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 closer attention, as over the long term, the non-AI memory market is likely to reach supply-demand balance more efficiently than the AI memory market. 4. Had Apple defined memory as a “supply constraint” rather than a “cost pressure” during last week’s call, it likely would have benefited non-AI memory stocks—an outcome some short-term/bullish investors were looking for. Coincidentally, non-AI memory stocks across various markets saw a correction following the earnings call. Yet, few recent discussions have interpreted this pullback through the lens of Apple’s guidance/implications for demand. 5. Given Apple’s strong bargaining power and the fact that iPhone DRAM and NAND flash consumption accounts for around 20–25% of the global smartphone memory market, Apple’s perspective may not be an ideal proxy for broader non-AI/smartphone memory demand. In comparison, outlooks from Qualcomm and MediaTek provide a more useful reference for evaluating medium- to long-term non-AI/smartphone memory demand trends. 6. In my view, valuations of non-AI memory stocks may not yet fully reflect the strong earnings momentum expected this year. Looking further ahead to 2027–2028, AI memory demand should continue to offer better visibility.

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。

A few quick thoughts on Apple/iPhone memory price hikes: 1. The 1Q26 LPDDR price hikes mentioned in the news are pretty close to what I’ve heard. NAND flash increases, however, are a bit lower. 2. iPhone memory pricing is now negotiated quarterly instead of every six months, so expect another hike in 2Q26. Right now, the 2Q26 QoQ increase looks similar to 1Q26. 3. For most non-AI brands, even if you’re willing to pay up, there’s no guarantee you’ll get the supply. The fact that Apple can lock in a deal like this shows just how much leverage they have. 4. Higher memory costs will hit iPhone gross margins. But Apple’s playbook is clear: use the market chaos to their advantage—secure the chips, absorb the costs, and grab more market share. They’ll make it back later on the services side. 5. The cost pressure from memory could be a hot topic for investors and analysts at Apple’s earnings call this week. What Apple says could actually shake up other industries’ stocks more than Apple’s own or its suppliers'. 6. Apple’s current plan for 2H26 new iPhone 18 models is to avoid raising prices as much as possible—at least keep the starting price flat, which is helpful for marketing. 7. Apple has realized that it’s not just memory and T-glass—other components could also run short as the AI server boom continues to squeeze the rest of the supply chain. https://t.co/yskVn1BDo4

簡單聊一下Apple/iPhone的記憶體漲價： 1. 這篇新聞提到的1Q26 LPDDR漲價幅度跟我認知的接近，NAND Flash的漲價幅度略低。 2. iPhone的記憶體報價現在是按季談，而非半年，所以iPhone的記憶體報價在2Q26還會再漲一次。目前看2Q26的QoQ漲價幅度約跟1Q26接近。 3. 絕大部分非AI產業的品牌客戶，就算願意付錢也不見得能得到記憶體供應保證，所以Apple能談成這樣已經算很強了。 4. 記憶體報價上漲會影響iPhone的毛利率。但Apple的策略是，趁記憶體市場混亂時，透過確定能拿到貨與吸收成本的優勢，提升市佔率，後續再用服務業務賺回來。 5. 記憶體報價上漲導致成本上升的議題，應該是投資人與分析師對本週Apple法說的關注重點之一。Apple對這件事的說法，可能對其他產業的股價造成的波動，會遠高於對Apple自身與其供應鏈股價的影響。 6. Apple目前對2H26新款iPhone 18的定價策略是「盡可能不漲價」，至少起始價不動，有利行銷宣傳。 7. Apple已經意識到，在記憶體跟T-glass後，可能還會有其他零組件也會受到AI伺服器產業的影響而供應短缺。 https://t.co/yskVn1BDo4

