←
郭明錤|Ming-Chi Kuo1,382 posts
郭明錤|Ming-Chi Kuo avatar

郭明錤|Ming-Chi Kuo

@mingchikuo

香港天風國際證券分析師,分享科技產業趨勢觀察|TF International Securities (HK) analyst sharing tech trend insights

◷ Joined March 2011▧ 126 Photos and videos
郭明錤|Ming-Chi Kuo avatar
郭明錤|Ming-Chi Kuo@mingchikuo·Nov 16

Google TPU update: New OpenAI and Anthropic orders extend shipment visibility to 2029; v10 lineup overhauled (now Quillfish); MediaTek and Broadcom lead as the top suppliers to watch Key takeaways from my latest industry survey and research conclusions: 1. Order visibility and shipment plans: Strong new orders from OpenAI and Anthropic have extended Google TPU shipment visibility to 2029, ahead of Nvidia GPUs and other AI ASICs (mostly 4Q27–2Q28). (1) From 2028, v9 will become the mainstay of Google’s own compute buildout, so v8-series orders (8t / 8i or their design variants) can be viewed as coming from OpenAI and Anthropic. (2) Combined 2028–2029 demand: about 12M units for OpenAI (8t, designed by MediaTek) and about 14M for Anthropic (8i, designed by Broadcom), with both growing year over year. OpenAI buys outright, while Anthropic uses a mix of leasing and outright purchases. (3) Given supply chain constraints, more reasonable TPU shipment estimates for 2027–2029 are 8–9M, 14–15M, and 16–18M units, respectively. The market broadly expects AI capex to peak in 2028, yet this plan shows TPU shipments still growing in 2029, driven by direct orders from AI service leaders OpenAI and Anthropic. 2. v10 lineup and supply chain changes: v10 has changed from Icefish and Merope / Bladerunner to Quillfish, with mass production expected in 2029. MediaTek has won the design service order. (1) Icefish is now v11, with mass production expected in 2030 and the design service order also going to MediaTek. Merope has lower visibility: no supplier has been chosen, and it may be canceled. (2) MediaTek will be the sole supplier for v9 (Humufish & Triggerfish) and v10 (Quillfish), accounting for about 45–50%, 50–55%, and 55–60% of total TPU supply in 2027–2029, respectively. TPU’s contribution to its revenue and profit should rise each year. (3) From v9 onward, all new TPUs use Intel’s EMIB-T advanced packaging. Given limited initial production capacity, the v8 series should remain the bulk of TPU shipments through 2029. v8 is expected to be the main source of TPU profit growth for MediaTek and Broadcom over the next few years, thanks to significant accumulated learning-curve benefits and the fact that both handle procurement of key components and materials. 3. Market attention and supplier outlook: Recent events should raise and sustain market attention on Google TPU, with key design service providers MediaTek and Broadcom likely to be the supply chain’s top focus. (1) Uncertainty over Nvidia’s CEO succession helps Google in discussing post-2028 long-term plans with customers. OpenAI and Anthropic are both key TPU customers, so their IPOs should also raise and sustain attention on TPU, and MediaTek and Broadcom are the suppliers the market most readily associates with it. (2) MediaTek is clearly the biggest TPU winner of the next few years. Its position as sole design supplier from v9 onward, together with a supply share that rises each year alongside TPU shipments, will expand its TPU revenue. The high-margin 8t series, with shipment visibility through 2029, is a key profit source. (3) Because Broadcom leads the 8i design, its gross margin is even higher than the 8t’s, supporting Broadcom’s profit growth. Note, however, that Broadcom’s focus over the next few years will be non-TPU XPU and AI networking chip orders, with non-TPU revenue expected to surpass TPU revenue as early as 2027.

