Ming-Chi Kuo

Nvidia LPU/LPX racks set for order-of-magnitude growth: application trends, ecosystem integration, and a new PCB cycle 1. My latest supply-chain checks suggest that following Nvidia’s investment in Groq, LPU shipment forecasts have moved significantly higher. Total shipments in 2026–2027 are estimated at 4–5 million units (with 30–40% in 2026 and 60–70% in 2027). Relative to historical annual volumes, this implies order-of-magnitude growth, roughly 10×+. 2. Fast-growing LPU demand reflects two key factors: (1) Deep integration with Nvidia’s ecosystem (e.g., CUDA), significantly lowering barriers to application development and deployment. (2…

Nvidia LPU/LPX機櫃將迎來數量級成長：應用趨勢、生態整合與PCB新成長週期 1. 我的最新產業調查顯示，Nvidia投資Groq後，LPU出貨規劃大幅上修。2026-2027年LPU共出貨預估約400–500萬顆（2026約30–40%、2027約60–70%），相較過去年度出貨量，將出現10倍以上的數量級成長。 2. LPU需求快速成長主要來自兩項因素： (1) 與Nvidia生態系 (如CUDA) 高度整合，大幅降低應用開發與部署門檻。 (2) 超低延遲推論需求快速增加，包括AI agents（如coding agents）以及正在興起的real-time、consumer-facing與physical-AI等類型應用。 3. 為維持推論decode階段的超低延遲優勢，並因應長文本推理帶動的KV cache需求快速成長，Nvidia預計將每機櫃LPU數量由目前64顆提升至256顆，以擴大記憶體容量並維持超低延遲效能。新架構機櫃預計於4Q26–1Q27量產，2026與2027年機櫃出貨量分別約300–500與15,000–20,000個。 4. Nvidia生態整合LPU的三個關鍵觀察重點： (1) 網路架構：NVLink Fusion與RealScale的機櫃互聯。 (2) 開發者介面：Nvidia NIM是否讓開發者在部署時無需區分GPU與LPU。 (3) 編譯整合：TensorRT-LLM是否支援LPU的compile-first架構。 5. LP…

Supply chain check update: Nvidia and WUS Printed Circuit have begun testing next-gen CCL material M10, which could trigger the next upgrade cycle in PCB materials for future AI servers. Key points: 1. The program suggests WUS has an early lead in PCB development for Nvidia’s next-gen rack Kyber and the new Rubin Ultra / Feynman platforms. This could support the company’s future growth momentum. 2. Sampling began in 1Q26, with preliminary test results expected in 2Q26. 3. M10's target applications include orthogonal backplanes (midplanes) designed to replace the current cartridge architecture, as well as switch blade motherboards for the R…

供應鏈調查更新：Nvidia與滬電股份已開始測試次世代CCL材料M10，將帶動未來AI伺服器的PCB材料新一輪升級週期。幾個重點： 1. 此測試計畫意味著滬電股份在Nvidia次世代機櫃Kyber與新平台 Rubin Ultra / Feynman的PCB之開發上有領先優勢，此優勢有助於該公司未來的營運動能成長。 2. 1Q26已開始送樣與打樣，預計2Q26將取得初步測試結果。 3. M10的主要應用包括用來取代Cartridge架構的正交背板 (midplane / orthogonal backplane)，以及Rubin Ultra / Feynman平台的Switch blade主板。 4. 與M9僅台光電通過認證不同，目前M10測試共三家CCL供應商。除台光電外，另新增一家中國廠與一家台系廠，未來有助於提升Nvidia的CCL供應鏈管理彈性。 5. 若測試順利，M10 CCL與PCB預計將於2H27開始量產。 6. 從量產性與商業化考量，M10目前使用的石英布 (Quartz Cloth)，未來也可能改為Low Dk-2。 7. 滬電股份同時也是Nvidia用於AI推理的超低延遲LPU/LPX機櫃之52層PCB主要供應商 (預計在4Q26–1Q27量產)。在同時具備高層數PCB製造能力與先進材料開發能力之優勢下，該公司可望受益於Nvidia未來AI伺服器架構升級。

Apple the only bright spot among notebook brands in 2026: Neo, Neo 2, OLED roadmap & strategy