My latest supply-chain surveys indicate that Carl Zeiss SMT will significantly expand its EUV and higher-ASP immersion DUV optical system capacity by 20-25% YoY and 40-50% YoY, respectively, in 2027 to meet robust demand from ASML. Coupled with stronger-than-expected shipment outlooks for 2026, ASML's revenue is projected to reach €38–40bn in 2026 and €45–47bn in 2027, outperforming market consensus of ~€34–36bn and ~€41–43bn. Key drivers behind the upside vs. consensus are as follows: 1. 2026 Growth Drivers: ➢ EUV and DUV shipments are estimated at 67 and 355 units, surpassing consensus of 53–55 and 310–320 units. ➢ Driven by strong 2nm demand, TSMC has upwardly revised its 2026 EUV orders twice: from an initial 22 units to 25 units last October, and currently to 28 units. ➢ To capture robust demand from Chinese memory makers, ASML plans to launch a new immersion DUV model NXT:1965i in 4Q26. As a down-spec version of the 1980i series, it complies with U.S. export controls while addressing Chinese clients' needs, serving as a key growth driver for 4Q26 and 2027. 2. 2027 Growth Drivers: ➢ 2027 capacity for both EUV and immersion DUV is currently fully booked thanks to the strong demand; further shipment upside will hinge on ASML’s supply-side improvements. ➢ Based on Zeiss SMT’s expansion, ASML’s 2027 shipments are projected to reach 80–85 EUV systems and 380–400 DUV systems. ➢ Boosted by the new 1965i launch, China’s procurement of high-ASP immersion DUV is expected to grow ~40% YoY in 2027. 3. Additional Upside Potential to Monitor: ➢ Potential ASP hikes for EUV and DUV systems amid prolonged supply tightness. ➢ Ongoing upward revisions for the new 1965i model from Chinese memory suppliers. ➢ Potential Intel Upside: The above estimates exclude incremental EUV demand from Intel. Although Intel has not officially placed orders (explaining its previously lower-than-expected Capex guidance), it has entered discussions with ASML for additional bookings. Intel's incremental EUV demand is projected at 20–30 systems over 2026–2027, of which 3–5 are High-NA EUV.

我最新的供應鏈調查顯示，Carl Zeiss SMT (蔡司半導體) 為滿足ASML的強勁需求，將顯著擴增2027年EUV與較高ASP的浸潤式DUV光學系統產能分別約20-25% YoY與40-50% YoY，加上2026年EUV與DUV出貨優於預期，故預估ASML 2026與2027年營收將分別達到約380-400億與450-470億歐元，優於市場共識的340-360億與410-430億歐元。 以下是優於市場共識的關鍵原因： 1. 2026年關鍵動能： ➢ 預估EUV與DUV出貨分別達67台與355台，優於共識的53-55台與310-320台。 ➢ 台積電因2nm強勁需求，接連兩次上修2026年EUV訂單，從最初的22台，於去年10月上修到25台，再到目前的28台。 ➢ ASML為滿足中國記憶體廠商的強勁需求，預計在4Q26推出新款浸潤式DUV NXT:1965i。此新款DUV由1980i系列降規而來，同時符合美國出口規範並更能滿足中國記憶體業者需求，將貢獻4Q26與2027年動能。 2. 2027年關鍵動能： ➢ 受益於強勁需求，2027年的EUV與浸潤式DUV目前已銷售一空，能否再提升出貨量取決於ASML的供應改善狀況。 ➢ 根據Zeiss SMT的擴產規劃，ASML在2027年的EUV與DUV出貨將分別達到80-85台與380-400台。 ➢ 受益於新款DUV 1965i，中國對高單價浸潤式DUV採購量量在2027年成長約40% YoY。 3. 需關注潛在上調可能： ➢ ASML因供應緊張而調升EUV與DUV價格。 ➢ 來自中國記憶體廠商強勁的需求，新款浸潤式DUV 1965i訂單仍在上修中。 ➢ 以上分析未計入來自Intel的新EUV需求。Intel雖未就新需求正式下單 (故先前法說的資本支出低於預期)，但已跟ASML開始協調加單事宜。目前預估Intel對EUV額外需求在2026與2027年共20-30台 (其中3-5台是High-NA EUV)。

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 2026 assembly GM: Beyond the aforementioned Nvidia AI server design trends, rising component prices will also dilute reported GM. 4. Wistron and its largest AI server customer, Dell, rallied together last September on buy-side expectations of Wistron’s market share gains and GM expansion (from ~5.5–6% to 6%+). While Wistron’s shares have remained elevated in recent months following an earlier rerating, it is worth noting that Dell’s share price has retraced to its September starting point amid profitability/GM concerns, following sell-on-news after a sharp backlog increase.

