Power semiconductors hit 30-week lead times
- Industry sources said on April 27 that lead times for some power semiconductors have stretched to about 30 weeks as AI server demand rises. - TrendForce said power-management chip lead times for general servers are now moving from 21-26 weeks toward 35-40 weeks as suppliers favor AI orders. - AI server shipments are still forecast to rise more than 28% in 2026, keeping pressure on mature-node power chips. (trendforce.com)
Power semiconductors — the chips that convert and regulate electricity inside servers — are turning into a new bottleneck for AI infrastructure. Industry sources said April 27 that lead times for some parts have stretched to about 30 weeks. (semimedia.cc) These are not the headline chips like Nvidia graphics processors or high-bandwidth memory. They are the switches and controllers that move power from the wall to racks, boards and processors without wasting too much energy as heat. (ti.com) (infineon.com) The squeeze is tied to how AI servers are built. Bigger graphics clusters draw far more power than standard machines, pushing data centers toward 48-volt and 800-volt direct-current designs that need denser power conversion hardware. (ti.com) (semimedia.cc) Infineon has said older 54-volt infrastructure is nearing its limits for fast-growing AI workloads. Texas Instruments has been pitching chips and reference designs for 12-volt, 48-volt and 800-volt architectures, including a 30-kilowatt AI server power-supply design. (semimedia.cc) (ti.com) TrendForce said on April 15 that the shortage is no longer confined to premium AI boxes. It cut its 2026 server shipment growth forecast to 13% from nearly 20%, saying lead times for power-management integrated circuits are extending as capacity is steered to higher-margin AI systems. (trendforce.com) For those power-management chips, TrendForce said lead times are expected to move from 21-26 weeks to 35-40 weeks. It said AI servers require higher power density, so suppliers are allocating more 8-inch Bipolar-CMOS-DMOS capacity to AI-focused parts. (trendforce.com) That matters because many of these chips are made on older 8-inch wafer lines, not the newest factories used for cutting-edge processors. TrendForce said in January that global 8-inch capacity is projected to decline 2.4% in 2026 as TSMC and Samsung cut output. (trendforce.com) At the same time, demand is still rising. TrendForce said global AI server shipments are expected to grow more than 28% in 2026, with spending from Google, Amazon Web Services, Meta, Microsoft and Oracle helping drive the buildout. (trendforce.com) The immediate risk is not that AI server demand disappears. It is that ordinary server programs, board revisions and upgrade cycles wait longer for the less glamorous chips that make the power system work. (trendforce.com) (theregister.com) The result is a supply chain where the limiting factor is increasingly electricity hardware, not just compute silicon. In 2026, the race to add AI capacity is also a race to secure the chips that keep those racks powered. (semimedia.cc) (trendforce.com)