The processor, display size, and casing color options of MacBook Neo are almost exactly in line with what I predicted about nine months ago (https://x.com/mingchikuo/status/1939558413459177897 Neo entered small-volume production at the end of Dec. 2025 (vs. my earlier expectation of Oct.). Because mass production is about three months later than previously expected, 2026 shipments are slightly lower than prior estimates, around 4.5–5 million units (with about 2–2.5 million in 1H26). For a single laptop model, that’s still a very strong number. Neo 2 was originally expected to feature a touch panel to compete with Chromebooks (50%+ of which …

Apple是2026筆電品牌中唯一亮點：Neo展望、Neo 2、OLED機型與Apple策略

MacBook Neo的處理器、螢幕尺寸與顏色，與我約9個月前的預測幾乎一致 (https://x.com/mingchikuo/status/1939558393875943719 Neo自2025年12月底開始小量生產 (原先預期為10月)。因量產時程延後約3個月，2026年出貨量略低於先前預估，約450–500萬台 (其中1H26約200–250萬台)。以單一筆電機型來看，這樣的出貨量仍然相當強勁。 原先預期Neo 2會配備觸控面板以與Chromebook (50%+配備觸控) 競爭，但我的最新產業調查顯示Neo 2可能不採用。 目前廣達為Neo的獨家組裝廠。鴻海近期可能成為第二供應商。需注意的是，過去1-2年Windows筆電組裝業務成長快速、目標是成為全球最大筆電組裝廠的立訊，正積極爭取Neo 2的NPI。 Neo出貨動能預期將逐季穩定成長，原因有二：1) Back-to-School與年末Holiday旺季需求、2) 競爭對手去年囤積的較低成本記憶體庫存預計在1H26中期耗盡，因此2Q26起更多筆電機型可能調漲售價以反映記憶體成本上升，不利與Neo競爭。 重申我先前預測：MacBook Pro將在4Q26末或1Q27初採用觸控OLED面板之重大升級。 MacBook Air預期將在2028或2029年配備OLED面板。 預估2026年MacBook出貨量約2,500萬台 (20–25% YoY)，有機會重返新冠疫情時期的高峰，並成為2026年筆電品牌中的…

Apple’s Q1 2026 earnings call provided a number of insights. Below are the key takeaways from a memory sector perspective: Earnings call transcript: https://t.co/R2cjXyjBym 1. Apple highlighted two major uncertainties in its hardware supply chain: advanced nodes and memory. The former is defined as a "supply constraint," while the latter is viewed as a "cost pressure." 2. Apple’s comments aligns with my pre-earnings analysis (see quotation below): memory pricing has increased significantly, with supply guaranteed. 3. For AI memory investors, Apple’s comments are relatively less relevant. In contrast, non-AI memory investors should pay clo…

Apple上週法說 (Q1 2026) 資訊量豐富，以下從記憶體角度整理重點： 法說會議記錄： https://t.co/R2cjXyjBym 1. Apple硬體供應鏈的兩大不確定性為先進製程與記憶體；前者屬「供應限制」 （supply constraint），後者為「成本壓力」 （cost pressure）。 2. Apple的說法與我在法說前的分析一致 (見下面引用)：記憶體報價顯著上漲，但能得到供應保障。 3. AI記憶體投資人可相對忽略Apple的觀點，非AI記憶體投資人則需關注，因長期來看，非AI記憶體市場達到供需平衡的效率應該高於AI記憶體市場。 4. 若Apple在上週法說中定義記憶體是「供應限制」，而非「成本壓力」，應有利於非AI記憶體股，這是部分短線/樂觀投資人想要看到的結果。很巧的是，Apple法說後不同市場的非AI記憶體股價同時出現修正，而在近期非AI記憶體股價修正相關討論中，較少看到從Apple法說/需求面來解讀的觀點。 5. 考量到Apple具備絕強的議價力，且iPhone的DRAM與NAND Flash用量均約佔手機記憶體市場的20–25%，故較不適合單憑Apple的觀點解讀非AI／手機記憶體需求。相較下，Qualcomm與聯發科對手機市場的展望，更具中長期非AI／手機記憶體需求趨勢之參考價值。 6. 我認為非AI記憶體的股價，可能還沒充分反映今年的強勁獲利動能。不過，展望2027-2028，AI記憶體的需求應該還是有較好的能見度。