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。

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 only clearly defined for P1–P9, and Jensen Huang’s ultimate goal was to secure the P12 land as well. 4. Before Huang raised the idea of securing P10 and P11 land, buy-side consensus for TSMC’s 2026 capex was US$45–50bn. 5. Even with tight advanced-node and packaging capacity, TSMC stays highly disciplined on demand approval, often approving less than customers’ most optimistic requests. Want the most aggressive capacity? Put more cash down, book the land, and pay to grow capacity—just as Huang did. That said, with resources increasingly constrained across the board, even if other customers try to follow Huang’s playbook, securing the most aggressive capacity is becoming harder and harder.

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。

AI server assembler Wiwynn’s recent 4Q25 gross margin miss (7.2% vs. consensus estimates of 8–8.3%) has triggered a share price correction and reignited investor concerns. While rising component costs can mechanically dilute reported gross margins, it is more critical to examine the underlying profitability trends in server assembly through the lens of structural changes in AI server design. Nvidia continues to increase the level of design integration in AI servers to boost token output per unit of space and power, addressing the constraints of limited data-center space and scarce electricity supply. Highly integrated designs also help improve manufacturing efficiency, which in turn benefits supply-chain management and reduces maintenance costs. VR200 NVL72 is a representative example of this high-integration approach. Based on my industry surveys, its compute tray adopts a cable-less design, resulting in an ~40% reduction in the number of component items vs. GB300 NVL72. However, higher integration also comes with a greater share of Nvidia-specified components and a shrinking scope for customization, both of which are unfavorable to assembler profitability. My industry survey indicates that components classified as “customization not allowed” in the VR200 NVL72 compute tray have risen to 20–22%, well above the 5–7% level seen in GB300 NVL72. Moreover, VR200 NVL72 has directly eliminated "ODM-designed" component items entirely. In fact, this trend toward higher integration began with GB300 NVL72. My industry checks suggest that white-box GB200 NVL72 assembly gross margins were still around 18–20% in 1Q25, but by 1Q26, white-box GB300 NVL72 margins are expected to decline sharply to 6–8%. Importantly, this trend is likely to persist. Looking ahead, Kyber rack designs are expected to feature an even higher degree of integration and face greater margin pressure than the current Oberon generation. Assemblers will not remain passive in the face of these challenges. Beyond straightforward improvements in manufacturing efficiency, there are currently three practical approaches to addressing margin pressure in the server business, although their effectiveness remains to be seen: 1. optimizing product mix by increasing exposure to higher-margin ASIC AI servers; 2. securing NPI opportunities to enhance bargaining power with customers during the early stages of mass production; and 3. pursuing vertical integration strategies—ranging from qualifying in-house or group-affiliated server components through Nvidia validation, to expanding the business scope toward providing full data-center build-out solutions.

