Knowledge article

SemiVision Weekly Intelligence | 2026 July Week 4

更新 2026-08-01 公開知識

本頁目錄

• SemiVision Weekly Intelligence | 2026 July Week 4
  • 中文翻譯與英文原文（雙語）
• SemiVision Weekly Intelligence | 2026 July Week 4
  • Original Article By SemiVision Research [Reading time: 13 mins]

SemiVision Weekly Intelligence | 2026 July Week 4

• Source: SemiVision
• Author: SEMIVISION
• Published: 2026-08-01T11:21:01+08:00
• URL: https://tspasemiconductor.substack.com/p/semivision-weekly-intelligence-2026

中文翻譯與英文原文（雙語）

SemiVision 每週情報 | 2026 年 7 月第 4 週 --- SemiVision Weekly Intelligence | 2026 July Week 4

SemiVision Weekly Intelligence | 2026 July Week 4

SemiVision 每週情報 | 2026 年 7 月第 4 週

Original Article By SemiVision Research [Reading time: 13 mins]

原文作者：SemiVision Research [閱讀時間：13 分鐘]

Aug 01, 2026

∙ Paid∙ 付費

Share分享

The AI semiconductor industry is crossing an important inflection point.

AI 半導體產業正跨越一個重要的轉折點。

Over the past year, the market focused primarily on how many GPUs Nvidia could produce, how many accelerators cloud service providers could secure, and which companies could build the largest AI compute clusters first.

過去一年，市場焦點主要集中在 Nvidia 能生產多少 GPU、雲端服務供應商能取得多少加速器，以及哪些公司能率先建構最大規模的 AI 運算集群。

But as GPU supply gradually expands, a new set of constraints is emerging. Even when companies can obtain the chips, do they have enough HBM, advanced packaging capacity, testing capability, networking equipment, power infrastructure, and cooling systems to turn those chips into fully operational AI factories?

但隨著 GPU 供應逐漸擴張，一組新的瓶頸正在浮現。即使企業能夠取得晶片，它們是否擁有足夠的 HBM 記憶體、先進封裝產能、測試能力、網路設備、電力基礎設施與散熱系統，將這些晶片轉化為全面運作的 AI 工廠？

This week, the latest signals from Samsung, SK hynix, Lam Research, KLA, Teradyne, Microsoft, Meta, and others all point in the same direction:

本週，來自 Samsung、SK hynix、Lam Research、KLA、Teradyne、Microsoft、Meta 等公司的最新訊號皆指向同一方向：

The center of AI competition is shifting from GPU procurement to control and integration of the entire infrastructure supply chain.

AI 競爭的核心正從 GPU 採購轉向對整個基礎設施供應鏈的掌控與整合。

This means the next winners may not simply be the companies with access to the most powerful GPUs. They may be the companies that can secure memory, packaging, equipment, testing, optical interconnects, power, and data center resources—and convert massive capital expenditure into usable computing output and sustainable cash flow.

這意味著，下一階段的贏家可能不再是單純擁有最強大 GPU 的公司，而是那些能夠確保記憶體、封裝、設備、測試、光學互連、電力及資料中心資源，並將龐大資本支出轉化為可用運算產出與永續現金流的企業。

Key Takeaways關鍵要點

1. The primary AI bottleneck is shifting from GPUs to the overall system

2. AI 的主要瓶頸正從 GPU 轉向整體系統

GPUs remain critical, but HBM, CoWoS, testing, networking, optical interconnects, power, and cooling are increasingly determining the actual pace of AI infrastructure deployment.

GPU 仍至關重要，但 HBM、CoWoS、測試、網路、光學互連、電源與散熱，正日益成為決定 AI 基礎設施部署實際進度的關鍵因素。

1. Memory has evolved from a component into a strategic resource

2. 記憶體已從單一元件演變為戰略資源

HBM, server DRAM, and enterprise SSDs remain in tight supply. The market is no longer asking only whether demand exists. It is increasingly asking whether suppliers can deliver on customer schedules.

