I predicted three months ago - Super Micro would hit another purchasing delay in 2Q25, and that's exactly what happened.Super Micro: Shortfall stem from specification changes from a major new customer that delay revenue recognition.Transcript：https://t.co/Smp7fMXTQB https://t.co/ULN2VlFQl7郭明錤 (Ming-Chi Kuo) (@mingchikuo):Let’s take a quick look at Super Micro’s (SMCI) preliminary results. Long story short: this quarter’s numbers are likely to stay under pressure.AI servers currently split into three main ty ...

Super Micro's Performance & Forecast - Super Micro is expected to face deferred procurement again in Q2 2025, aligning with previous predictions [1] - Revenue shortfall stems from specification changes from a major new customer, delaying revenue recognition [1] - The current quarter's financial report may continue to be under pressure [1] AI Server Market Dynamics - There are three main types of AI servers: ASIC (e g, Google TPU servers), Nvidia GB200/300 NVL72, and lower-end Nvidia servers (e g, x86 HGX series) [1] - The primary growth drivers in the AI server industry are ASIC and NVL72, with demand originating from CSPs (Cloud Service Providers) [1]

Unlikely to be iPhone assembly — more likely a subassembly for a small-volume product, or assembly for niche products or Apple’s in-house servers.Source:https://t.co/61ZcNACuRP https://t.co/tn8fTPpRu1 ...

不大像是iPhone組裝，可能是某規模較小產品之半成品生產，或是利基產品或Apple自用的伺服器之組裝。來源：https://t.co/61ZcNACuRP https://t.co/YaK6uXqTV4 ...

Trump說台積電要投資美國3,000億美元1. 台積電官方說的投資金額是1,650億美元，Trump先前說的金額是2,000億美元。Trump進一步將投資金額提高到3,000億美元，像是為即將到來的半導體關稅暖身，符合Trump自己說過的「先提出一個極高的數字，再往下談」的做法。2. 近期市場傳言美台政府談判關稅時，Trump/美方要求台積電入股Intel 49%或追加投資4,000億美元。然而，我的理解是，美台灣關稅談判並沒有提到台積電。Trump單方面提到台積電要投資美國300億美元的發言，也算驗證了我的觀點 - 如果有需要，美國政府可以單方面發言或直接找台積電。3. 如果上述市場傳言為真，台積電總共要投資美國5,650億美元，遠高於Trump說的3,000億美元，代表市場傳言非事實。4. 美國政府部分人士先前的確將台積電入股，作為振興Intel的多種潛在方案之一，但隨著美國政府越來越理解Intel實際上面臨的挑戰，以及現任CEO陳立武上台後，目前這樣的聲音已經顯著減少了。來源：https://t.co/FF2GfGIe2V ...

Trump says TSMC will invest $300 billion in the US1. TSMC's officially announced investment is $165 billion, while Trump previously stated $200 billion. His further escalation to $300 billion seems to be laying the groundwork for upcoming semiconductor tariffs, consistent with his stated approach of "starting with an extremely high number, then negotiating down."2. Recent market rumors suggest that during US-Taiwan tariff negotiations, Trump/the US demanded TSMC take a 49% stake in Intel or invest an additi ...

Investment & Financial Analysis - TSMC's official investment in the US is $165 billion, while Trump initially mentioned $200 billion, and later increased it to $300 billion [1] - Market rumors suggested that during tariff negotiations, the US requested TSMC to invest an additional $400 billion or acquire 49% of Intel [2] - If the market rumors were true, TSMC's total investment in the US would reach $565 billion, significantly higher than Trump's $300 billion, indicating the rumors are likely false [3] Government & Policy Impact - Trump's increased investment figure from TSMC could be a prelude to upcoming semiconductor tariffs, aligning with his strategy of "proposing a very high number first, then negotiating down" [1] - The US government might unilaterally make statements or directly engage with TSMC if needed, regardless of formal negotiations [2] - Some within the US government previously considered TSMC's investment as a potential solution to revitalize Intel, but this idea has diminished with a better understanding of Intel's challenges and the appointment of its new CEO [4]