Wistron, a leading AI server assembler, reported its FY25 financial results. Gross margin (GM) for FY25 and 4Q25 came in at 6.1% and 5.6%, respectively, both below street consensus of 6.5% and 6.8%. 1. These results validate my earlier view on the structural impact of AI server design trends on assembler profitability. GM was also pressured by the product transition (GB200 NVL72 → GB300 NVL72). 2. 2026 assembly GM positives to watch: 1) ~200% YoY shipment growth for the GB NVL72 series in 1H26, where scale should be supportive; and 2) Mass production of new ASIC AI servers with higher assembly GM starting in 2H26. 3. On the downside for 20…

AI伺服器組裝大廠緯創公布2025年財務業績，關鍵之一的毛利率在2025年與4Q25分別為6.1%與5.6%，均低於市場共識的約6.5%與6.8%。 1. 緯創的財務業績除驗證我先前提到AI伺服器設計趨勢對組裝廠獲利之結構性影響，另一原因是產品轉換 (GB200 NVL72 ⭢ GB300 NVL72) 對組裝廠的毛利率也有壓力。 2. 2026年組裝廠的毛利率利多可觀察：1) GB NVL72系列在1H26出貨顯著成長約200% YoY，規模擴大應有幫助、與2) 組裝毛利率較高的新款ASIC AI伺服器在2H26開始量產。 3. 2026年組裝廠的毛利率挑戰方面，除上述提到Nvidia AI伺服器設計趨勢外，還有零組件漲價會稀釋帳面毛利率。 4. 緯創與其最大AI伺服器客戶Dell在去年9月股價一同起漲，反應法人調升緯創的市占率與毛利率 (2025年的5.5-6% ⭢ 6%+)。緯創股價雖受益於重新評價 (rerating) 故近期仍在高檔，但需關注Dell股價在顯著成長的積壓訂單 （Backlog） 利多出盡後，因市場對其獲利能力/毛利率擔憂，股價已回到去年9月起漲點。

Recent market chatter suggests HP may raise its 2026 notebook shipment outlook. My latest supply-chain surveys indicate this is not the case. HP has only verbally asked suppliers to extend lead times from about one month to six months, while full-year shipment volume remains largely unchanged. Since there’s no upfront payment, suppliers are unlikely to truly build inventory at that level—though some incremental stocking may occur, creating the appearance of order growth. Two key takeaways: 1. A notable increase in HP notebook shipments is unlikely with no meaningful improvement in memory supply or end-market demand. This looks more like a mo…

近期市場傳言HP要調升2026年筆電出貨預估。我最新的產業調查顯示並非如此，HP僅是口頭上要求供應鏈提前備貨時間從約1個月增加到6個月，但全年出貨量無太大變化。對供應鏈來說，因為HP沒有付錢，所以不大可能真的這樣備貨，但多少會因為HP的要求多備一些，故有加單的假象。 兩個重點： 1. 因記憶體供應與市場需求無顯著改善，故不預期HP筆電出貨量會明顯成長。HP此舉較像是因應未來數個零組件價格可能上漲故提前備貨。 2. 如果其他品牌也跟進HP提前備貨，那筆電零組件短期內加單的假象會更顯著。

TSMC announced 2026 capex of US$52–56bn at today’s earnings call, with strong Nvidia demand as the key driver. 1. Ahead of the call, sell-side analysts and media generally expected 2026 capex of US$45–52bn, while the buy side was already at US$53–56bn. 2. TSMC is typically conservative on capex guidance, so final 2026 spending is very likely to beat buy-side pre-call expectations. 3. Most customers negotiate for guaranteed capacity. Nvidia CEO Jensen Huang negotiates for land. In November last year, he visited Tainan and told TSMC he was willing to pay to secure the reserved P10 and P11 land next to Fab 18. At the time, Fab 18 plans were o…