HBM、伺服器 DRAM 與企業級 SSD 持續供不應求。市場關注的焦點已不再只是需求是否存在，而是供應商能否依照客戶時程如期交貨。

1. Equipment orders suggest the AI investment cycle is not over

2. 設備訂單顯示 AI 投資週期尚未結束

The outlooks from Lam Research, KLA, and Teradyne indicate that demand for wafer-fabrication, process-control, and testing equipment remains strong. The more complex AI chips become, the greater the equipment intensity required for each unit of capacity.

Lam Research、KLA 與 Teradyne 的展望顯示，晶圓製造、製程控制及測試設備的需求依然強勁。隨著 AI 晶片日益複雜，每單位產能所需的設備密集度也隨之提升。

1. Advanced packaging is moving from a single dominant route to a multi-platform competition

2. 先進封裝正從單一主導路徑轉向多平台競爭

CoWoS remains the leading platform, but EMIB, FOPLP, 3D integration, and hybrid bonding are accelerating toward commercialization. The next phase of competition will depend on who can integrate chiplets, memory, substrates, bonding, testing, and thermal management most effectively.

CoWoS 仍是領先平台，但 EMIB、FOPLP、3D 整合與混合鍵結技術正加速邁向商業化。下一階段的競爭將取決於誰能最有效地整合小晶片（chiplets）、記憶體、基板、鍵結、測試與熱管理。

1. The AI supply chain is becoming increasingly regionalized

2. AI 供應鏈正日益區域化

The United States, Japan, and China are all building domestic capabilities across wafer manufacturing, packaging, materials, equipment, and system integration. The global supply chain is shifting away from pure cost optimization toward greater autonomy, security, and resilience.

美國、日本與中國皆正積極建構涵蓋晶圓製造、封裝、材料、設備及系統整合的本土能力。全球供應鏈正從單純的成本優化，轉向追求更高的自主性、安全性與韌性。

1. Capital markets are beginning to demand proof of AI returns

2. 資本市場開始要求驗證 AI 的投資回報

Microsoft is showing that AI can contribute to cloud revenue growth, while Meta’s free-cash-flow pressure demonstrates how capital-intensive AI data centers can be. The key question is shifting from “How much are you investing?” to “When will the investment generate returns?”

微軟展現了 AI 如何帶動雲端營收成長，而 Meta 面臨的自由現金流壓力則凸顯了 AI 資料中心的高度資本密集特性。關鍵問題已從「你投入了多少資金？」轉變為「這項投資何時能產生回報？」

1. Bottleneck capacity is becoming the truly scarce asset

2. 瓶頸產能正成為真正稀缺的資產

The first phase was about securing GPUs. The second phase is about securing HBM, packaging, equipment, testing, and power. The next AI winners will be the companies best positioned to control critical bottlenecks and coordinate the broader supply chain.

第一階段的重點在於取得 GPU。第二階段則轉向確保 HBM 記憶體、先進封裝、製程設備、測試技術以及電力供應。下一波 AI 勝出者，將是那些最能掌握關鍵瓶頸並協調整體供應鏈的企業。

本週文章僅供付費訂閱者閱讀。

What Happened?發生了什麼事？

Why Does It Matter?為何這很重要？

What Happens Next?接下來會發生什麼？

SemiVision InsightSemiVision 洞察

What Happened?發生了什麼事？

The most important developments this week can be organized into five major themes.

本週最重要的發展可歸納為五大主題。

1. Memory Has Become One of AI’s Largest Bottlenecks

2. 記憶體已成為人工智慧最大的瓶頸之一

The latest earnings from Samsung and SK hynix once again confirm that demand for AI servers has not meaningfully weakened.

三星與 SK 海力士的最新財報再次證實，AI 伺服器的需求並未顯著減弱。

On the contrary, HBM, server DRAM, and enterprise SSDs remain among the strongest segments in the memory market.

相反地，HBM、伺服器 DRAM 及企業級 SSD 仍是記憶體市場中表現最強勁的領域。

Samsung stated during its earnings call that hyperscale data centers continue to increase procurement, while the rise of agentic AI is driving demand across both conventional servers and AI servers. Although higher component prices may weigh on smartphones and PCs, incremental server demand is still sufficient to offset part of the weakness in consumer electronics.

