 3).
As a result, WBG semiconductors have become the ideal choice for next-generation AI PSU designs. In addition to increasing rated power and efficiency, they can also enhance dynamic response capabilities to handle load variations generated by GPUs during operation. In NVIDIA’s recently announced list of 18 Power Revolution partners (Note 4), more than half of the semiconductor companies are involved with WBG technology, including Infineon, ROHM, STMicroelectronics, ON Semiconductor, Navitas, and China Innoscience, drawing significant market attention.
The next-generation AI PSUs may adopt three-phase topologies, such as three-phase PFC and LLC architectures, which require an increased number of power components. By integrating diodes, transistors, and other components into a single module, it is possible to further enhance thermal management, power density, parasitic parameters, and reliability. This is particularly relevant as AI PSUs move toward liquid-cooling solutions, where the design of the cooling system reduces available internal space. In such cases, WBG semiconductor-based power modules provide an optimal solution. Moreover, by integrating additional peripheral circuits and components into the module, the performance of AI PSUs can be further elevated. Currently, Delta is internally developing high-performance WBG power modules suitable for AI PSU applications, aiming to optimize power transmission efficiency in data centers, reduce the Power Usage Effectiveness (PUE), and further promote sustainable, net-zero development.
The third-generation WBG semiconductor material – silicon carbide (SiC) powder – is transformed into SiC wafers through a series of processes, including crystal growth and epitaxy.
Note 1: McKinsey: The cost of compute: A $7 trillion race to scale data centers
Note 2: Delta: Grid-to-Chip Power Solutions for Gigawatt-Scale AI Data Centers
Note 3: DIGITIMES: China’s SiC substrate market continues to grow; price pressure drives new application opportunities
Note 4: anue - NVIDIA Promotes 800V HVDC Architecture, 18 Suppliers Join the List
BRAND NEWS ARCHIVE