Serenity @ aleabitoreddit, the White Haired Stock God The 4-layer transmission framework of AI industry chain According to the order of fund rotation: Tier 1: Computing power/GPU (fully priced) -Subject: NVDA, TSM, AVGO -Logic: The computing power foundation for AI training, the starting point for all requirements -Status: From 2023 to mid-2025, Fwd PE 17.6x is no longer expensive, but the market value elasticity of $5.4T is limited -Bottleneck location: CoWoS packaging, HBM production capacity Layer 2: Storage/HBM (in progress) -Subject: MAmeriaNDK, WDC, Samsung, SK Hynix -Logic: HBM is the "memory bottleneck" of GPU, and each H100/B200 needs to be paired with 6-8 HBMs -Status: The period from the second half of 2025 to the first half of 2026 is the main upward trend, and MU and others have already significantly outperformed -Bottleneck location: HBM3E production capacity, DRAM wafer supply Layer 3: Optical Communication/Interconnection (2026H2 Maximum Alpha) -Subject: MRVL (Switch Chip), AVGO (SerDes), NOK (Remote) AAOI、AXTI、LITE、COHR -Logic: Marvel CEO Computex 2026 makes it clear that 'the next bottleneck is not computing power or storage, but connectivity' -Reason: Physical limits of copper cables → Explosive demand for silicon optical/CPO → InfiniBand → Ethernet switching requires a large number of optical modules -Status: Institutions are in the initial stage of building positions, with low attention from retail investors, which is exactly what Serenity refers to as the "third layer of potential excess returns" Level 4: Electricity/Infrastructure (Long term but most certain) -Subject: BE, GEV, OKLO, UUUU, CCJ -Logic: The electricity consumption of AI data centers doubles every 2 years, and the combination of nuclear energy, natural gas, and energy storage is a hard demand -Status: Narrative stage, waiting for the actual power generation project to be implemented from 2027 to 2028 --- What does Serenity think of herself? I have roughly summarized the complete AI supply chain bottleneck ranking of Serenity: InP substrate (photonics)>optical transceiver/CPO packaging>HBM3E>advanced packaging (CoWoS)>power. The "225x return" case of his core is IntelliEPI (InP substrate supplier) - because Serenity identified InP as the "upstream bottleneck" of optical communication in 2024, and no one looked at it at that time. The CEO only publicly acknowledged in Q1 2026 that the shortage of InP is a bottleneck for AI infrastructure - market confirmation and stock price explosion. Key insights: -The bottleneck for 2024 is HBM/CoPoS (storage+encapsulation layer) -The bottleneck of 2025-2026 has shifted to the optical communication/interconnection layer (copper cable → fiber optic → silicon optical) -The bottleneck of 2026-2027 is moving towards the power/energy layer --- core conclusion Which is the most important in the AI framework? Answer: There is no fixed "most important", the bottleneck lies in dynamic transfer - who is stuck and who is the most expensive Where is the biggest excess return now? Answer: Optical communication/interconnect layer (MRVL/NOK/AOI/photonics), the third layer organization is just beginning to build warehouses Can storage still be purchased? Answer: Watch and see - MU and others have been fully priced and need to wait for the next catalyst (HBM4, cycle recovery confirmation) Is it the most certain in the long run? Answer: Electricity (nuclear energy+energy storage) -2027-2028 is just the peak period for performance, and the layout is too early but worth tracking One sentence: First, identify the super trend (AI infrastructure), break down the physical supply chain (4 layers), find the bottleneck where the current funds have not been fully priced (internet/optical communication), and stand in the wheel position in advance