三星於財報會議中指出，超大型資料中心持續增加採購量，而代理型 AI（agentic AI）的興起則帶動了傳統伺服器與 AI 伺服器的雙重需求。儘管零組件價格上漲可能對智慧型手機與 PC 造成壓力，但伺服器需求的增量仍足以抵銷部分消費性電子產品的疲軟。

More importantly, Samsung believes that capacity expansion may continue to lag demand, potentially widening the supply-demand gap next year.

更重要的是，三星認為產能擴張可能持續落後於需求，進而導致明年供需缺口進一步擴大。

SK hynix also delivered strong revenue and profit growth. However, the market’s attention is shifting away from whether AI demand exists and toward a more practical question:

SK 海力士亦實現了強勁的營收與獲利成長。然而，市場的關注焦點正從「AI 需求是否存在」轉向一個更實際的問題：

Can HBM capacity, yields, and delivery schedules keep pace with the ramp-up of AI GPUs?

HBM 的產能、良率與交期能否跟上 AI GPU 的放量速度？

This means the market’s evaluation framework has changed.

這意味著市場的評估框架已發生改變。

Previously, the key question was:

過去，關鍵問題是：

Are there enough orders?訂單量是否充足？

Now the question is:現在的問題是：

Can suppliers deliver on time?

供應商能否按時交貨？

HBM is not a standard memory product. It requires advanced DRAM processes, through-silicon vias, die stacking, packaging, testing, and thermal management. A problem in any one of these steps can constrain final shipments.

HBM 並非標準記憶體產品。它需要先進的 DRAM 製程、矽穿孔（TSV）、晶粒堆疊、封裝、測試及熱管理技術。其中任何一個環節出現問題，都可能限制最終出貨量。

The faster GPU demand grows, the more pressure it places on HBM supply.

GPU 需求成長越快，對 HBM 供應造成的壓力就越大。

As AI workloads continue to scale, the memory capacity, bandwidth, and stack height required per GPU are also likely to increase. From HBM3E to HBM4 and beyond, memory is no longer a supporting component attached to the GPU.

隨著 AI 工作負載持續擴展，每顆 GPU 所需的記憶體容量、頻寬與堆疊高度也可能隨之增加。從 HBM3E 到 HBM4 乃至更未來的世代，記憶體已不再是附屬於 GPU 的輔助元件。

It has become a core factor determining system performance and shipment volume.

它已成為決定系統效能與出貨量的核心因素。

7月24日7 月 24 日

1. Wafer-Fabrication Equipment Demand Is Not Cooling

2. 晶圓製造設備需求未見降溫

Judging only by recent technology-stock volatility, some investors may assume that AI capital expenditure is beginning to slow.

僅從近期科技股的波動來看，部分投資人可能會認為 AI 資本支出正開始放緩。

The equipment suppliers are telling a very different story.

設備供應商則講述了截然不同的故事。

This week, Lam Research issued guidance above market expectations, KLA reported record revenue, and Teradyne raised its outlook.

本週，Lam Research 發布的財測高於市場預期，KLA 營收創下歷史新高，而 Teradyne 則上調了其業績展望。

These companies represent three critical areas of semiconductor manufacturing:

這些公司代表了半導體製造的三個關鍵領域：

Lam Research: etch, deposition, and advanced process equipment

Lam Research：蝕刻、沉積及先進製程設備

KLA: process control, metrology, and defect inspection

KLA：製程控制、量測與缺陷檢測

Teradyne: semiconductor and system testing

Teradyne：半導體與系統測試

Their combined message is clear:

他們傳達的訊息非常明確：

Foundries, memory manufacturers, and advanced packaging companies are still placing significant equipment orders.

晶圓代工廠、記憶體製造商與先進封裝業者仍持續下達大量設備訂單。

The more complex AI chips become, the greater the demand for manufacturing equipment.