近期AI伺服器組裝廠商緯穎公布低於預期的4Q25毛利率 (7.2% vs. 市場預期的8-8.3%)，導致股價下跌，此事又引起投資人關注。雖零組件漲價也會稀釋帳面上的毛利率，但更重要的是從AI伺服器設計的本質上去檢視組裝的獲利能力趨勢。 Nvidia持續提升AI伺服器的設計整合程度，以提升每單位空間與電力的Token產出，來因應資料中心有限空間與稀缺電力挑戰。此外，高度整合的設計也有助於改善生產效率，進而有利供應鏈管理與降低維修成本。 VR200 NVL72就是設計高度整合的例子，我的產業調查顯示，其Compute tray導入無線化設計（Cable-less），零組件項目顯著減少約40% (vs. GB300 NVL72)。 但隨整合程度提高，伴隨而來的是Nvidia指定用料比重提升，與客製化空間也遭到壓縮，這些都不利組裝廠商的獲利能力。我的產業調查顯示，VR200 NVL72 compute tray「不允許客製化規格」的零組件項目佔比提升至20-22%，遠高於GB300 NVL72的5-7%，且VR200 NVL72更直接取消「ODM設計」的零組件項目。 事實上，提高整合程度的設計趨勢自GB300 NVL72就已開始。我的產業調查指出，白牌GB200 NVL72在1Q25的組裝毛利率尚維持在18-20%，但到1Q26，白牌GB300 NVL72的毛利率預估將顯著下滑至6-8%。 值得注意的是，上述趨勢只會繼續延續，未來Kyber機櫃的設計整合程度與毛利率壓力將比現在的Oberon更高。 組裝廠商們當然不會做坐以待斃。面對上述伺服器事業的毛利率挑戰，除單純改善生產效率，實務上目前有3種做法應對，不過實際幫助仍需觀察。3種做法是： 1. 改善產品組合：提升毛利率較高的ASIC AI伺服器出貨比重、 2. 爭取NPI以提高量產初期對客戶的議價力、與 3. 垂直整合策略：小至爭取自家/集團內的伺服器零組件通過Nvidia驗證，大至將事業範疇拓展至提供資料中心建置方案。

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 currently has lower visibility, potentially due to delays in mass production, but the current target spec remains 896K3. 4. Nvidia’s AI servers continue to push the technological frontier, making extensive testing entirely normal, including previously rumored PTFE from Thaifu and Taichem. 5. This event doesn't change the long-term trend of ongoing upgrades in PCB material spec. 6. The short-term stock price volatility reflects Nvidia’s trade-offs and compromises between real-world constraints and ideal targets. Similar situations are likely to recur in the future.

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 improves deployment flexibility under data-center power constraints. It also suggests NVIDIA is increasingly factoring real-world physical infrastructure limits into product specifications. 2. VR200 NVL72 upgrades the GPU thermal solution for both MaxQ and MaxP. ➢ Both use a micro-channel cold plate (MCCP) paired with a gold-plated lid. ➢ The market has been expecting a move to a micro-channel lid (MCL) once GPU TGP reaches 2.3 kW; however, MCL mass production is unlikely before 2H27. 3. Benefiting from upgrades across the GPU, HBM, and NVLink Switch, VR200 NVL72 delivers roughly 3.5×/5× the training/inference AI computing power of GB300 NVL72, driving a sharp increase in rack-level power demand. ➢ VR200 NVL72 upgrades power shelves to 3*3U 110 kW (6*18.3 kW PSUs), vs. the most common GB300 NVL72 configuration of 8*1U 33 kW shelves (6*5.5 kW PSUs). ➢ VR200 NVL72 power shelves adopt an 3+1 redundancy design. ➢ VR200 NVL72 raises the power whip rating to 100A (vs. 60A for GB300 NVL72), tightening data-center power infrastructure requirements (e.g., busway and tap-off boxes). 4. VR200 NVL72 relies even more heavily on liquid cooling. ➢ Both the compute and NVSwitch trays adopt a fanless design. ➢ Rack technology cooling system flow (TCS flow) nearly doubles (~+100%) vs. GB300 NVL72, benefiting spec and/or count upgrades for CDUs, manifolds, cold plates, and quick disconnects (QDs). ➢ Rack airflow requirements fall by roughly 80% (in CFM) vs. GB300 NVL72. 5. VR200 NVL72 compute tray adopts a midplane for the first time, enabling a truly cableless design. ➢ Key midplane specs include 44 layers (22+22), M9 CCL (EM896K3), and an approximate size of 420×60 mm. ➢ Jensen noted that this design can reduce compute-tray assembly time from roughly two hours to about five minutes (vs. GB300 NVL72). 6. Rubin CoWoS are estimated at 300–350k wafers in 2026, with pilot production expected in early 1Q26 and mass production by late 2Q26. VR200 NVL72 rack assembly is expected to enter mass production by the end of 3Q26; factoring in yield ramp, 2H26 rack shipments are estimated at ~5,000–7,000 units. 7. Once rack power approaches ~200 kW, 54V distribution begins to face material constraints in space (busbars/cabling) and conversion efficiency. As a result, VR200 NVL72 can be viewed as the final generation of the Oberon rack architecture. To address the rising rack-power requirements driven by continued AI compute scaling, NVIDIA is expected to transition to the next-generation Kyber rack design, which supports 800V HVDC.