郭明錤|Ming-Chi Kuo avatar
郭明錤|Ming-Chi Kuo@mingchikuo·Nov 16

Google TPU 重點更新:OpenAI 與 Anthropic 新訂單推升出貨能見度至 2029 年、v10 產品線大幅調整 (改為 Quillfish)、聯發科與 Broadcom 為供應商中之領先關注焦點 我最新的產業調查與研究結論: 1. 訂單能見度與出貨規劃: 受惠於 OpenAI 與 Anthropic 強勁新訂單,Google TPU 出貨能見度已達 2029 年,優於 Nvidia GPU 與其他 AI ASIC(多數為 4Q27–2Q28)。 (1) 自 2028 年起,v9 將成 Google 自建算力的主力方案,v8 系列(8t/8i 或其設計變種)訂單可視為來自 OpenAI 與 Anthropic。 (2) 2028–2029 年合計需求數量:OpenAI 約 12M(聯發科設計的 8t),Anthropic 約 14M(Broadcom 設計的 8i),兩者需求均呈逐年成長趨勢。OpenAI 採買斷,Anthropic 則是租賃與買斷都有。 (3) 考量供應鏈限制,2027–2029 年較合理的 TPU 出貨估算分別為 8–9M、14–15M、16–18M。市場普遍預期 AI 資本支出於 2028 年見頂,但此規劃顯示 TPU 在 2029 年仍具出貨成長動能,關鍵在於 AI 服務龍頭 OpenAI 與 Anthropic 的直接訂單。 2. v10 產品線與供應鏈調整: v10 由原先的 Icefish 與 Merope/Bladerunner 改為 Quillfish,預計 2029 年量產,設計服務訂單由聯發科取得。 (1) Icefish 改為 v11,預計 2030 年量產,設計服務訂單同由聯發科取得。Merope 能見度較低,供應商未定,也可能取消。 (2) 聯發科將是 v9(Humufish & Triggerfish)與 v10(Quillfish)的獨家供應商,2027–2029 年占整體 TPU 供應比重分別約 45–50%、50–55%、55–60%。TPU 對其營收與利潤的貢獻預計逐年增加。 (3) 自 v9 起新款 TPU 均採 Intel EMIB-T 先進封裝。考量初期量產能力,預期至 2029 年仍以 v8 系列為 TPU 出貨主力。v8 預期將是聯發科與 Broadcom 未來數年的 TPU 事業獲利成長的主要來源,因長期累積的學習曲線效益明顯,加上雙方負責關鍵零組件與材料採購。 3/ 市場關注度與供應商展望: 近期事件有助提升並維持 Google TPU 的市場關注度,關鍵設計服務商聯發科與 Broadcom 可望成為供應鏈首要焦點。 (1) Nvidia CEO 接班的不確定性,有利 Google 與客戶洽談 2028 年後的長期規劃。OpenAI 與 Anthropic 均為 TPU 關鍵客戶,其 IPO 也有助提升並維持 TPU 關注度,而聯發科與 Broadcom 是市場最容易聯想到的供應商。 (2) 聯發科無疑是未來數年 TPU 的最大贏家,身為 v9 後的獨家設計供應商,加上供應比重與 TPU 出貨成長同時逐年提升,均有利 TPU 事業營收規模擴大。高毛利率的 8t 系列出貨能見度達 2029 年,為關鍵利潤來源。 (3) 因 Broadcom 主導 8i 設計,故毛利率更高於 8t,有利該公司利潤成長。不過,需注意的是,Broadcom 未來數年的營運重點為非 TPU 的 XPU 與 AI 網路晶片訂單,預計最快 2027 年,非 TPU 營收將超越 TPU。

郭明錤|Ming-Chi Kuo avatar
郭明錤|Ming-Chi Kuo@mingchikuo·Nov 12

Trump says Tim Cook is the best person to take Jensen Huang’s place as Nvidia CEO, but I think where Cook’s strengths would shine most is cost management. Having watched Apple’s supply chain management from the sidelines for more than 10 years, and comparing that with what I’ve learned tracking and researching Nvidia in recent years, I think Nvidia still has clear room for improvement in supply chain and cost management. As for everything else, especially the most important part, planning and developing new technologies, this isn’t just about Cook. So far, the only CEO who has clearly proven he can do this well, bringing humanity’s most cutting-edge technologies from across fields into AI chips and rack solutions while keeping up a fast pace of commercial iteration, is Jensen Huang. (Of course, others will likely prove they can do this well in the future too.)

