The race to build artificial intelligence data centers is encountering a less visible bottleneck than GPUs, HBM memory, or network switches: power semiconductors. These components are much less media-focused, but without them, a server cannot convert, stabilize, or distribute the energy its processors, accelerators, memories, and network cards need. And now they are starting to become scarce.
The pressure is already evident in delivery timelines. According to TrendForce, lead times for some PMICs, the integrated circuits managing power within servers, have gone from 21-26 weeks to 35-40 weeks. The firm has also lowered its forecast for global server shipment growth by 2026: from around 20% year-over-year to about 13%. Demand remains strong, but the supply chain cannot keep up at the same pace.
AI servers consume more power and require more electrical control
The problem stems from a simple reality: AI servers consume much more energy than conventional servers. A rack with the latest accelerators not only needs more total power but requires much finer, more stable, and multi-level electrical delivery. Each GPU, CPU, memory, NIC, or board demands specific voltages and quick responses to sudden load changes.
This is where PMICs, MOSFETs, and other power and analog components come into play. A MOSFET functions as a highly precise electronic switch: enabling rapid and efficient opening or closing of current flow. In an AI server, thousands of such tiny electrical decisions happen constantly. If these components are delayed, the entire server stalls—even if GPUs are available.
The architecture of new systems also increases dependence on these chips. Multi-stage power designs, direct power feeds to high-consumption modules, and the need to minimize thermal losses mean each server incorporates more power management components than before. AI not only demands more computation but also requires more sophisticated power electronics.
This shift is affecting suppliers like Infineon, Texas Instruments, and onsemi, but it also opens opportunities for second- and third-tier manufacturers in Taiwan and other Asian markets. When major suppliers prioritize higher-margin products for AI servers, unmet demand shifts toward less known alternatives, especially in analog and power components.
The new scarcity reaches less visible silicon
In recent years, the debate has centered on NVIDIA, AMD, HBM, TSMC, or ASML. That makes sense, as these are associated with the most expensive and complex parts of AI infrastructure. But the case of power semiconductors reminds us that the actual supply chain is much broader. A data center isn’t built solely with accelerators; it also needs boards, power supplies, controllers, materials, cooling, substrates, capacitors, cabling, networks, and available energy.
TrendForce explains that suppliers are allocating capacity to higher-value products linked to AI servers. Meanwhile, Samsung plans to close its S7 8-inch fab in Korea, which could add pressure on the available capacity for PMICs produced in mature processes. The paradox is clear: a critical part of AI depends on manufacturing nodes that aren’t always at the cutting edge but are hard to ramp up quickly.
8-inch wafers remain vital for many analog, power, and management components. They lack the glamour of 2nm or 3nm nodes but support much of industrial electronics. When demand surges simultaneously in servers, electric vehicles, renewables, defense, and automation, mature fabs also face saturation.
Infineon is already seen capitalizing on this opportunity. The German company has raised its forecasts for 2026, driven by demand for power solutions in AI data centers, expecting revenues from AI data center applications to reach approximately €1.5 billion in fiscal year 2026 and around €2.5 billion in 2027. This is not a marginal figure: it shows that power management has become a growth business within AI infrastructure.
South Korea aims to reduce dependence on SiC and GaN
South Korea is also taking action. The Korean government has focused on next-generation power semiconductors, especially those based on silicon carbide (SiC) and gallium nitride (GaN). These materials better withstand high voltages, elevated temperatures, and fast switching than traditional silicon, making them attractive for electric vehicles, power grids, defense, renewable energy, and data centers.
In March, the Ministry of Trade, Industry, and Energy launched a task force for next-gen power semiconductors, aiming to increase domestic technological self-sufficiency from 10% to 20% by 2030. The strategy includes a technological roadmap, R&D projects, regulatory reviews, and deployment in sectors like national power grids, AI data centers, and defense systems.
Busan is positioning itself as a key part of this strategy. The city was selected for two projects by the Ministry and secured 20 billion won in government funding. The local plan envisions a total investment of 28.6 billion won to expand 8-inch compound semiconductor infrastructure, including a second SiC fab scheduled for 2027.
This makes industrial sense. SiC and GaN are not just technical upgrades; they reduce losses, improve energy control, and enable more compact designs. In AI data centers, where every watt counts and cooling is a major cost, boosting power conversion efficiency can directly impact operational density and total consumption and rack density.
PwC also points in this direction in its 2026 global semiconductor report, forecasting that the worldwide market will surpass one trillion dollars by 2030. The server and network segment is expected to be the fastest-growing, with an 11.6% annual rate. It also warns that data center power consumption could more than double by 2030, fueled by AI demand.
Thus, power semiconductors are no longer fringe components. Their shortages delay servers; inefficiencies increase energy costs; and an overly concentrated supply chain exposes AI infrastructure to new bottlenecks.
AI is forcing a comprehensive view of the entire tech supply chain. Initially centered around GPUs, then HBM memory, then networks, substrates, cooling, and power. Now, attention shifts to the chips that ensure this power reaches hardware reliably. Though less glamorous, these components are beginning to determine who can deploy AI capabilities timely and who will face nearly year-long delays for seemingly invisible parts.
Frequently Asked Questions
What are PMICs, and why do they matter in AI servers?
PMICs are power management integrated circuits. They control the electrical supply to different server components and are essential for stability, efficiency, and safety in high-power systems.
What is a MOSFET?
A MOSFET is a transistor that functions as an electronic switch. It allows control of current flow with speed and precision, necessary in power supplies, converters, and energy control systems.
Why have delivery times extended so much?
The surge in demand for AI servers has increased the need for power chips, while manufacturing capacity in mature processes like 8-inch wafers isn’t growing at the same rate.
Why is South Korea investing in SiC and GaN?
Because these materials are more suitable than traditional silicon for high-power, high-frequency, and energy-efficient applications like data centers, electric vehicles, power grids, and defense.
via: biz.chosun and PWC