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，VR200 NVL72的AI訓練/推理算力約是GB300 NVL72的3.5/5倍，機櫃的供電也因此大幅提升。 ➢ VR200 NVL72的power shelf升級到3*3U 110kW (6×18.3kW PSU) (vs. GB300 NVL72最普遍的8*1U 33kW (6×5.5kW PSU)) ➢ VR200 NVL72 power shelf採3+1備援設計 ➢ VR200 NVL72的電力進線 (power whip) 提升到100A (vs. GB300 NVL72的60A)，這對資料中心供電要求更高 (如母線 (busway) 與插接箱 (tap-off box) 等設計)。 4. VR200 NVL72的散熱設計更依賴液冷方案。 ➢ Compute & NVSwitch tray均採用無風扇設計 (fanless) ➢ 機櫃冷卻液流量 (TCS flow) 幾乎增加100% (vs. GB300 NVL72)，有利CDU、分歧管、水冷版與QD規格與/或數量升級 ➢ 機櫃風量 (Air flow) 需求則降低約80% (以CFM計) (vs. GB300 NVL72)。 5. VR200 NVL72的Compute tray首度採用midplane，落實無線設計 (cableless)。 ➢ Midplane關鍵規格包括：44層 (22+22)、M9 CCL (EM896K3)、尺寸約420*60 (mm)。 ➢ 黃仁勳提到，受益於此設計，Compute tray的組裝時間可由2小時顯著降低到5分鐘 (vs. GB300 NVL72)。 6. 2026年Rubin CoWoS投片量約30-35萬片，預計在1Q26初開始小量試產並於2Q26末量產。VR200 NVL72的機櫃組裝量產預計在3Q26底，考慮到良率爬坡，預計2026/2H26機櫃出貨量約5,000-7,000櫃。 7. 當機櫃功耗達約200kW，54V配電會開始遭遇顯著的空間占用 (銅排/線材) 與轉換效率瓶頸，故VR200 NVL72可視為Oberon機櫃的最後一代方案；後續面對AI算力提升帶來的挑戰，將採用支援800V HVDC的次世代Kyber機櫃設計因應。

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獻，雖Intel在未來數年仍無法與台積電競爭，但這意味著IFS事業最壞將過，未來14A或更先進的製程，或將取得更多Apple與其他一線客戶的訂單，長期展望轉正向。

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 for 1Q26. Apple's plan is for Intel to begin shipping its lowest-end M processor, utilizing the 18AP advanced node, as early as 2Q–3Q27, but the actual timeline remains contingent on development progress following the receipt of PDK 1.0/1.1. Apple’s lowest-end M processor is currently used in the MacBook Air and iPad Pro mainly, with combined shipments of roughly 20 million units for 2025. As MacBook Air shipments in 2026 may be impacted by a new more-affordable MacBook model using an iPhone-class processor, shipments of lowest-end M processor in both 2026 and 2027 are expected to be 15–20 million units. In absolute terms, order volumes for the lowest-end M processor are relatively small and virtually no material impact on TSMC’s fundamentals or its technology leadership over the next several years. However, the signaling and trend implications for Apple and Intel are meaningful: 1. For Apple: In addition to showing strong support for the Trump administration’s strongly promoted “Made in USA” policy, Apple, while clearly expected to remain highly dependent on TSMC’s advanced nodes for the foreseeable future, still needs to secure a second source to meet supply-chain management requirements. 2. For Intel: The significance of winning Apple’s advanced-node orders far exceeds the direct revenue and profit contribution from this business. Although Intel will still be unable to compete head-to-head with TSMC over the next several years, this suggests that the worst may soon be over for the IFS business. Looking ahead, the 14A node and beyond could capture more orders from Apple and other tier-one customers, turning Intel’s long-term outlook more positive.

