# Nomura report : You're not deep enough in the AI Semiconductor stack yet. You...
Canonical: https://social-archive.org/yena/NTZYIqysa9
Original URL: https://x.com/glocalinvestor/status/2072011259675824388
Author: Arvind Srinivas
Platform: x
## Content
Nomura report : You're not deep enough in the AI Semiconductor stack yet. You're fishing with the crowd. Buying GPU, CPU or foundry are a crowded trade now. Nomura released the map to the depths. Here's the rundown: You'll surely want to bookmark this Investment implications run much deeper than GPU demand. For decades, chip performance gains came from one lever: shrink the transistor. At 3nm and below, that lever is breaking. > New growth model: three pillars: 3D transistors (GAA evolving to stacked cFET), backside power delivery (separating power and signal networks, reducing IR drop), and materials substitution. The capex ramp starts 2026. Mass production hits 2027. > Photolithography: High-NA EUV tools cost $400M each and require metal-oxide photoresists that run $10,000–$40,000/gallon. Current EUV resists cost $5,000/gallon. That's a 2–8x unit price step-up in a material most investors have never heard of. This is before high-NA even hits mass production (2029). > Advanced packaging: SoIC hybrid bonding replaces microbumps with direct Cu-to-Cu connections, taking AI chip bandwidth from 200 GB/s to 1 TB/s. It requires nanometer-level wafer flatness, meaning CMP steps increase from 45–55 to 55–70 per chip (20–30% jump). BESI is positioned as the dominant equipment supplier here. > Glass substrates: Glass-core substrates replace ABF organic substrates for large, high-power AI chips. Lower CTE, better flatness, lower signal loss. Broadcom adopts first for switch ASICs in 2027. Intel follows. The bottleneck is TGV (Through-Glass Via) formation - laser drilling, etching, metal fill, planarization. Companies with proprietary TGV process IP (LPKF, E&E) are entering this supply chain now. > Photonics: Indium phosphide supply stays tight 2025–2027 due to export controls and yield constraints. The alternative is photonic SOI wafers: 25% the cost of InP, scalable for CPO. Soitec owns ~70% of global photonic SOI share. Demand enters rapid growth by 2027. > Silicon wafers: Three technology waves (backside power delivery, wafer-bonded NAND, photonic SOI) stack onto baseline demand growth of 5%, adding 4–6 percentage points annually. Total 12-inch silicon wafer demand approaches 10% growth p.a. Supply cannot keep pace. Gap opens in 2027 Pricing power returns to GlobalWafers, Shin-Etsu, SUMCO. > TSMC the amplifier: $70B capex in 2027 alone. 26 advanced fabs globally. Increasing localized procurement across lithography consumables, CMP, specialty gases, wafers, and packaging materials. Regional suppliers certified now see disproportionate order flow. Which layer of this supply chain do you think is most mispriced right now - equipment, materials, substrates, or optical? 2026 is ramp setup, 2027 is the inflection, 2028–2030 is full adoption. Best investment can be made by knowing which player is essential for which phase. Repost this. Most people following the AI trade are one layer too shallow. Your followers will thank you.