The Server — Supply Chain Decomposition

⚠ CHOKEPOINTSiC substrate supply highly concentrated — Wolfspeed and Resonac control >70% of high-quality SiC epi capacity. Long lead times constrain power module ramps for next-gen server PSUs.

⚠ CHOKEPOINTChina controls ~85% of rare earth processing. HDD voice-coil magnets and inductor cores depend on NdFeB supply chains routed through Chinese state-linked processors.

⚠ CHOKEPOINTShin-Etsu and Sumco together supply >55% of 300mm wafers globally. Any capacity constraint cascades into DRAM, NAND, and timing IC fabs simultaneously.

⚠ CHOKEPOINTSK Hynix supplies ~70% of HBM3E to NVIDIA. Samsung's yield issues on HBM3E delayed qualification. Micron ramping but still minor share. HBM is the single most concentrated supply chokepoint in AI server builds.

💡 INSIGHT48V power delivery (from rack PDU to GPU via 48V→1V direct conversion) is replacing 12V. Vicor's Factorized Power Architecture and MPWR's Titan controller are competing to be the standard, enabling >1kW-per-chip GPU power delivery.

⚠ CHOKEPOINTJapan (Epson, NDK, KDS) controls the quartz crystal resonator supply chain. MEMS (SiTime) is displacing quartz for high-spec AI server timing, but quartz still dominates volume. SiTime's MEMS fab is outsourced to TSMC, adding indirect TSA risk.

💡 INSIGHTCustom silicon is bypassing merchant switch ASICs. Hyperscalers (Google TPU, AWS Trainium/Inferentia, Microsoft Maia) are fabbing proprietary ASICs at TSMC, reducing dependence on NVIDIA GPUs and Broadcom switch ASICs for internal AI clusters.

⚠ CHOKEPOINTJapan and Taiwan dominate passive components. Murata alone holds ~40% of global MLCC capacity. AI GPU boards require 10x-20x more high-cap MLCCs than standard server boards, creating persistent supply tightness.

💡 INSIGHTCXL 2.0/3.0 memory expansion (CXL Type-3 memory expanders) will decouple DRAM capacity from CPU sockets, enabling pooled memory fabrics. Micron, SK Hynix, and Samsung are all shipping CXL DIMMs. This changes server memory topology by 2026-27.

⚠ CHOKEPOINTDelta Electronics and Lite-On (both Taiwan-listed) together supply an estimated 60-70% of AI server PSU units to SMCI, Dell, and HPE. Bel Fuse (BELFA) is the only meaningful US-listed pure-play but has much smaller share.

💡 INSIGHTPCIe Gen 6 (128 GT/s) will double GPU-to-NIC bandwidth by 2026, enabling 800G NICs to saturate PCIe. Marvell is positioned as a custom ASIC designer for hyperscaler NICs; NVIDIA ConnectX-8 will integrate NIC+GPU fabric management.

⚠ CHOKEPOINTNVIDIA GPU allocation is the real Level 0 constraint — not assembly capacity. SMCI, Dell, and HPE are GPU system integrators whose shipment volumes are dictated entirely by NVIDIA's H100/H200/GB300 wafer-out at TSMC CoWoS. GB300 adds HBM3E→NVLink→NVSwitch complexity further concentrating risk.

💡 INSIGHTCustom silicon (Google TPU, AWS Trainium 2, Microsoft Maia 100, Meta MTIA) is structurally displacing merchant GPU volume at hyperscalers for inference. This bifurcates the Level 0 market: merchant-GPU servers (SMCI/Dell/HPE) for enterprises; custom-ASIC ODM servers (Quanta/Wiwynn) for hyperscalers.