Direct:
The hard part of building AI infrastructure increasingly extends beyond GPU supply. With the four largest North American cloud providers pushing roughly $650 billion of AI infrastructure capex in 2026, semiconductor output forecasts keep climbing, but more of the binding constraints now sit downstream in packaging, substrates, materials, and power.
Advanced packaging is the clearest case. CoWoS capacity has become a primary gating constraint, with HBM availability a coupled limit on the same accelerators. Below that, ABF substrate supply is tight enough that memory and AI customers are locking multi-year contracts, and PCB lead times have pushed past twenty weeks.
The upstream materials are where supply is thinnest. High-end ABF substrates use low-CTE T-glass cloth in their laminate structure to control package warpage and reduce solder-joint failure risk. Top-tier 2.8%-CTE T-glass has effectively one qualified supplier, Nittobo (TSE: 3110), whose new capacity does not come online until the end of 2026 with shipments in 2027. Taiwanese substrate makers expect a 10 to 20% T-glass gap in 2026, and more than 40% for high-end glass cloth in the second half. Copper foil, glass fiber cloth, and even drill bits are showing the same strain.
Then there is power. US interconnection queues have ballooned past 2,100 gigawatts, beyond total grid capacity, and analysts expect 30 to 50% of planned 2026 data center capacity to slip to 2028. TSMC (NYSE: TSM) has said AI demand is stressing the entire supply chain, well beyond chipmakers.
The read across to consumer hardware is indirect but real. The same HBM and substrate capacity feeding AI accelerators is what memory makers are prioritizing, which is part of why DRAM stays scarce and expensive. None of these bottlenecks clear quickly. T-glass shipments are a 2027 issue, and while new packaging capacity may land sooner, none of it relieves the squeeze in the near term.
Drafted with AI assistance against parallel reporting.
Sources:
- DigiTimes, “AI supply chain shortages shift from chips to equipment” (June 8, 2026, paywalled stub: headline and lead only)
- tech-insider.org, “Big Tech's $650B AI Capex Surge” (2026): hyperscaler 2026 AI infrastructure capex
- Manufacturing Dive / Omdia, “The great data center delay” (2026): 30-50% of planned 2026 data center capacity slipping to 2028
- Fusion Worldwide, “Understanding the T-Glass Shortage and Its Role in AI Growth” (2026): Nittobo as sole top-tier maker (2.8% CTE), capacity end-2026 / shipments 2027, 10-20% gap and >40% high-end glass cloth gap in H2 2026
- DigiTimes, “IC substrate shortage to persist in 2026” (Nov 20, 2025) and “AI chip demand tightens ABF substrate supply, three-year upcycle in sight” (Apr 8, 2026): ABF tightness, long-term contracts
- CNAS, “American AI Companies Can't Get Enough Chips” (2026): power interconnection queues >2,100 GW, materials constraints
- DigiTimes, “TSMC says AI demand is straining entire supply chain, not just chipmakers” (2026): TSMC supply-chain strain comment
Reddit: https://ift.tt/DEoJM3w