PCB/Substrate 瓶颈 ⟪LINK21⟫
PCB/Substrate 是当前最清晰的 AI 基础设施瓶颈交易
- Rubin 架构推升价值量: Rubin 架构是 PCB 交易的“唤醒时刻”：单机架 PCB 价值量 GB300 → Rubin 提升 233% ($35k → $117k)[1]Terminal Junkie · 05-23"…CB is the biggest downstream content increase in VR200, up *+233%* vs GB300. PCB content per rack rises from *$35k to $117k*,…"，ABF 载板提升 82% ($11k → $20k)[1]Terminal Junkie · 05-23"…es from *$35k to $117k*, while ABF substrate content rises *+82%*, from *$11k to $20k*. Honestly, why are we not buying more…"。
- ABF 载板长期短缺: 市场低估了 ABF 载板的紧缺程度，预计到 2030 年缺口扩大至 22%[1]Terminal Junkie · 05-23"…here. MS now expects the ABF supply/demand gap to widen to *22% undersupply by 2030*, versus *15% previously*. Personally, I think even…"。本轮周期由 AI 驱动，定价权强，Unimicron (首选) 和 NYPCB (现货定价杠杆) 将显著受益[1]Terminal Junkie · 05-23"*Random Thoughts:* 1. *PCB/Substrate is still one of the cleanest AI infra bottleneck trades:* For those who have been f…"。
- 上游材料同样紧缺: T-glass (Nittobo) 和高频铜箔 (Co-Tech) 产能接近满载。Kingboard 因具备玻璃纤维、铜箔和环氧树脂的垂直整合能力，在抢占产能分配上具有优势[1]Terminal Junkie · 05-23"*Random Thoughts:* 1. *PCB/Substrate is still one of the cleanest AI infra bottleneck trades:* For those who have been f…"。
- 设备端二阶受益: 层数增加 + 良率要求提升 → 利好 Ta Liang, Han’s CNC 和 Dtech 等设备商，它们可能是更干净的二阶受益者[1]Terminal Junkie · 05-23"*Random Thoughts:* 1. *PCB/Substrate is still one of the cleanest AI infra bottleneck trades:* For those who have been f…"。
仓位轮动 ⟪LINK22⟫
半导体多头拥挤度创历史新高 —— 资金持续从软件抽血
- 半导体权重创新高: 对冲基金多头组合中半导体权重创历史新高，而软件权重降至 2019 年来最低[2]H-Cap · 05-23"Hedge funds and mutual funds have continued their recent portfolio rotations away from Software and toward Semis. The we…"。公募基金在软件 (除 MSFT 外) 的低配幅度达到 2012 年来最大[2]H-Cap · 05-23"Hedge funds and mutual funds have continued their recent portfolio rotations away from Software and toward Semis. The we…"。
- 共同重仓股: BA, MA, MRVL 和 V 成为对冲基金和公募基金本季度的“共同最爱”，今年以来跑赢等权 S&P 500 指数 3 个百分点[2]H-Cap · 05-23"…rned 10% YTD, outperforming the equal-weight S&P 500 by 3 pp."。
Cerebras 速度测试 ⟪LINK23⟫
Cerebras 运行 Kimi 速度惊人，但延迟指标被指伪命题
- 速度碾压 GPU: Cerebras 运行 Kimi K2.6 达到 981 tokens/s，比第二快的 GPU 云提供商快 6.7 倍[3]Arun Dhar · 05-23"…clocked in at 981 output tokens per second, making Cerebras 6.7 times faster than the next-fastest GPU-based cloud provider and 2…"。
- Agentic 延迟是伪命题: 群友指出，在 Agentic 场景下，单纯追求延迟 (latency) 是一个 red herring (转移注意力的伪命题)[4]JP · 05-23"imo, agentic latency is a red herring"。