AI 晶片越趨複雜，對製程設備的需求就越高。

Advanced logic processes require more etch, deposition, and cleaning steps. HBM requires additional TSV, stacking, and packaging processes. Chiplets and 2.5D or 3D integration require increasingly precise metrology, alignment, and defect inspection.

先進邏輯製程需要更多的蝕刻、沉積與清洗步驟；HBM 則需額外的 TSV、堆疊與封裝製程；而 Chiplet 以及 2.5D 或 3D 整合技術，則對量測、對準與缺陷檢測提出更精準的要求。

As semiconductor manufacturing moves from planar scaling toward three-dimensional integration, process steps multiply and the cost of defects rises.

隨著半導體製造從平面微縮轉向三維整合，製程步驟倍增，缺陷所造成的成本也隨之攀升。

A high-end AI accelerator may combine multiple compute dies, HBM stacks, silicon interposers, and complex packaging structures. A failure in one die, one interconnect, or one packaging layer can render the entire module unusable.

高端 AI 加速器可能結合多個運算晶粒、HBM 堆疊、矽中介層以及複雜的封裝結構。只要其中一個晶粒、互連或封裝層發生故障，整個模組便無法使用。

This is why rising AI-chip demand does not benefit lithography alone. It also drives demand for etch, deposition, metrology, inspection, and testing equipment.

正因如此，AI 晶片需求的增长不僅利好微影製程，也帶動了蝕刻、沉積、量測、檢測與測試設備的需求。

The market sees GPU sales.

市場看到的是 GPU 的銷售表現。

Equipment suppliers see the process complexity behind those sales.

設備供應商看到的則是這些銷售背後的製程複雜度。

7月30日7 月 30 日

1. AI Packaging Is Beginning to Diversify Beyond a Single Path

2. AI 封裝開始朝向多元路徑發展

Over the past two years, discussions about AI advanced packaging have largely been synonymous with TSMC’s CoWoS platform.

過去兩年，關於 AI 先進封裝的討論幾乎等同於台積電的 CoWoS 平台。

CoWoS remains the most mature and scalable packaging solution for high-end AI GPUs. However, as AI chips continue to expand in size, power consumption, bandwidth, and chiplet count, no single packaging architecture is likely to serve every application.

CoWoS 仍是高階 AI GPU 最成熟且可擴展的封裝解決方案。然而，隨著 AI 晶片在尺寸、功耗、頻寬及小晶片（chiplet）數量上持續擴張，單一封裝架構恐怕難以滿足所有應用需求。

Several alternative approaches gained more visibility this week, including:

本週有幾種替代方案受到更多關注，包括：

Intel EMIB

Samsung Advanced Packaging

Samsung 先進封裝

Fan-Out Panel-Level Packaging

扇出型面板級封裝

Hybrid Bonding混合鍵合

3D Chiplet Integration3D 小晶片整合

The sharp increase in orders at hybrid-bonding equipment supplier Besi is particularly important. It suggests that hybrid bonding is moving from research and technical qualification toward commercial deployment.

混合鍵合設備供應商 Besi 的訂單大幅增長尤為重要。這顯示混合鍵合技術正從研發與技術驗證階段，邁向商業化部署。

Hybrid bonding can reduce interconnect pitch, increase I/O density, and lower transmission power. As AI processors integrate more compute dies, memory, and functional modules, conventional micro-bump interconnects may increasingly encounter density and power limitations.

混合鍵合可縮減互連間距、提升 I/O 密度，並降低傳輸功耗。隨著 AI 處理器整合更多運算晶粒、記憶體及功能模組，傳統的微凸塊互連技術可能日益面臨密度與功耗的限制。

The future of AI packaging will therefore not be defined by a single standard.

因此，AI 封裝的未來將不會由單一標準所定義。

CoWoS may continue to dominate high-end GPU and HBM integration, while EMIB, Foveros, I-Cube, X-Cube, FOPLP, and hybrid bonding may each gain traction depending on product architecture, package size, cost, and customer requirements.

