GS 核心逻辑：Higher for longerGS 认为 2027-2028 年供应依然紧张，HBM 消耗 3-4 倍常规 DRAM 产能[1]H-Cap · 06-06"*In today's GS Weekend Macro call:* Giuni Lee, GS Korea • Memory stocks priced at mid-single digit P/E because market do…"。客户要求 3-5 年 LTA 并付预付款，GS 给出 9x PE 目标估值[1]H-Cap · 06-06"*In today's GS Weekend Macro call:* Giuni Lee, GS Korea • Memory stocks priced at mid-single digit P/E because market do…"。
JPM 数据支撑需求: JPM 预计 2026 年存储占 CSP 硬件 capex 达 52%，2027 年达 73%[2]Terminal Junkie · 06-06"Random Thoughts: 1. Broadcom numbers were okay-ish, but the call was still a bit disappointing: Management is still poin…"。AI CPU 驱动 DRAM 需求上修，eSSD 需求 2026 年将超 500EB[2]Terminal Junkie · 06-06"Random Thoughts: 1. Broadcom numbers were okay-ish, but the call was still a bit disappointing: Management is still poin…"。
反面声音：预算占比不可持续群内指出，CSP 将超 50% capex 花在存储上不可持续，认为“必定会出问题”[3]John · 06-06"Good piece. Point 3 should be negative to memory. My own estimate using MU’s last ER numbers is over 33% of hyperscalers…"。同时担忧 2028 年周期见顶风险[4]H-Cap · 06-06"does the market bother to price their P/E going from 6x to 9x? When you know the party could be over come 2028"。
算法创新对长期需求的冲击存疑: 关于 TurboVec 等算法创新压缩内存使用的讨论，有人认为这会降低长期终端价值[5]davii · 06-07"good reminder that a lot of things can change technologically/architecturally over 3-4 years (especially if you actually…"，但也有人指出历史上的算法创新往往不影响总体需求[6]techbro · 06-07"Lol, how often do we have to have algorithmic innovations that compress memory usage but have no effect on demand. Seems…"。
业绩电话会平淡: 管理层指引 FY2027 出货近 10GW，但未能强力推销 AI 网络故事[2]Terminal Junkie · 06-06"Random Thoughts: 1. Broadcom numbers were okay-ish, but the call was still a bit disappointing: Management is still poin…"。
MTK 威胁显现: 预计 FY2027 定制 ASIC 营收 $94.2B，但 MTK 势头强劲，被认为是当前更好的选择[2]Terminal Junkie · 06-06"Random Thoughts: 1. Broadcom numbers were okay-ish, but the call was still a bit disappointing: Management is still poin…"。
开源模型大爆发: 单周发布 25+ 知名开源模型，涵盖 LLM（如 NVIDIA Nemotron 3 Ultra 550B）、图像（Ideogram 4）、音频及视频等多模态[7]Maverick · 06-06"Before the week ends, let's acknowledge one of the most INSANE week ever for open AI, with 25+ notable open-weight drops…"。
前沿实验室估值受压: Chamath 预计主要前沿实验室的 run rate 营收增速将“有意义地下降”[8]AD · 06-06"Chamath the Famous Bubble Pimp expects growth rates of the major frontier labs to decline “meaningfully”"，引发如果增速下滑估值也将承压的讨论[9]AD · 06-07"Not freaking out, just stating what he said. And if the run rate does decline meaningfully, so do the valuations"。
对 Jensen 话语权的分歧: 有人认为 Jensen 只是卖芯片，不应过度评价 AI 走向[10]Maverick · 06-06"It amazes me that all of these smart techies are coming around to this conclusion now. Not tooting my own horn bc I'm no…"；但也有人反驳其过去几年预测准确率极高，信誉远超他人[11]Chris · 06-06"Jensen track record so far going back several years might be 99.9% correct spot on. So his credibility and foresight is …"。
前所未有的融资规模: David Solomon 指出 GOOG 融资规模可能达 900 亿美元，是投资者对这种规模需求的首次具体数据点[12]RI · 06-06"David Solomon on Odd Lots discussing GOOG equity issuance “We upsized it and the Green Shoe, my guess is given the way t…"。
更多股权发行预期: 由于未来 5-10 年资本计划庞大，预计更多公司会选择股权而非纯债务融资以控制杠杆风险[12]RI · 06-06"David Solomon on Odd Lots discussing GOOG equity issuance “We upsized it and the Green Shoe, my guess is given the way t…"。
🟦正将 NAND 从存储转为 AI 内存层次 (CM/GP/LC 三层产品) · 产品路线图/存储转型 Kioxia is trying to turn NAND from storage into part of the AI memory hierarchy... CM Series for KV cache, GP Series with XL-Flash for ultra-low latency, LC Series QLC for high-capacity storage
🟦LTA 需求看至 2029 年后, 目标覆盖 50% 业务 · LTA需求/业务覆盖 Kioxia is seeing LTA demand beyond 2029... wants LTAs to cover roughly 50% of the business over the next two years
🟦需求预测 FY2027 11mn GB, FY2028 17mn GB · 需求预测 AI CPU DRAM demand alone is forecast at 11mn GB / 17mn GB in FY2027/FY2028, or 19%/24% of total market demand, per JPM
🟥▲96GB 模块主流化可推动总 LPDDR 需求提升 10-20% (非利空) · 内存需求/SoCAMM2 If 96GB modules become mainstream and increase nearly 6x... total SoCAMM2 LPDDR demand can still move 10-20% higher versus prior forecast. This looks like NVIDIA optimizing Vera Rubin shipments under
🟥▼John 根据 MU 财报估算已超 33%, 认为不可持续 · DRAM支出占比/不可持续 My own estimate using MU's last ER numbers is over 33% of hyperscalers capex is on DRAM. It's not sustainable.
🟥▲Computex 上亚洲投资者确信其必将涨价 (尽管官方否认) · 涨价预期/涨价 Most of the Asian investors I spoke to at computex were quite convinced Murata et al will raise prices despite rhetoric.
🟥▲如何与 GOOG 募资 $80B 并存? (若 capex 即将见顶, 为何大举融资?) · 资本支出/融资信号 How to reconcile 2028 peak with the fact that Google is raising $80B in equity? They wouldn't be making that decision if they thought the capex would peak in the next 1-2 years.