郭明錤|Ming-Chi Kuo avatar
郭明錤|Ming-Chi Kuo@mingchikuo·Nov 12

雖然 Trump 說 Tim Cook 是最適合接任黃仁勳 Nvidia CEO 位置的人選,但我覺得最能發揮 Cook 所長的應該只有成本管理。根據我側面觀察 Apple 供應鏈管理超過 10 年以上的經驗,來對比這幾年追蹤與研究 Nvidia 的心得,我覺得 Nvidia 在供應鏈成本管理上,還有明顯的改善空間。 其他方面,特別是最重要的新技術規劃與開發,不要說 Cook 了,目前已經證明能在維持快速商業化迭代更新下,把人類最先端的各領域技術導入 AI 晶片與機櫃方案的 CEO,這件事做得明顯好的只有黃仁勳(當然未來應該還會有其他人也能證明可以把這件事做很好)。

郭明錤|Ming-Chi Kuo avatar
郭明錤|Ming-Chi Kuo@mingchikuo·Nov 12

Key takeaways on Nvidia CEO Jensen Huang’s resignation: 1. Near term, the bad news is priced in: For the stock market, this could be a one-time clearing event. Huang has now fully cut ties with Nvidia, so Congress summoning him and any follow-up investigations should have less impact on Nvidia’s stock from here. As for how Huang’s next moves might affect Nvidia, it’s too early to call before the investigation results are out. 2. Google’s TPU is likely the biggest beneficiary: With Nvidia facing succession uncertainty, Google’s TPU has a chance to win more trust from customers and the supply chain. Its key partners, MediaTek and Broadcom, could be the market’s near-term focus. 3. Succession impact shows up in 2028 at the earliest: CFO Colette Kress is now interim CEO. The formal CEO transition and the organizational changes that come with it will likely start affecting Nvidia’s product development and competitiveness in 2028 at the earliest. 4. Capacity pre-booking and spec upgrades are the key focus for supply chain checks: Huang’s product strategy pushed the supply chain hard in two ways: capacity pre-booking and spec upgrades. It’s critical to watch for any changes in capacity expansion and new tech development across the supply chain. These are among the core reasons Nvidia’s supply chain trades at high P/E multiples. 5. China demand and export controls: The share of Nvidia’s AI chip revenue that ultimately ends up with Chinese buyers may be higher than what the company has officially stated. That’s why some investors worry Nvidia’s revenue will drop sharply once the US government tightens export controls. But China’s end demand won’t disappear because of export controls. Even under tighter rules, as long as overseas resale channels exist, Chinese buyers may still pay the premium, even at several times the price, and get Nvidia chips/racks through indirect routes.

郭明錤|Ming-Chi Kuo avatar
郭明錤|Ming-Chi Kuo@mingchikuo·Nov 12

重點分析 Nvidia CEO 黃仁勳辭職: 1. 短期利空出盡: 對股市而言,短期內可能是一次性的利空出盡。黃仁勳目前已與 Nvidia 完全切割,接下來國會傳喚黃仁勳及後續調查,對 Nvidia 股價的影響應該會較小。至於黃仁勳離職後的動向對 Nvidia 的影響,在調查結果出爐前,現在就下判斷仍言之過早。 2. Google TPU 為直觀下的最大受益者: 在 Nvidia 面臨接班不確定性下,Google TPU 可能有機會進一步取得客戶與供應鏈信任,其核心合作廠商聯發科與 Broadcom 或為短期市場焦點。 3. 接班影響最快自 2028 年浮現: 目前 CFO Colette Kress 擔任代理執行長。後續正式 CEO 交接與相應的組織改革,對 Nvidia 產品開發與競爭力的影響,預期最快可能自 2028 年開始浮現。 4. 產能預訂與規格升級為供應鏈研究焦點: 黃仁勳的產品策略對供應鏈有兩大激進影響:「產能預訂」與「規格升級」。未來需密切關注供應鏈在擴產與新技術開發方面有無變化,這是支撐 Nvidia 供應鏈高本益比的核心因素之一。 5. 中國需求與出口管制: Nvidia 最終流向中國買家的 AI 晶片營收比重,可能高於過去對外的正式聲明。因此,有些投資人擔心,美國政府實施更嚴格的出口管制後,Nvidia 營收將顯著衰退。不過,中國終端需求並不會因出口管制而消失。即使出口管制加嚴,只要海外仍存在轉售管道,即使轉售價格高出數倍,中國買家仍有可能承擔溢價並輾轉取得 Nvidia 晶片/機櫃。

