報告_MS_AI供應鏈_20260810

PDF 原檔：報告_MS_AI供應鏈_20260810_original.pdf

內容說明

本篇為 MS「Idea」短篇快評（記憶體短缺下的 GPU 架構最佳化），trimmed 版僅含開頭 HBM de-spec／Kyber 延後兩段；完整版另含 CoWoS 產能分配、HBM/晶圓 TAM、ASIC-TPU 進度、Blackwell/Rubin 晶片與機櫃出貨、AI 半導體股價倍數等豐富附表——多數數字與既有庫內報告_MS_AI供應鏈CoWoS分配_20260708、分析_AI_capex與電力瓶頸_香港投資人回饋_MS_20260809 已收錄內容高度重疊（同分析師團隊延續追蹤），本次 ingest 已交叉核對、僅將淨新增內容（HBM 多 SKU 分層、Kyber 8/10 最新確認、中國 NPO SerDes 技術原因、KYEC Google 營收占比、TSMC AI 營收拆分 CAGR）寫入對應 lib 頁，重複數字不重寫。

圖片清單（已驗證 2026-08-11）

檔名	size	分類	親眼所見內容
`260810_ms_ai-supply-chain_001.png`	65KB	裝飾·banner	"Asia Summer School 2026" 廣告橫幅，藍色泳池背景＋信封圖示
`260810_ms_ai-supply-chain_002.png`	25KB	真資料圖	圓餅圖：2027 GPU/ASIC 總耗電量佔比，NVIDIA 16、AMD 7、Google 9、AWS 2、Microsoft 1、Meta 1、Others 1（合計對應 Exhibit 2，38GW）
`260810_ms_ai-supply-chain_003.png`	162KB	真資料圖	表格圖：AI chip vendor 別 HBM 需求明細（NVIDIA/AMD/Google/AWS/Microsoft/Meta 各產品之 CoWoS 配額、HBM 顆數/容量/世代/供應商），合計 Total HBM demand 48,618 mn Gb（Exhibit 3）
`260810_ms_ai-supply-chain_004.png`	30KB	真資料圖	長條圖：全球 CoWoS 產能分配（k wafers）2026e vs 2027e，依 NVIDIA/Broadcom/AMD/Xilinx/MediaTek/AWS-Annapurna/AWS-Alchip/Marvell/GUC/Cisco 分色（Exhibit 5）
`260810_ms_ai-supply-chain_005.png`	31KB	真資料圖	堆疊長條圖「CoWoS Supply Capacity Breakdown (By Year End)」：TSMC vs Non-TSMC（Amkor/UMC/ASE）2023-2027e，2027e 合計 280kwpm（TSMC 200＋Non-TSMC 80）（Exhibit 8）
`260810_ms_ai-supply-chain_006.png`	19KB	真資料圖	堆疊長條圖「Global CoWoS capacity demand by key customer」2023-2027e，依客戶別分色（Exhibit 9）
`260810_ms_ai-supply-chain_007.png`	31KB	真資料圖	雙軸線圖：Blackwell/Rubin 季出貨量（mn units）vs NVL72 機櫃數（k racks），標示 10k／15k NVL72 節點（Exhibit 12，舊版）
`260810_ms_ai-supply-chain_008.png`	35KB	真資料圖	雙軸線圖：Blackwell 晶片量 vs GB NVL72＋HGX equivalent 機櫃，標示 13k／19-20k NVL72 equivalent 節點（Exhibit 13，新版）
`260810_ms_ai-supply-chain_009.png`	175KB	真資料圖	雙軸線圖：Rubin 晶片季出貨量 vs NVL72 機櫃數，標示 3k／12-13k NVL72 節點，橫軸 2Q26e-4Q27e（Exhibit 14）
`260810_ms_ai-supply-chain_010.png`	75KB	真資料圖	折線圖：AI 半導體 P/E 倍數趨勢（GP GPU／NVIDIA、Alternative AI Semis、AI Semi Enablers 三線），2022-11~2026-08（Exhibit 15）
`260810_ms_ai-supply-chain_011.png`	83KB	真資料圖	長條＋折線圖：NVIDIA+AMD 資料中心/HPC 半導體營收（US$mn）2014Q1-2027Q3e，疊加 YoY 成長率折線（Exhibit 16）
`260810_ms_ai-supply-chain_012.png`	45KB	真資料圖	堆疊長條圖「TSMC AI revenue breakdown」：General-purpose AI／Custom AI chips (ASICs)／CoWoS-wafer test／AI server CPU，2021-2029e，標示 2024-2029e CAGR 60%（Exhibit 17）
`260810_ms_ai-supply-chain_013.png`	39KB	真資料圖	折線圖：AI GPU H100 每小時租金（GCP H100 A3-HIGH vs AWS H100 p5.48xlarge），2024/03-2026
`260810_ms_ai-supply-chain_014.png`	34KB	真資料圖	折線圖：16x Trainium（trn1n.32xlarge）每小時租金，2024/03-2025/07，中段有明顯尖峰
`260810_ms_ai-supply-chain_015.png`	54KB	真資料圖	折線圖：NVIDIA 4090／5090D 顯卡淘寶經銷價（人民幣）2024/07-2026/08，另疊加官方標價虛線

