None of the three scenarios for how the AI power component squeeze plays out through 2028 put lead times back where they were before 2024. In the base case, 65% probability, silicon carbide and microcontroller lead times hold at 30 to 45 weeks through 2028, with power device deployment lagging GPU delivery by one to two quarters. Even the bull case, 15% probability, only gets lead times down to 20 to 26 weeks, still above pre-2024 norms.
That is not a forecast procurement teams can wait out. It is one they need to plan around now.
This scenario modeling comes directly from "The AI Power Shortage" — read the full report, Fusion Worldwide's Q2 2026 State of the Industry Report.
Vertiv entered the second quarter of 2026 with a project backlog already above $15 billion on AI-driven data center demand alone. Eaton's own Q1 2026 earnings call put a number on the scale of what is competing for capacity: 32 gigawatts of U.S. data center capacity under construction, about 70% of it AI-driven, against a total backlog of 228 gigawatts, roughly twelve years of build at 2025 rates. This is not a theoretical risk. It is already on the books at the companies building this infrastructure.
These backlog figures are cited alongside Fusion Worldwide's own sourcing data in "The AI Power Shortage" — read the full report.
Extend the horizon. Move forecasting and purchase order (PO) placement out to 52 weeks for power devices, SiC MOSFETs, and 32-bit MCUs. A six-month forecast no longer secures capacity, and buyers trying to fast-track 12-month POs into 18-month blanket orders mid-cycle are getting told the line closed months ago.
Qualify a second source before you need one. Single-sourcing any part on the PDU bill of materials, PMICs, inductors, or connectors, is now an unacceptable risk. Nearly every panicked call Fusion Worldwide gets right now traces back to a buyer with exactly one approved vendor on the PMIC line.
SECOND-SOURCING, IN PLAIN TERMS: Qualifying an alternate, approved supplier for a component that a design currently sources from only one manufacturer. On a constrained bill of materials, a qualified second source is the difference between a delayed shipment and a stalled rack when the primary supplier's lead time slips or allocation tightens.
Buffer strategic inventory. The carrying cost of safety stock on wide bandgap semiconductors is small next to the cost of a stalled AI deployment. The customers holding six months of SiC and MCU safety stock right now are not the ones calling asking for miracles.
Build for agility, not just performance. Work with engineering to qualify drop-in replacements for high-risk legacy MCUs, and treat an independent distribution partner as part of the sourcing strategy, not a fallback plan. The organizations that get this right will be the ones still shipping racks on schedule in 2027. The ones that don't will find a three-dollar microcontroller standing between them and a multi-million-dollar deployment.
Agility also means giving engineering a standing list of the highest-risk parts on the bill of materials, not a one-time review. Renesas, Microchip, STMicroelectronics, and NXP are issuing accelerated end-of-life and last-time-buy notices on legacy MCU families faster than most design teams are cycling their qualification lists. A quarterly review that flags newly announced EOL parts against the current design gives engineering a head start on requalifying a drop-in replacement before the last-time-buy deadline forces a rushed decision.
These four moves are the full strategic-sourcing playbook from "The AI Power Shortage" — read the full report, Fusion Worldwide's Q2 2026 State of the Industry Report.
The math behind these four moves is not complicated. Carrying an extra few months of safety stock on a SiC MOSFET or a 32-bit MCU costs a fraction of a percent of the rack it protects. A stalled AI deployment costs the full value of that rack, sitting idle, plus the revenue the cluster was built to generate, calculated the way Nvidia's own tokens-per-watt-times-gigawatts formula measures it. Every week a rack sits waiting on a component that costs less than a cup of coffee is a week of lost throughput on hardware that already cleared the far more expensive GPU and HBM line items.
That asymmetry is why the four moves below are not a hedge against a remote risk. They are a response to a squeeze that is already showing up in vendor backlogs, foundry utilization data, and the panicked calls Fusion Worldwide fields from buyers who single-sourced the wrong line.
For the broader context on why power components, not GPUs, became the constraint, see "The Real Bottleneck in AI Infrastructure Isn't the GPU." For a closer look at the single riskiest part on this list, see "What Is a PMIC, and Why Is It Holding Up AI Data Centers?"
All three pieces are drawn from "The AI Power Shortage" — read the full report, Fusion Worldwide's Q2 2026 State of the Industry Report.