郭明錤|Ming-Chi Kuo avatar
郭明錤|Ming-Chi Kuo@mingchikuo·Nov 12

Jensen Huang matters to Nvidia because he can take technologies his engineers believe are possible and turn them into products the whole company commits resources to, takes risks on, and has to deliver. That's why the supply chain gets such strong order visibility and is willing to keep up with aggressive spec upgrades. It also forces competitors to follow, which drives technology upgrades and supply chain investment across the entire AI industry. Between the two, deciding on aggressive spec upgrades and making them happen is the harder part, compared with giving the supply chain order visibility. Huang doesn't need to be the person at Nvidia who knows the most about new technologies like HVDC, CPO/NPO, or CCL/PTFE. His value is that he weighs strategy, organization, technology, and supply chain all at once, then makes the call quickly and puts resources behind it. Based on what's public so far, it looks hard for Huang to come out of this unscathed. If so, the practical move may be to gradually reduce Nvidia's reliance on Huang's personal decision-making while preparing a succession plan. Along the way, two things are worth watching to gauge where AI technology and investment are heading: whether Nvidia can keep developing new products and technologies as efficiently as before, and how competitors respond. Supply chain research will be an important basis for that judgment.

郭明錤|Ming-Chi Kuo avatar
郭明錤|Ming-Chi Kuo@mingchikuo·Nov 12

黃仁勳對 Nvidia 的重要性,在於他能把工程團隊認為可能實現的技術,轉化成全公司願意投入資源、承擔風險,而且必須交付的產品。正因如此,供應鏈才能有極高的訂單能見度,並配合激進的規格升級。而這也帶動了競爭對手的跟進,進一步推動整體 AI 產業的技術升級與供應鏈投資。 如果硬要比較,做出並落實激進規格升級的決策,比提供供應鏈訂單能見度更困難。黃仁勳不必是 Nvidia 內部最懂 HVDC、CPO/NPO 與 CCL/PTFE 等新技術的人,但他的價值,在於同時考慮到策略、組織、技術與供應鏈,並迅速拍板定案與投入資源執行。 就目前公開資訊,黃仁勳似乎很難全身而退。在這前提下,比較務實的做法,可能是階段性降低 Nvidia 對黃仁勳個人的決策依賴,同時布局接班計畫。而在這過程中,Nvidia 能否延續高效的新產品與新技術研發,以及競爭對手如何因應,將是觀察整體 AI 產業技術發展與投資趨勢的重點,而供應鏈研究則能提供重要的判斷依據。

