THE COMPONENT SIGNAL

Issue #009 · Monday, March 10, 2026 · AI Hardware + GaN + Thermal-EMI

GaN power devices — allocation tightening. Infineon's CoolGaN portfolio is on 16–22 week lead times, up from 12 weeks in January. EPC (Efficient Power Conversion) is slightly better at 10–14 weeks. The demand driver: 48V power stages for AI server racks and EV onboard chargers. GaN switches at 1 MHz+ switching frequencies generate EMI spectral content well into the GHz range — which means the shielding requirements around GaN-based converters are fundamentally different from silicon MOSFET designs. If you're switching from Si to GaN, re-evaluate your EMI mitigation before layout, not after.

Aluminum substrates — AI thermal demand. IMS (insulated metal substrate) boards are seeing 20% lead time increases. AI accelerator cards are consuming IMS capacity for power stage thermal management. If your design uses aluminum-backed PCBs for LED drivers or motor controllers, plan for longer lead times through H1.

Ferrite beads — sleeper shortage. Murata and TDK ferrite bead inventories are thinning on high-current automotive grades (>3A rated). Not yet on allocation, but distributor stock is below 8 weeks for popular 0805 and 1206 sizes. These typically co-spec with EMI shields — if your design uses ferrites at the IC level and shields at the board level, make sure both are on order.

NVIDIA's Blackwell Ultra (B300) pushes 800W per GPU. A standard 72-GPU rack draws 57.6 kW. The thermal challenge is well-documented. The EMI challenge isn't.

What nobody's talking about: Each B300 module has a VRM (voltage regulator module) running at 500 kHz–1 MHz, switching 100+ amps. The conducted EMI from these VRMs propagates through the power distribution network into adjacent boards and backplanes. At 800W, the current transients during load steps (GPU idle → full compute) generate broadband noise from 1 MHz to 500 MHz.

The shielding implication: Traditional open-frame rack designs assumed air cooling was the priority and EMI was secondary. At 800W per GPU, liquid cooling frees up the chassis for shielding — but almost nobody is designing for it yet. We've had three RFQs in the last month from companies building networking cards that sit adjacent to GPU trays. They're all discovering the same thing: their board passed EMC in isolation and fails in the rack.

Why this matters to you: If you're designing anything that sits in an AI rack — networking cards, storage controllers, BMC modules — the EMI environment just got dramatically worse. The GPU next to your board is dumping broadband noise into shared power rails. Design your power filtering and shielding for the neighbor, not just your own circuit.

When a thermal interface material (TIM) is placed between a shield can and a heatsink, it creates an uncontrolled impedance between the shield and the thermal management system. If the TIM is electrically conductive (graphite, copper-filled), it couples the heatsink to the shield — and the heatsink becomes part of the EMI antenna.

If the TIM is electrically insulating (silicone-based), the shield is thermally connected but electrically isolated. This is usually what you want — but only if the heatsink is also grounded independently.

The worst case: A conductive TIM connecting a shield can to an ungrounded heatsink. The heatsink floats at the shield potential, picks up noise, and re-radiates it at a frequency determined by the heatsink's geometry. We've seen this add 10–15 dB of radiated emissions at 1–3 GHz.

Fix: Decide upfront whether your heatsink is part of the shielding system or not. If yes: use conductive TIM and ground the heatsink. If no: use insulating TIM and ensure the heatsink has its own ground path that doesn't route through the shield.

Customer had a motor controller with a shield can and a heatsink separated by a graphite TIM. Radiated emissions were 12 dB worse with the heatsink installed than without it. The graphite pad was coupling the shield's internal noise to the heatsink, which re-radiated it. Switched to a silicone insulating pad. Emissions dropped back to shielded levels. The heatsink was helping thermally and hurting electrically — simultaneously.

If your design puts a heatsink on or near a shield can, verify whether the thermal interface material is electrically conductive. If it is, and the heatsink isn't grounded, you may be adding a radiator instead of removing heat. Check this before EMC testing — it's a 5-minute measurement that can save a retest cycle.