Data Center Power Demand and Power Device Trends — Key Issues for 2026

Large-scale investment in generative AI infrastructure is driving a sharp increase in data center power consumption. According to IEA estimates, global data center electricity consumption is expected to exceed 400 TWh in 2025, up from approximately 220 TWh in 2022. For data center operators where energy costs represent 30–40% of operating expenses, even a few percentage points of improvement in power conversion efficiency translates directly into billions of yen in annual cost savings. This rising bar of requirements is significantly raising the specifications for power devices used in PSUs (power supply units) and UPS (uninterruptible power supplies).

To put the scale in perspective, GPU racks equipped with AI accelerators in the H100/H200 class can draw 50–100 kW per rack in some configurations. Compared to conventional general-purpose server racks (5–10 kW), power density has increased by more than tenfold. As a result, demand is surging for a complete redesign of the entire power conversion infrastructure — including incoming power equipment, UPS, and PSUs — to handle high power density.

The Scale of the Power Surge — How Much Has It Grown?

The primary driver of this surge is the mass deployment of GPU clusters for generative AI. Major cloud operators — Microsoft, Google, Meta, and Amazon — executed data center investments on the order of hundreds of billions of yen during 2024–2025, and this trend is expected to continue through 2026 and beyond. NVIDIA's next-generation GPUs (post-Blackwell) are pushing power density even higher, with the design ceiling for power infrastructure being rewritten every two years.

In Japan, domestic data center development is also accelerating under the government's policy to strengthen the AI and semiconductor industry base. Large-scale DC investments are underway in Hokkaido (leveraging surplus power near the TSMC site), Kumamoto, Osaka, and Kyushu, driving a corresponding expansion in domestic demand for power equipment. Hokkaido in particular benefits from the combination of cold-climate cooling cost advantages and renewable energy, making it a preferred location for large-scale GPU-focused data center construction.

Concrete Changes in Power Device Requirements

Data center power conversion follows a multi-stage conversion path: utility grid → transformer → UPS → switchboard → PDU → PSU → voltage regulator (VR) → processor. Because losses accumulate at each conversion stage, improving overall conversion efficiency requires performance improvements in power devices at every stage.

The 48V Power Architecture in Detail — The Core of Design Change

The 48V DC power architecture led by OCP (Open Compute Project) is approaching de facto standard status in data center design among major cloud operators. The technical rationale: for the same delivered power, a 48V architecture requires only one-quarter the current of a 12V architecture, reducing resistive losses in wiring and connections by a factor of sixteen (P = I²R).

This transition significantly changes power device design requirements. While silicon MOSFETs optimized for low voltage and high current were the dominant choice in 12V systems, GaN-on-Si and GaN-on-SiC MOSFETs are increasingly being adopted in the 48V → 12V conversion (POL: Point of Load) stage. GaN MOSFETs in the 40–65V breakdown voltage class from EPC, Navitas, and GaN Systems are optimized for this application, enabling significant size reduction compared to silicon devices of equivalent die area.

Supplier Landscape — Who Will Capture This Demand?

For GaN-on-Si PSU devices, EPC (Efficient Power Conversion), Navitas Semiconductor, and GaN Systems (now under Infineon) have emerged as leading suppliers. For SiC UPS applications, Wolfspeed, STMicroelectronics, Rohm Semiconductor, and onsemi have established customer bases.

Japanese power supply manufacturers (Sanyo Denki, TDK, Cosel, etc.) hold meaningful market positions in data center PSUs and UPS systems and are in the process of transitioning to GaN and SiC devices. However, as procurement volumes for GaN and SiC devices surge, supply tightness for GaN devices in particular has become partially visible, increasing the need for long-term procurement contracts.

The pace of specification change in the data center segment is rapid — designs from two years ago can already be obsolete. Confirming with suppliers whether they have a roadmap for compatibility with next-generation GPUs is an effective way to contain design revision costs.

Procurement Implications — Can Supply Keep Pace with Change?

Connecting Data Center Power Demand to Renewable Energy Procurement

Major cloud operators (Microsoft, Google, Meta, Amazon) have committed to RE100 and carbon neutrality targets, meaning the surge in data center power demand is simultaneously creating large-scale demand for renewable energy. From a power device perspective, this creates a synergistic effect in which demand for solar and wind power conversion (PCS) and data center power supply demand expand simultaneously.

In Japan, plans for large-scale DC campus construction in Hokkaido, Kyushu, and the Chubu region are materializing between 2025 and 2026, making these facilities an important domestic demand source for domestic power equipment and power module manufacturers. The domestic power supply market for data centers is projected to grow at an annual rate of 15–20%, in contrast to the volatility of the EV-focused SiC market, characterized by more stable and predictable growth.

Power device demand for data centers is one of the few high-growth segments that will expand steadily as long as AI infrastructure investment continues. Unlike the fluctuations in EV-driven demand, this demand is supported by large-scale capital expenditure from cloud operators and is highly predictable. Incorporating the acceleration of SiC and GaN adoption in this segment into supplier competitive assessments is becoming increasingly important.