郭明錤|Ming-Chi Kuo avatar
郭明錤|Ming-Chi Kuo@mingchikuo·Sep 30

My latest supply chain checks indicate that Nvidia has begun testing glass-free copper-clad laminate (CCL) based primarily on hydrocarbon (HC) resin as a replacement for glass-free PTFE-based CCL tested previously, paired with glass-free HC-based PP. The goal is to determine whether this material combination can meet the switch tray PCB requirements for Rubin Ultra NVL576, an eight-rack interconnected solution expected to enter mass production in 2H27. Preliminary tests conducted as part of efforts to improve PCB manufacturability show that glass-free HC CCL meets the switch tray’s high-frequency electrical requirements. Its electrical performance trails that of the original PTFE solution but exceeds that of M9-class solutions and even surpasses M10’s stated specifications. These results do not represent a final material selection, but they indicate that NVIDIA continues to seek higher PCB manufacturing yields and greater production efficiency while meeting electrical requirements. 1. Key Takeaways and Analysis (1) Previous NVL576 switch tray tests used glass-free PTFE CCL paired with glass-free HC PP. The latest round replaces the PTFE CCL with glass-free HC CCL while retaining the same HC PP part number. (2) Shengyi Technology is currently the leading supplier in the NVL576 switch tray material evaluations. The materials involved are SG5300N (glass-free PTFE CCL), SG1030N (glass-free HC CCL), and SIF09 (glass-free HC PP). (3) The primary objective of testing glass-free HC CCL as a replacement for glass-free PTFE CCL is to determine whether PCB manufacturing yields and production efficiency can be improved while still meeting high-frequency electrical requirements. (4) To further improve PCB manufacturability, subsequent tests may also evaluate replacing glass-free HC PP with woven-glass-reinforced HC PP. The challenge is whether the glass-reinforced material can still meet signal integrity and 0.4 dB/in loss requirements. (5) Notably, both SG1030N and SIF09 still contain small amount of PTFE, although HC is the primary material. These tests therefore do not represent a complete shift away from PTFE. Rather, they involve adjusting the primary resin system and material combination to balance high-frequency electrical performance with PCB manufacturability. 2. Conclusions (1) The long-term trend toward broader use of PTFE in advanced PCB materials, including CCL and PP, remains clear. However, the pace of shipment growth over the next two years remains uncertain. Key variables include NVL576 switch tray PCB’s final material selection, NVL576 adoption rate, and development progress on Kyber-like rack designs that use orthogonal backplanes. (2) HC use in advanced CCL and PP continues to expand. Beyond NVL576 switch trays, near-term growth drivers include the production ramp of M9-class materials beginning in 4Q26, followed by the adoption of M10-class materials. (3) Once specifications are finalized and mass production begins, it's expected that Shengyi Technology to accelerate sourcing from Chinese HC suppliers. These include existing suppliers such as Dongcai and Shengquan, as well as potential ones such as GCH Technology. This should help alleviate supply shortages and optimize its cost structure.

郭明錤|Ming-Chi Kuo avatar
郭明錤|Ming-Chi Kuo@mingchikuo·Sep 30

我最新的產業調查顯示,Nvidia 開始測試以碳氫樹脂(hydrocarbon resin,HC)為主的無布 CCL,取代先前測試的 PTFE 無布 CCL,並搭配 HC 無布 PP,評估能否滿足預計於 2H27 量產的 Rubin Ultra NVL576(8 機櫃互連方案)之 switch tray PCB 需求。 在提升 PCB 生產性的目標下,HC 無布 CCL 的初步測試結果符合 switch tray 的高頻電性要求(表現不及原 PTFE 方案,但優於 M9 等級方案,甚至優於 M10 的帳面規格)。上述測試不代表最終定案,但顯示 Nvidia 仍在滿足電性要求的前提下,尋求更好的 PCB 製造良率與生產效率。 1. 重點解讀與分析: (1) 先前測試 NVL576 switch tray 的組合,是 PTFE 無布 CCL 搭配 HC 無布 PP。近期的新一輪測試,則是以 HC 無布 CCL 替代原本的 PTFE 無布 CCL,HC 無布 PP 仍使用相同料號。 (2) 生益科技是目前 NVL576 switch tray 材料測試的領先供應商。相關材料包括 SG5300N(PTFE 無布 CCL)、SG1030N(HC 無布 CCL),以及 SIF09(HC 無布 PP)。 (3) 將 PTFE 無布 CCL 改為 HC 無布 CCL,主要是評估在改善 PCB 的製造良率與生產效率的目標下,同時滿足高頻電性要求。 (4) 為進一步優化 PCB 可製造性,後續也可能測試將 HC 無布 PP 換成 HC 有布 PP。導入 HC 有布 PP 的挑戰在於,能否仍達到訊號完整性與 0.4 dB/in 的損耗要求。 (5) 值得注意的是,SG1030N 與 SIF09 雖以 HC 為主要材料,但配方中仍含少量 PTFE。因此,這次測試並非完全去 PTFE 化,而是調整材料主體與組合,以兼顧高頻電性與 PCB 可製造性。 2. 結論: (1) PTFE 導入高階 PCB 材料(CCL 與 PP)的長期趨勢明確,但未來兩年的出貨成長節奏仍有變數。關鍵在於:NVL576 switch tray PCB 的最終材料方案、NVL576 的採用比重,以及採用正交背板的 Kyber-like 機櫃設計之研發進展。 (2) HC 在高階 CCL/PP 的應用持續擴大。短期成長動能除了 NVL576 switch tray,還包括自 4Q26 開始放量的 M9 等級材料,以及後續的 M10。 (3) 預期未來規格定案與開始量產後,生益科技將加速導入中國的 HC 供應商(如既有的東材與聖泉,潛在供應商呈和等),有助於供應缺口改善與優化成本結構。