台積電在今天的法說公布2026年資本支出為520-560億美元，Nvidia的強勁需求是推升關鍵。 1. 分析師與媒體在法說前對台積電的2026年資本支出共識普遍為450-520億，不過法人共識早已調高至530-560億。 2. 台積電向來對宣布資本支出略微保守，意味著2026年最終資本支出結果極有可能擊敗法人在法說前的預期。 3. 一般客戶跟台積電談的是包產能，Nvidia CEO黃仁勳跟台積電談的是包地。去年11月黃仁勳到台南對台積電提出「願意掏錢包下Fab 18旁的P10與P11預定用地」之要求。當時台積電Fab 18較明確的廠區規劃只有P1-P9，黃仁勳的目標是爭取包下P12用地。 4. 在黃仁勳提出包下P10與P11用地前，法人對台積電2026年的資本支出共識約450-500億美元。 5. 先進製程與封裝產能緊張已是廣為人知，但台積電仍嚴格審視客戶提出的需求，故常常提供低於客戶「樂觀版本」的產能。如果客戶真的想要台積電提供最樂觀版本的產能也不是不行，像黃仁勳一樣，掏更多錢出來承諾包地養產能就有機會。不過，在各方資源緊張下，即便有其他客戶仿效黃仁勳，要得到最樂觀版本產能的難度也越來越高。

A quick take on the Apple–Google AI partnership I was on the MM Podcast about a month and a half ago, where we talked about the Apple–Google AI partnership. Here are the main takeaways from that part of the conversation (from ~38:50 to 45:20). I’ve lightly edited the text below to make it easier to read: 1. Apple is facing two short-term challenges in its in-house AI development that are effectively pushing it toward partnering with Google: (1) this year’s WWDC AI showcase can’t afford another disappointment, and (2) as cloud AI continues to advance, user expectations have risen to a point where even fully delivering on past promises around…

重點快速分析Apple與Google合作AI 我在一個半月前接受MM Podcast放談時，當中提到關於Apple與Google合作AI的重點如下 (約自38:50開始到45:20，以下內容為方便閱讀，有做部分文字潤飾)： 1. Apple發展自有AI的2個短期挑戰，迫使Apple跟Google合作： (1) 今年WWDC的AI展示不允許再失敗。 (2) 用戶對AI服務的品質標準隨雲端AI發展提升，就算做到過去對Apple Intelligence與Siri的承諾，現在看來也不夠，故需要更強的裝置AI模型。 2. 雖不預期裝置AI在短期內能推升硬體出貨量，但此合作可緩解Apple短期來自各方的壓力。不過，AI成為硬體、作業系統與用戶體驗之核心是遲早的事，故Apple在長期仍需面對提升對AI核心技術的掌控能力之挑戰。 3. Tesla的FSD就是AI成為裝置與用戶體驗核心的好例子。如果Tesla的FSD算法來自其他廠商，則Tesla的競爭優勢與本益比評價肯定會比現在低。 4. Apple自研的AI伺服器晶片預計在2H26量產，自有資料中心預計在2027年開始建置並運營，這可能代表Apple的裝置AI需求預期在2027年開始會有較顯著成長。

Regarding recent discussions among CCL-focused investors on whether the key CCL spec for VR200 NVL72 could be downgraded from M9, below are my latest surveys and views: 1. Nvidia previously began testing 896K2 and 892K2 and prototyping PCBs, which likely gave rise to the market rumors. Test results are expected by the end of 1Q26. 2.Nvidia’s current mass-production target remains 896K3. While this has not been finalized, the uncertainty alone is sufficient to impact stock prices that have already priced in this positive expectation. 3. Discussing CPX based on Nvidia’s CES 2026 keynote is somewhat odd, as Jensen didn't mention CPX. CPX curren…

關於最近關注CCL的投資人討論的VR200 NVL72關鍵CCL規格會否自M9降規，我的最新調查與看法： 1. 先前Nvidia開始測試896K2、892K2並打樣PCB，這可能是市場傳言起源。預計1Q26底前會有測試結果。 2. 目前Nvidia的量產目標依舊是採用896K3。此尚未定案，但此種不確定性足以對已反映此利多的股價造成影響。 3. 從Nvidia的2026 CES演講內容討論CPX是很奇怪的事，因為黃仁勳並沒有提CPX。CPX的能見度較低，因量產可能延期，但目前目標依舊是採用896K3。 4. Nvidia AI伺服器持續挑戰人類技術最前沿，有各種測試很正常，包括先前市場傳言的泰福與台虹的PTFE。 5. 此事件不影響PCB材料規格持續升級的長期趨勢。 6. 短期股價波動來自Nvidia對現實與理想的取捨與妥協，未來這樣的案例可能還會持續發生。

