AI需求神話被戳破！ 分析師點名「3大瓶頸黑馬」逆勢衝新高

市場持續看好超大型雲端業者擴大AI資本支出，但資深分析師Mr. P對所謂「10年需求能見度」提出質疑。他認為，超大型雲端業者甚至難以準確預測2年後的需求，更難掌握長達10年的變化，AI基礎設施支出最終仍可能受到周期性影響。

長期追蹤AI基礎設施建設的匿名分析師、半導體科研機構P Equity Research背後的操刀人Mr. P近日以嘉賓身分出席Logan Jastremski的播客節目。Mr. P在節目中指出，短期來看，AI推理需求仍推動HBM、DRAM與NAND等記憶體產品需求。他估計，明年超大型雲端業者資本支出可能達到1.1兆至1.2兆美元，其中記憶體支出占50%至60%，約5000億至7000億美元；瑞銀甚至估計，明年超大型雲端業者單在記憶體上的支出可能達到9000億美元。

市場也出現長期供應協議（LTA）需求，包括三星已開始收到長達10年的合約要求。不過，Mr. P引述曾任職三星與超微的業界人士說法指出，這類長約的實際約束力可能被過度解讀；若需求突然下滑，部分合約可能在協商下取消，過多庫存也可能影響後續營收。

AI算力供應的限制也不再只是GPU數量。Mr. P指出，H100二手價格仍可達2.5萬美元，B系列GPU每小時租金也由約5美元升至7至8美元，部分晶片使用年限甚至可能達到10年。

但從AI基礎設施的物料清單來看，真正吃緊的環節已延伸至電力、高階封裝、記憶體與ABF載板。其中ABF載板產能相當緊張，供應短缺可能持續至2028年甚至2030年以後；戴爾、惠普、輝達與博通在近期的財報中也將DRAM、NAND、ABF載板及晶圓列為產能限制因素。

能源供應同樣形成限制。三菱、西門子及GE Vernova的燃氣輪機訂單已排到2030年以後，若目前下單，可能要到2030年才能取得設備。

資料中心網路連接方面，Mr. P認為銅纜與光通訊未來數年仍將並存。2027年以前，雲端業者仍會盡可能延長銅纜使用周期。

至於CPO（共同封裝光學），目前仍面臨成本、良率與散熱等問題，NPO（近封裝光學）可能在2027年率先出現較多應用，CPO則可能在2028年至2029年開始擴大產能，但滲透率仍將維持較低水準。Mr. P預期，CPO真正取得主導地位可能要等到2030年以後。

從「10年需求能見度」到記憶體、ABF載板、電力與封裝等限制，相關觀察顯示，AI基礎設施擴張的關鍵已不只是GPU供應，而是整體供應鏈能否同步跟上需求。

◎免責聲明：本文的個股、ETF與專家觀點，非任何投資建議與參考，投資人應獨立判斷，審慎評估風險，自負盈虧。

Dropping a podcast with @pequityresearch Mr. P has been doing some great work breaking down the AI buildout and following where hyperscaler capex actually goes. In this episode I wanted to walk through the full stack with logic, memory, power, and networking. We get into why memory could become the largest line item in the AI bill, what old GPU rental prices tell us about compute demand, and where value is going to accrue as the physical constraints get harder to solve. At the center of the conversation is his view that AI spending cannot grow forever. Long-term contracts might change the shape of the next memory downturn, but they don’t eliminate the cycle. And signing a 10-year contract does not mean anyone can actually see 10 years of demand. We also get into why he’s excited about optics, where NAND and HBF fit as agents use more memory, and his views on Chinese open source and the future of US model development. We discuss: - Why he thinks 10-year demand visibility is bullshit - Memory’s growing share of hyperscaler capex, and why estimates vary so much - Why older GPUs are still renting and what that says about compute demand - How long-term agreements, pricing floors, and prepayments actually work - Power as a bottleneck, and why identifying a constraint isn’t the same as finding an investment - Copper vs optics, and where networking value accrues - Agents, NAND, and where HBF fits in the memory hierarchy - CXMT, Chinese open source, and the risks of slowing frontier model development Timestamps: 0:00 – Why AI Spending Can’t Grow Forever 1:13 – P Equity Research’s Background 5:48 – Where Hyperscaler Capex Actually Goes 8:00 – Is Compute Still Tight? 12:08 – Memory’s Share of the AI Bill 17:20 – Why the Memory Cycle Isn’t Dead 18:37 – Inside a Long-Term Agreement 28:30 – What Happens If Customers Cancel? 32:17 – The Rising Cost of the AI Buildout 33:43 – Copper vs Optics 38:49 – Power and Gas Turbines 39:29 – US Models and Chinese Open Source 42:36 – Agents and NAND 47:00 – Where HBF Fits 52:23 – What P Is Most Excited About 56:27 – CXMT and China’s Memory Industry 1:04:00 – Closing Thoughts Enjoy!