郭明錤|Ming-Chi Kuo avatar
郭明錤|Ming-Chi Kuo@mingchikuo·Sep 1

MediaTek’s latest announcement on the deepening of its long-term partnership with Nvidia repeatedly uses the term “rack-scale.” This is the key to understanding the announcement, and it directly validates the central conclusion of my research from more than two months ago: MediaTek has upgraded the strategic positioning of its AI business from "IC / ASIC design" to "system-level design." The announcement link: mediatek.com/press-room/nvid… The announcement also explains how MediaTek plans to put this new positioning into practice. By working with Nvidia, MediaTek can bring custom chips and XPUs into Nvidia’s mature rack-scale ecosystem, so that once chip development is complete, they can be integrated more quickly into rack-scale systems for deployment in data centers and AI factories. For MediaTek, Nvidia’s mature rack-scale systems ecosystem means it does not need to start from scratch, giving this new strategy a much clearer path to execution. For Nvidia, MediaTek’s custom silicon design capabilities and customer base can help extend Nvidia’s rack-scale ecosystem further into the custom silicon market. This also explains why “rack-scale” is the key term in the announcement: it is where the two companies need each other most. MediaTek can move from chips to systems, while Nvidia can extend its rack-scale ecosystem into the custom silicon market, creating a win-win that opens up new growth opportunities for both.

Post attachment
郭明錤|Ming-Chi Kuo avatar
郭明錤|Ming-Chi Kuo@mingchikuo·Sep 1

聯發科最新公告宣布深化與 Nvidia 的長期合作,其中反覆出現的「rack-scale」,正是理解這份公告的關鍵,也直接驗證了我兩個多月前研究的核心判斷: 聯發科內部已將 AI 事業的策略定位,從「IC / ASIC 設計」提升至「系統級別設計」。 公告網址: mediatek.com/press-room/nvid… 這份公告也進一步說明,聯發科將如何執行這項新定位:透過與 Nvidia 合作,把客製化晶片 / XPU 導入 Nvidia 已成熟的機櫃級生態,讓客製化晶片在完成開發後,能更快整合成可部署於資料中心與 AI 工廠的機櫃級系統。 對聯發科而言,Nvidia 已成熟的機櫃級系統生態,使其不必從零開始,也讓新策略更具可執行性。 對 Nvidia 而言,聯發科的客製化晶片設計能力與客戶基礎,可協助其機櫃級生態進一步延伸至自研晶片市場。 這也說明了為何 rack-scale 是這份公告的關鍵字:這正是雙方最需要彼此的地方。聯發科藉此從晶片走向系統,Nvidia 則將機櫃級生態延伸至客製化晶片市場,形成雙方都能開拓新成長空間的雙贏局面。