Technology & Production Challenges - CoWoP faces significant uncertainties and challenges in mass production and commercialization [1] - Implementing SLP in CoWoP is far more challenging than Apple's use case, involving approximately 10,000 times the system power, less than half the line width and spacing, more than three times the layer count, and 100 times the area [2] - Expectation that CoWoP will reach mass production and be deployed in Rubin Ultra by 2028 seems overly optimistic without concrete test data [3] Strategic & Competitive Landscape - Apple invested in SLP R&D as early as 2013, with mass production starting in 2017, requiring four years of collaboration across the supply chain [1] - Nvidia may not possess stronger control over technology and the supply chain than Apple did during SLP adoption [2] - TSMC is developing CoPoS, targeting mass production around the same time as CoWoP (post-2028), potentially prioritizing it due to its focus on manufacturing efficiency [3] - Introducing two major, unproven technologies (CoWoP and CoPoS) within the same year carries considerable risk, posing a challenge to CoWoP's 2028 mass production goal [3]

CoWoP是近期AI伺服器產業的焦點，這是好技術並值得持續關注，但也不能忽略量產/商業化的高度不確定性與挑戰。網路上已經有很多關於技術優勢與製造挑戰的分析，我從另外兩個角度來分析:1. 首先是用Apple的例子來對比根據臻鼎的年報，可推論Apple至少從2013年就開始投入SLP研發，至2017年才開始量產並用於新款iPhone (X、8與8 Plus)。這四年內，Apple、材料商、製造商、與設備商合作，共同解決研發與量產問題，這不只是單一技術開發，而是整個產業生態升級。現在的PCB產業技術當然遠勝十年前，但Nvidia對技術與供應鏈的掌控能力不見得勝過10年前那時的Apple，且CoWoP要導入SLP的挑戰也遠勝iPhone案例 (粗略看，前者是後者約萬倍的系統功耗、一半以下的線寬線距、3倍以上的層數、百倍的面積)。在沒有具體的實際測試結果前，CoWoP要在2028年量產並用於Rubin Ultra是很樂觀的預期。2. CoWoP與CoPoS同時量產與商業化的挑戰艱鉅台積電有另一個次世代封裝技術CoPoS，也預計在2028後量產。CoWoP在理論上可以改善傳輸效率並簡化供應鏈，但CoPoS要解決的是很實際 ...

Strategic Partnership & Manufacturing Advantage - Tesla gains real-world foundry experience at an exceptionally low cost, enhancing chip design capabilities and manufacturing knowledge [1] - Acquiring core manufacturing expertise becomes a strategic advantage for Elon Musk's businesses due to increasing demand for advanced chips [1] - Samsung's new Texas fab will be dedicated to making Tesla's next-generation AI6 chip, highlighting its strategic importance [4] - Partnership presents manageable downside and strong upside potential for both Tesla and Samsung [3] AI Chip Development & Production - Tesla's AI6 chip is scheduled for mass production in 2027 using Samsung's 2nm node (SF2) [2] - Samsung's SF2 yield is currently 40-45%, lower than TSMC's N2 (over 70%) and Intel's 18A (50-55%) [2] - TSMC will make AI5, which just finished design, initially in Taiwan and then Arizona [5] Risk Mitigation & Alternative Scenarios - If AI6 production falls short of expectations, Tesla could shift the order back to TSMC, absorbing resulting delays [3] - Tesla's edge in real-world AI could significantly reduce the risk of AI6 delays [3] Competitive Landscape - Chip design and manufacturing could become a core competitive advantage across Elon Musk's businesses if AI6 reaches mass production smoothly [4] - Samsung may not fully catch up with TSMC in advanced nodes but has discovered a new business model involving customers in the manufacturing process [4]