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 to major customers rose from an average of 23.8% in FY2020–2024 to 65% in Q3 FY26. (3) It is entirely reasonable for DSO to rise as receivables become more concentrated in a handful of large customers. This reflects the bargaining power of key accounts—especially cloud service providers (CSPs), which have historically operated with longer payment terms. (4) Another flaw in the critics’ analysis is the peer group the view chose. It compared Nvidia’s DSO to companies whose primary customers are not CSPs. A more appropriate comparison set is suppliers whose main customers are the same CSPs—such as Arista, Celestica, and Vertiv—whose DSOs typically run above 60–70 days. Against that backdrop, Nvidia’s 53-day DSO looks entirely reasonable, not anomalous. 2. Inventory (1) Critics describe the 32% QoQ increase in Nvidia’s Q3 FY26 inventory as a “paradox,” and cite Q2 FY23 as a counterexample, claiming that inventory declined 18% QoQ when the H100 was launched. (2) There are two factual errors in this argument: – First, Q2 FY23 inventory did not decline; it increased by roughly 23% QoQ to USD 3.889 billion. – Second, the 32% increase in Q3 FY26 inventory is broadly in line with the ramp in upstream capacity: TSMC’s CoWoS average monthly output is estimated to have grown by roughly 25–30% QoQ to around 60kwpm in 3Q25. Within Nvidia’s Q3 FY26 inventory, the largest item was Work-in-Process (WIP), which represented 44.2% of total inventory and surged about 98% QoQ to USD 8.735 billion. This is best understood as reflecting the ramp and mass production of the new Blackwell B300 GPU starting in 3Q25. Far from being evidence of fraud, it shows that Nvidia had already been preparing for strong demand for the B300.

在Nvidia公布Q3 FY26財報後，有觀點質疑該財報中的應收帳款週轉天數（Days Sales Outstanding；DSO）與庫存（Inventory）異常且疑似詐欺。但我認為，該財報結果反而符合財報常識與產業現況。 1. 應收帳款週轉天數： (1) 質疑者認為Nvidia的DSO由FY2020－2024的平均46天增加到Q3 FY26的53天，此惡化是金融異常與詐欺。 (2) 然而，上述並沒有考慮到應收帳款的客戶集中度變化，從FY2020－2024的平均23.8%，顯著提升到Q3 FY26的65%。 (3) DSO隨應收帳款的客戶集中度顯著提升而增加很合理，這反映了大客戶的議價力，特別Nvidia目前面對的是付款條件向來較久的各大CSP。 (4) 質疑者的另一錯誤是將Nvidia與其他主要客戶非CSP的廠商之DSO來比較。事實上，其他主要客戶為各大CSP的廠商，如Arista、Celestica與Vertiv等，其DSO普遍在60－70天以上，這是另一個驗證Nvidia DSO為53天的合理證明。 2. 庫存： (1) 質疑者將Nvidia的庫存金額在Q3 FY26增加32% QoQ解釋為庫存悖論，並舉出Q2 FY23的庫存金額隨H100推出而季減18%作為反例。 (2) 質疑者的論點有兩個錯誤： – 首先，Q2 FY23的庫存金額是成長約23% QoQ至38.89億美元，而非減少。 – 其次，Q3 FY26庫存金額增加32%，與台積電CoWoS平均月產能在3Q25成長25－30% QoQ至約60kwpm趨勢一致，而Q3 FY26庫存金額中佔比最高(44.2%) 的在製品 (Work-In-Process) 顯著成長約98％ QoQ至87.35億美元，則是反應新款Blackwell B300晶片在3Q25開始爬坡與量產，這非但不是詐欺，反而證明Nvidia為B300晶片強勁需求做準備。