分類三選一：真資料圖 / 裝飾·logo·banner / 文字卡。<40KB 者本次仍全數逐張 Read 驗證（非預設略過）。

原始內容

M August 10, 2026 09:00 PM GMT

Asia-Pacific Technology | Asia Pacific

AI Supply Chain: GPU Architecture Optimization Given Memory Shortage and Supernode Challenge

Our checks indicate Rubin Ultra may adopt tiered HBM spec to balance supply, cost & workload requirements; shorter testing times could ease near-term capacity constraints.

GPU: Hardware Configuration Optimization

1. HBM de-specification - investors cross-checking whether this is true:

We think Nvidia's Rubin Ultra will feature several SKUs with high-end version adopting HBM4e 8Hi lower-end versions using HBM4 12Hi or 8Hi. Based on discussions with memory coverage analysts, we see this reflecting: 1) HBM capacity shortages, 2) cost pressures, and 3) optimization for different workloads based on memory capacity requirements.
Nvidia should make final decision before end-3Q26. We continue to gather feedback on the performance impact of changes to HBM specifications. Our checks suggest HBM de-specification will have a more meaningful impact on decoding than pre-fill workloads, particularly for long-context-window LLMs.
However, reducing memory content per chip may increase overall chip shipments, boosting volume-driven suppliers. We do not rule out major ASIC projects adopting different HBM configurations for similar reasons.

2. Kyber server rack delays with no concrete timeline, while inter-rack CPO connectivity remains Nvidia's preferred direction:

We believe Nvidia's (analyst: Joe Moore) long-term vision for AI server racks is maximizing compute capacity/GPU density and rack-level performance, with GPUs interconnected through scale-up architectures for LLM training & large-scale inference. Since 2023, we have seen the transition from Hopper HGX (8-GPU, or NVL8) systems to Blackwell NVL72.
To further increase GPU density per rack, Nvidia first disclosed its Kyber blade server design at GTC 2025. However, given ongoing PCB and thermal challenges, first-generation Rubin Ultra server racks will likely continue to use the Oberon solution for NVL72 while adopting CPO/NPO technologies to enable NVL576-scale deployments.
We remain optimistic about Nvidia's adoption of CPO as a key differentiator in scale-up performance versus China's supernode architectures (see below).

Idea

Morgan Stanley Taiwan Limited+ Charlie Chan Equity Analyst Charlie.Chan@morganstanley.com	+886 2 2730-1725
Daniel Yen, CFA Equity Analyst Daniel.Yen@morganstanley.com Morgan Stanley Asia Limited+	+886 2 2730-2863
Daisy Dai, CFA Equity Analyst Daisy.Dai@morganstanley.com	+852 2848-7310
Morgan Stanley Taiwan Limited+ Tiffany Yeh Equity Analyst Tiffany.Yeh@morganstanley.com	+886 2 7712-3032
Lucas Wang Research Associate Lucas.Wang@morganstanley.com Morgan Stanley Asia Limited+	+886 2 2730-2875
Ethan Jia Research Associate Ethan.Jia@morganstanley.com	+852 3963-2287

Morgan Stanley does and seeks to do business with companies covered in Morgan Stanley Research. As a result, investors should be aware that the firm may have a conflict of interest that could affect the objectivity of Morgan Stanley Research. Investors should consider Morgan Stanley Research as only a single factor in making their investment decision.