Post attachment
郭明錤|Ming-Chi Kuo avatar
郭明錤|Ming-Chi Kuo@mingchikuo·Aug 31

Just as the market had come to believe that Rubin CPX had been dropped from Nvidia’s product roadmap, my latest industry checks indicate that Nvidia has revived the program, with production expected to begin in 1Q27. Compared with the previous design, the revived Rubin CPX delivers stronger prefill performance and features major changes to both its GPU specifications and rack architecture, underscoring the high priority Nvidia places on prefill solutions. Key changes: 1. Rubin CPX GPU specs CPX delivers near-Rubin compute performance and matches Rubin’s maximum power rating of 2,300 W per GPU. CPX moves to 168 GB of HBM4, vs. 288 GB on Rubin and 128 GB of GDDR7 on the previous CPX design. 2. Rack design The new CPX uses a standalone MGX ETL rack rather than sharing a rack with Rubin, as in the previous design. Customers can opt for 64, 128, 192, or 256 CPX GPUs depending on their needs. Within a CPX rack, each group of 64 CPX GPUs forms a rack module comprising eight compute trays (eight CPX GPUs per tray) and one switch tray. 3. Scale-up and scale-out NVLink is used only for scale-up among the eight CPX GPUs within each tray, with 1–1.5 TB/s of NVLink bandwidth per CPX (vs. 3.6 TB/s per Rubin). Inter-tray scale-out within each rack module runs over Spectrum-6 Ethernet using all-copper L1 links. Across rack modules, scale-out is handled by each module’s Spectrum-6 switch over OSFP optical links. 4. How it works CPX must be paired with Vera Rubin NVL72, and Nvidia recommends a 1:1 ratio of CPX to Rubin GPUs. CPX handles prefill and builds the KV cache, which is then transferred to Rubin over Ethernet RDMA for decode. 5. Product positioning: Best performance per dollar for long-context prefill Over 50% of today’s AI inference workload comes from processing input context and building the corresponding KV cache. CPX therefore offers a more flexible, lower-cost way to handle prefill. Each eight-CPX tray has approximately 1.34 TB of HBM4, sufficient for most long-context prefill workloads and associated KV cache requirements.

郭明錤|Ming-Chi Kuo avatar
郭明錤|Ming-Chi Kuo@mingchikuo·Aug 31

正當市場認為 Rubin CPX 已被 Nvidia 從產品藍圖中移除之際,我最新的產業調查顯示,Nvidia 已重啟 Rubin CPX,預計於 1Q27 開始生產。相較舊版,重啟後的 Rubin CPX 擁有更強的預填充(prefill)效能,GPU 規格與機櫃架構也都有明顯改變,足以證明 Nvidia 對 prefill 方案的高度重視。 關鍵改變如下: 1. Rubin CPX GPU 規格 CPX 算力接近 Rubin,單顆 GPU 最高功耗也同為 2,300 W。CPX 記憶體改為 168 GB HBM4,低於 Rubin 的 288 GB HBM4,但高於舊版 CPX 的 128 GB GDDR7。 2. 機櫃設計 新版 CPX 採用獨立的 MGX ETL 機櫃,不再像舊版與 Rubin 共櫃。客戶可依需求配置 64、128、192 或 256 顆 CPX。在 CPX 機櫃中,每 64 顆 CPX 組成一個機櫃模組,每個機櫃模組配置 8 個運算托盤(每個托盤有 8 顆 CPX)與 1 個交換機托盤。 3. Scale-up 與 scale-out NVLink 僅用於同一托盤內 8 顆 CPX GPU 的 scale-up,每顆 CPX 的 NVLink 頻寬為 1–1.5 TB/s(vs. 每顆 Rubin 的 3.6 TB/s)。同一機櫃模組內的托盤間,透過 Spectrum-6 Ethernet(全銅 L1)進行 scale-out;跨機櫃模組時,則由各模組的 Spectrum-6 透過 OSFP 光纖進行 scale-out。 4. 運作方式 CPX 需與 Vera Rubin NVL72 搭配,Nvidia 建議 CPX 與 Rubin 的比例為 1:1。CPX 負責預填充(prefill)並建立 KV cache,再透過 Ethernet RDMA 傳給 Rubin 執行解碼生成(decode)。 5. 產品定位:長上下文 prefill 的最高性價比方案 目前 AI 推論工作量中,超過一半來自處理輸入內容(context)並建立相對應的 KV cache,因此,CPX 以更彈性的部署方式與更低成本承接 prefill。每個運算托盤(有 8 個 CPX)約有 1.34 TB HBM4,足以支援大多數長上下文 prefill 與 KV cache 建立需求。