RT @mingchikuo: @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CEO魏哲家先前曾對Musk說過「只要你肯付錢一定會有晶片」。台積電雖然對擴產向來謹慎，但從過去歷史看，市況好的時候台積電的產能常會緊張，但向來不是最大的供應缺口。 若未來有更多資訊，再針對TeraFab的供需做更精確的量化分析，這邊也先不分析Tesla能否順利蓋自己的晶片廠。以下先重點討論如果台積電在供應上沒問題，為什麼馬斯克還想蓋Tesla的晶圓廠。 首先是地緣政治。Musk公開表達過對半導體產能高度集中在台灣的擔憂。即便到2030年，台積電在美國的先進製程與封裝產能仍有限，特別是先進封裝，可能頂多佔全球的10%。我相信Musk也意識到這件事，所以他才會說「And then you go to solve memory and packaging too」。 其次是研發支援與生產彈性。預期Tesla未來將有很大的成長潛力，但對現在的台積電而言，Tesla就是一個二線客戶，因此Tesla在台積電得到的研發支援與生產彈性肯定低於一線客戶如Apple、Nvidia等，這也是為何Tesla將AI6訂單轉移至Samsung的主因之一。而解決這個問題的終極手段，就是擁有自己的晶片工廠。 最後是整合優勢。Musk的新技術與產品，往往位處人類技術前沿，若能依自身需求客製化關鍵設計與製造環節 (晶圓廠就是其中一個關鍵)，有利高度整合最終產出並最大化垂直整合效益。 參考資料： 1. Tesla最新股東會會議記錄 https://t.co/stoK3Zwq0J 2. 台積電CEO魏哲家對Musk說「只要你肯付錢一定會有晶片」 https://t.co/IuqqXewSRj 3. Musk公開表達過對半導體產能高度集中在台灣的擔憂 https://t.co/K7UXdHZT5J

RT @mingchikuo: @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 the plan ultimately executes as stated). 2. Musk wants Tesla to build its own chip production plants. This also validates my earlier view that a key reason for shifting AI6 orders to Samsung was to gain real-world foundry experience at an exceptionally low cost. It’s clear Musk expects that experience to help Tesla stand up chip production plants of its own. Musk is worried about future chip supply, hence his comment: “But even when we extrapolate the best case scenario for chip production from our suppliers, it’s still not enough.” However, I don’t think chip supply is the core issue—at least not at TSMC. TSMC is unlikely to be the primary bottleneck. TSMC CEO C.C. Wei has told Musk before, “If you’re willing to pay, there will be chips.” While TSMC is known for its cautious approach to capacity expansion and often runs tight when demand is strong, history shows it has rarely been the primary bottleneck in the supply chain. When more information emerges, I’ll provide a tighter quantitative read on TeraFab’s supply–demand balance. For now, I’ll leave aside whether Tesla can successfully build its own chip production plant and focus on why Musk still wants a Tesla chip production plant even if TSMC’s supply isn’t the constraint: The first factor is geopolitics. Musk has publicly voiced concerns about the concentration of advanced node capacity in Taiwan. Even by 2030, TSMC’s advanced-node and advanced-packaging capacity in the U.S. will remain limited—especially on the packaging side, likely no more than about 10% of its global capacity. I believe Musk realizes this as well, which is why he said, 'And then you go to solve memory and packaging too.' The second factor is R&D support and production flexibility. It's expected that Tesla has significant growth potential, but within TSMC today it is still a second-tier customer relative to the likes of Apple and Nvidia. That naturally means less priority on R&D support and production flexibility—one of the main reasons Tesla moved AI6 to Samsung. The ultimate remedy is to control its own chip production plant. Finally, there's the integration advantage. Musk's new technologies and products are often at the cutting edge. The ability to customize key design and manufacturing segments to his specific needs—with chip production being a critical one—enables a highly integrated final product and maximizes the benefits of vertical integration. References: 1. Tesla's 2025 Annual Shareholder Meeting Transcript https://t.co/stoK3Zwq0J 2. TSMC CEO C.C. Wei told Elon Musk, "If you’re willing to pay, there will be chips." https://t.co/IuqqXewSRj 3. Musk has publicly voiced concerns about the concentration of advanced node capacity in Taiwan. https://t.co/K7UXdHZT5J

As I predicted a year ago, M5 chip is the most important hardware upgrade in the new Vision Pro. https://www.apple.com/apple-vision-pro/

如我一年前所預測，M5晶片是新款Vision Pro最重要的硬體更新 https://www.apple.com/apple-vision-pro/