Key Updates on NVIDIA Vera Rubin/VR200 NVL72 (Integrating my latest supply chain checks and NVIDIA CEO Jensen Huang’s CES 2026 keynote) 1. NVIDIA has renamed the AI server VR200 NVL144 (die-based) to VR200 NVL72 (package-based). My supply chain checks indicate VR200 NVL72 will be offered in two power profiles: Max Q and Max P. ➢ Jensen referenced NVL72, rather than the previously used NVL144 naming. ➢ Max Q and Max P share the same hardware design. ➢ Max Q GPU/rack power (TGP/TDP): ~1.8/190 (kW). ➢ Max P GPU/rack power (TGP/TDP): ~2.3/230 (kW). ➢ Both are meaningfully higher than GB300 NVL72 at 1.4/140 (kW). ➢ Offering two power profiles imp…

Nvidia Vera Rubin/VR200 NVL72重點更新，整合我的最新產業調查與Nvidia CEO黃仁勳的2026 CES演講 1. Nvidia將AI伺服器VR200 NVL144 (die-based) 改名為VR200 NVL72 (package-based)。我的調查指出，VR200 NVL72將分為Max Q與Max P兩種power profile。 ➢ 黃仁勳提到了NVL72，而非先前的NVL144 ➢ Max Q與Max P硬體設計相同 ➢ Max Q GPU/機櫃功耗 (TGP/TDP)：約1.8/190 (kW) ➢ Max P GPU/機櫃功耗 (TGP/TDP)：約2.3/230 (kW) ➢ 兩者均顯著高於GB300 NVL72的1.4/140 (kW) ➢ 因應資料中心供電規格，不同power profile提升安裝彈性。此改變也意味著Nvidia開始注意到物理世界對建置AI伺服器的限制，並將其反映在產品規格上 2. VR200 NVL72 Max Q與Max P的GPU散熱設計升級。 ➢ 兩者均採用微通道冷板 (micro-channel cold plate；MCCP) 搭配鍍金散熱蓋 (gold-plated lid) ➢ 市場高度期待GPU TGP達2.3kW後開始採用微通道蓋板 (micro-channel lid；MCL)，但MCL最快需至2H27才會量產 3. 受益於GPU、HBM與NVLink Switch規格升級，…

Intel預計最快在2027年開始出貨Apple最低階M處理器 市場一直都有Intel成為Apple先進製程供應商的傳聞，但此傳聞始終欠缺能見度。不過，我最新的產業調查顯示，Intel成為Apple先進製程供應商的能見度近期顯著改善。 Apple先前與Intel簽署NDA並取得先進製程18AP的PDK 0.9.1GA，目前關鍵模擬與研究項目 (如：PPA等) 進展符合預期，並正等待Intel預計在1Q26釋出的PDK 1.0/1.1。Apple的規劃是Intel最快在2Q-3Q27開始出貨採用18AP先進製程的最低階M處理器，但實際狀況須視取得PDK 1.0/1.1後的開發進展而定。 目前Apple的最低階M晶片主要用於MacBook Air與iPad Pro，兩者在2025年的出貨量共計約2,000萬部。因MacBook Air的2026年出貨可能會被新款配備iPhone處理器的低價版MacBook所影響，故預期2026與2027年的最低階M處理器出貨量均在1,500-2,000萬顆。 最低階M處理器的訂單數量少，故對台積電的基本面與未來數年的領先優勢完全沒任何影響。不過，此訂單對Apple與Intel的宣示與趨勢意義重大。 1. 對Apple而言，除高度支持川普政府極力宣揚的「美國製造」政策外，雖可見未來仍需高度依賴台積電的先進製程，但仍須找到第二供應商以符合供應鏈管理所需。 2. 對Intel而言，取得Apple先進製程訂單的意義，遠遠高於此訂單對營收與利潤的實際貢…

