Signal & Supply
← Archive
October 21, 2026 POWER CHIPS

AI Servers Devour Dozens of One Unglamorous Chip Per Rack. Its Factory Lead Time Just Went From 21 Weeks to 40.

Every headline about the AI chip shortage is about GPUs. The chip that's actually running late is the one nobody brags about — the little part that regulates the power flowing into every server. In 2026, it became real enough to cut industry-wide server shipment forecasts.

Key takeaway TrendForce's April 15, 2026 report says lead times for power-management chips (PMICs) stretched from a 21–26-week baseline to 35–40 weeks, and for baseboard management controllers (BMCs) from 11–16 weeks to 21–26 weeks, as foundries funnel scarce 200mm ("8-inch") wafer capacity toward higher-margin AI-server orders first. TrendForce cut its 2026 total server shipment growth forecast from roughly 20% to 13% year-over-year as a direct result — even while holding AI-server shipment growth near 28–31%.

Run the capacity-rationing mechanism yourself

Two queues share one fixed pool of scarce 8-inch chip-factory capacity: AI-server orders and everyone else's. Drag the slider to reserve more of that capacity for AI orders, or press Run to animate today's approximate real-world tilt. Simplified model: the two queues' exact split isn't published — TrendForce reports one industry-wide PMIC lead-time range (21–40 weeks); this splits that single number into an illustrative "AI" vs. "general" pair to show the mechanism, with the ceiling pinned to the real reported figures.

0 10wk 20wk 30wk 40wk reported ceiling AI-PRIORITY ORDERS 23 wk GENERAL-SERVER ORDERS 23 wk
0%
AI share (slider)
0%
AI-order wait
23 wk
General-order wait
23 wk
Reported real range
21–40 wk

The plain version

Picture an AI server rack as a power-hungry little city block, with the GPU as the factory at its center. Before electricity from the wall can reach that expensive, sensitive chip, it has to pass through dozens of small conversion stations that step voltage down and clean it up. Officially, those parts are called power management ICs (PMICs), working alongside a quieter cousin called the baseboard management controller (BMC) — the rack's nervous system. An ordinary server needs a handful of these. A modern AI server, packed with GPUs pulling far more current at far more precise voltages, needs many times more of them.

Here's the twist: these chips aren't built in the shiny new factories you read about, the ones racing to etch ever-smaller transistors. They're built in older factories running 200-millimeter ("8-inch") silicon wafers — a factory format the industry mostly stopped expanding more than a decade ago, because everyone assumed the future was all about shrinking chips, not regulating power. Nobody poured money into more 8-inch capacity, because nobody expected AI servers to quietly become the biggest customer for the least glamorous chip in the building.

Now that they have, foundries are doing the obvious thing with a limited resource: handing their scarce 8-inch capacity to whoever pays the most, which is AI customers. Everyone else — ordinary servers, cars, appliances, anything else that needs a cheap power chip — gets pushed toward the back of the line. TrendForce, the industry's main forecasting shop, says the wait for these chips has nearly doubled this year, long enough that it's now cutting how many ordinary servers the whole industry can even ship in 2026. The GPU was never the only bottleneck. It just had the better PR.

The expert version

PMICs and BMCs are overwhelmingly fabricated on 200mm (8-inch) wafers using BCD (Bipolar-CMOS-DMOS) processes at mature-node geometries, typically in the 28nm–180nm-class range — not because the industry can't do better, but because there's no economic reason to. The power-delivery and high-voltage blocks that dominate these dies are sized by breakdown-voltage and headroom requirements set by the physics of the signal, not by transistor density. Shrinking the process doesn't shrink the power stage, and finer geometries raise sub-threshold leakage, which actively hurts standby efficiency on parts designed for multi-year product lifespans. That's why PMIC fabrication stayed parked on 200mm capacity that has barely grown in wafer-start terms in over a decade: most equipment vendors stopped building new 200mm tools years ago, pushing the industry onto a thin secondary market for refurbished "cores."

AI servers broke the assumption that this capacity was adequate. A GPU-dense rack needs far more point-of-load conversion stages, at higher current density, than a general-purpose server — multiplying PMIC and BMC demand per unit even before counting unit growth. TrendForce's April 15, 2026 release quantified the squeeze: PMIC lead times stretched from a 21–26-week baseline to 35–40 weeks, and BMC lead times from 11–16 weeks to 21–26 weeks, as foundries reallocated scarce 8-inch BCD capacity toward higher-margin AI-server orders. The firm cut its 2026 total server shipment growth forecast from roughly 20% to 13% year-over-year as a direct result, even as it held (and later raised) its AI-server shipment forecast to around 28–31% growth. Global 8-inch utilization is projected to climb to roughly 85–90% in 2026 from 75–80% in 2025, even as total 8-inch capacity contracts slightly — TSMC and Samsung have both been trimming 200mm lines, and unconfirmed reports of a Samsung 8-inch fab closure add further downside risk. It's a textbook capacity-rationing problem: fixed legacy-node supply, a new high-margin buyer, and a scarce process nobody expected to matter again.

Why it matters for tech + supply chain: The AI buildout's real pacing constraint in 2026 isn't the GPU everyone can name — it's a boring power chip made on factory technology nobody's expanded in years.

Why it matters for tech + supply chain: 8-inch BCD capacity, not leading-edge logic capacity, is the binding constraint on 2026 general-purpose server shipments — a legacy-node chokepoint created entirely by AI-server demand reallocation.

Get it in your inbox every morning