郭明錤|Ming-Chi Kuo avatar
郭明錤|Ming-Chi Kuo@mingchikuo·Aug 10

A recent media report claims that tight DRAM supply has left TSMC holding around US$1 billion worth of Apple 2nm processor work-in-process (WIP), described in the report as “processor wafers,” that cannot yet be packaged. As supporting evidence, it points to TSMC’s 2Q26 earnings call, where the company said the increase in inventory days was mainly due to the 2nm production ramp. My industry checks suggest that Apple has indeed scaled back its hardware shipment plans this year due to memory shortages. However, Apple plans its processor production at TSMC at least three months in advance, based on the amount of memory expected to be available, rather than having TSMC build large amounts of WIP ahead of time. This does not mean TSMC never builds WIP in advance and holds it in inventory. But if the bottleneck is not at TSMC, building ahead provides little benefit, so Apple would have little reason to pay TSMC extra for it. In other words, tight memory supply is real. But my understanding is that there has been no dramatic scenario in which TSMC first built up US$1 billion of WIP and then had to wait for memory to arrive before packaging could proceed. TSMC and Apple are both known for world-class execution, and such a dramatic development would be unusual given how closely the two companies coordinate their supply chains. Finally, public information alone offers a useful way to assess this claim. TSMC said on its 2Q26 earnings call that the increase in inventory days was mainly due to the 2nm production ramp. However, that inventory cannot be directly attributed to Apple processor WIP for three reasons: 1. TSMC's inventory days typically increase when a new advanced node enters its initial production ramp, so this is not unique to this year. 2. TSMC’s definition of inventory includes not only work-in-process, but also finished goods, raw materials, supplies, and spare parts. 3. Apple is not TSMC’s only 2nm customer this year. Other chip designers, including AMD and MediaTek, are also using 2nm.

郭明錤|Ming-Chi Kuo avatar
郭明錤|Ming-Chi Kuo@mingchikuo·Aug 10

市場近期傳聞,台積電因 DRAM 供應緊張,囤積約 10 億美元的 Apple 2nm 處理器半成品(傳聞稱為處理器晶圓)無法封裝,並以台積電 2Q26 法說提到「庫存天數增加主因是 2nm 量產爬坡」作為佐證。 我的產業調查顯示,Apple 今年的硬體出貨量確實因記憶體短缺而下調,但 Apple 在台積電的處理器訂單,會在至少約 3 個月前依可取得的記憶體數量規劃生產,而不是讓台積電提前大量生產半成品。 這不代表台積電不會提前生產並囤積半成品;但若瓶頸不在台積電,這麼做並無實益,因此,Apple 也沒有太大理由額外付錢要求台積電這樣做。 換言之,記憶體供應緊張是真,但我的理解是,並沒有出現「台積電先囤積 10 億美元半成品,再苦等記憶體到貨封裝」這種戲劇性變化。台積電與 Apple 都是全球執行力頂尖的企業,在供應鏈管理上,雙方緊密合作下也很難出現這種戲劇性的發展。 最後,憑公開資訊也能反思市場傳聞可能性。台積電在 2Q26 法說表示,庫存天數增加主要因 2nm 量產爬坡,但這裡的庫存不能直接視為由 Apple 處理器半成品造成,原因有三: 1. 台積電過往全新先進製程量產爬坡時,庫存天數通常也會增加,並非今年特有現象。 2. 台積電的庫存除半成品外,也包括製成品、原材料、物料與備品。 3. 今年 2nm 客戶不只有 Apple,還包括其他 IC 設計客戶(如 AMD 與聯發科)。