Intel expected to begin shipping Apple’s lowest-end M processor as early as 2027 There have long been market rumors that Intel could become an advanced-node foundry supplier to Apple, but visibility around this had remained low. My latest industry surveys, however, indicate that visibility on Intel becoming an advanced-node supplier to Apple has recently improved significantly. Apple previously signed an NDA with Intel and obtained the advanced-node 18AP PDK 0.9.1GA. The key simulation and research projects (such as PPA) are tracking in line with expectations, and Apple is now waiting for Intel to release PDK 1.0/1.1, currently scheduled fo…

After Nvidia reported its Q3 FY26 results, there was a view that its Days Sales Outstanding (DSO) and inventory levels were abnormal and potentially indicative of fraud. In my view, the reported figures are, in fact, consistent with basic financial common sense and the underlying industry structure. 1. Days Sales Outstanding (DSO) (1) Critics point out that Nvidia’s DSO increased from an average of 46 days in FY2020–2024 to 53 days in Q3 FY26 and claim this deterioration is a sign of financial irregularities or even fraud. (2) However, this view ignores the change in accounts receivable concentration: the share of receivables attributable …

VR200 NVL144 Rail Survey Update Although competition is everywhere—and rails are classified as a higher-replaceability Group C component in Nvidia’s AI server parts categories—my latest supply chain checks indicate that the VR200 NVL144 currently in development, testing, and assembly is using rails solely supplied by King Slide. The part-number prefix is the same as that of the GB300 NVL72.

VR200 NVL144滑軌調查更新 雖說競爭無所不在，且滑軌在Nvidia的AI伺服器零組件分類中屬於替換程度較高的Group C，但最新的供應鏈調查顯示，目前開發測試組裝中的VR200 NVL144，採用的滑軌是由川湖獨供，料號開頭與GB300 NVL72的相同。

@elonmusk remarks at the latest Tesla shareholder meeting validate my previous analysis and predictions regarding his semiconductor strategy. The following brief analysis will focus on why Musk wants Tesla to build its own chip production plants. 1. Musk said, “So I’m hopeful that we can within less than a year of AI5 starting production, we can actually transition in the same fab to AI6 and double all of the performance metrics.” I previously projected that AI6 would enter mass production in 2027. Many questioned whether only a one-year gap between AI5 and AI6 was realistic, but Musk’s comments validate that timeline (setting aside whether…

@elonmusk 在Tesla最新的股東會發言，驗證了我先前對馬斯克 (Musk) 在半導體佈局的分析與預測，以下簡短分析為何Musk希望Tesla能蓋自己的晶片生產工廠 1. Musk指出「So I’m hopeful that we can within less than a year of AI5 starting production, we can actually transition in the same fab to AI6 and double all of the performance metrics」 我先前預測AI6預計在2027年量產，當時不少人懷疑AI6與AI5生產時間僅相差一年過短，但Musk的說法驗證了我的預測 (先不討論屆時此目標能否順利執行)。 2. Musk希望Tesla能蓋自己的晶片生產工廠 這也驗證了我先前的分析，認為Tesla將AI6訂單轉移至Samsung的關鍵之一「能用極低的成本實際參與晶圓代工業務」。很顯然的，Musk預期此經驗將有助於Tesla蓋晶片工廠。 Musk擔憂未來晶片供應，因此才會說「But even when we extrapolate the best case scenario for chip production from our suppliers, it’s still not enough」。 然而，我認為至少對台積電而言，晶片供應不是問題，起碼台積電不會是供應鏈的最大瓶頸。台積電…

NVIDIA AI Server Power Roadmap: Kyber’s Next-Generation Strategy from GPU/Rack-Level to Data-Center Scale Full story https://t.co/4SIW7blFca https://t.co/mEHWaulpmm

NVIDIA AI伺服器電力規格藍圖：從GPU/機櫃走向資料中心的次世代Kyber戰略 閱讀全文 https://t.co/7r6D5j622S https://t.co/FqTmRkrBTS

iPhone Air demand has fallen short of expectations, leading the supply chain to begin scaling back both shipments and production capacity. Most suppliers are expected to reduce capacity by more than 80% by 1Q26, while some components with longer lead times are expected to be discontinued by the end of 2025. This indicates that the existing Pro series and standard models already cover the majority of high-end user demand well, leaving little room to carve out new market segments and positioning — as previous attempts with the mini, Plus, and now Air models have all failed to gain traction.