CoWoS 可能持續主導高階 GPU 與 HBM 的整合，而 EMIB、Foveros、I-Cube、X-Cube、FOPLP 及混合鍵結等技術，則將依據產品架構、封裝尺寸、成本及客戶需求等因素，各自獲得市場青睞。

The real competition is not simply about owning one packaging technology.

真正的競爭並非僅在於掌握單一項封裝技術。

It is about integrating:而在於整合：

wafer fabrication, chiplet design, interposers, substrates, bonding, testing, thermal management, and production yield.

晶圓製造、小晶片設計、中介層、基板、接合、測試、熱管理與生產良率。

Packaging is no longer just a back-end step performed after the chip has been manufactured.

封裝已不再只是晶片製造完成後的後段製程。

It is becoming part of the AI-chip architecture itself.

它正成為 AI 晶片架構本身的一部分。

1. The United States, Japan, and China Are All Building Regional AI Supply Chains

2. 美國、日本與中國皆在打造區域性 AI 供應鏈

For the past three decades, the semiconductor industry has focused on maximizing global efficiency.

過去三十年來，半導體產業一直專注於最大化全球效率。

Design may take place in the United States, wafer fabrication in Taiwan or South Korea, equipment may come from the United States, Japan, or Europe, materials may be supplied by Japanese companies, and assembly and testing may be distributed across Taiwan, China, and Southeast Asia.

設計可能在美國進行，晶圓製造在台灣或南韓，設備可能來自美國、日本或歐洲，材料可能由日本企業供應，而封裝與測試則可能分散於台灣、中國及東南亞。

That logic is now changing under rising geopolitical pressure, export controls, and supply-chain risk.

然而，在地緣政治壓力升高、出口管制加劇以及供應鏈風險增加的背景下，此一邏輯正發生改變。

If the key phrase of the previous era was:

如果上一個時代的關鍵詞是：

Global Supply Chain全球供應鏈

The key phrase of the next era may be:

下一個時代的關鍵詞可能是：

Regional Supply Chain區域供應鏈

This week, the United States, Japan, and China all signaled further efforts to strengthen domestic or regional semiconductor ecosystems.

本週，美國、日本與中國皆釋出訊號，表示將進一步強化本土或區域半導體生態系。

In the United States, Intel is increasing capital expenditure, Amkor is expanding advanced packaging capacity in Arizona, and Wistron is developing AI server and computing-system production in Texas.

在美國，英特爾（Intel）正增加資本支出，安靠（Amkor）於亞利桑那州擴充先進封裝產能，而緯創（Wistron）則在德克薩斯州發展 AI 伺服器與運算系統的生產。

The US objective is not merely to restore wafer-fabrication capacity. It is to build a more complete AI supply chain spanning semiconductors, advanced packaging, server assembly, and data-center deployment.

美國的目標不僅是恢復晶圓製造產能，更是要打造一條涵蓋半導體、先進封裝、伺服器組裝到資料中心部署的完整 AI 供應鏈。

In Japan, Tower Semiconductor is expanding its silicon-photonics activities, while the semiconductor cluster in Kumamoto continues to attract equipment, materials, engineering-service, and talent resources.

在日本，Tower Semiconductor 正擴大其矽光子業務，而熊本半導體聚落持續吸引設備、材料、工程服務及人才資源。

Japan’s strength may not lie in leading-edge logic alone. Its advantages include materials, equipment, sensors, power semiconductors, silicon photonics, and precision manufacturing. These capabilities are equally essential to AI infrastructure.

日本的優勢或許不僅在於先進製程邏輯晶片，其強項還涵蓋材料、設備、感測器、功率半導體、矽光子技術與精密製造。這些能力對人工智慧基礎設施同樣至關重要。

In China, CXMT’s listing and memory expansion have attracted significant attention, while localization efforts continue across semiconductor materials, equipment, EDA, and advanced packaging.

在中國，長鑫儲存（CXMT）的上市與記憶體產能擴張引發高度關注，同時半導體材料、設備、電子設計自動化（EDA）及先進封裝等領域的在地化努力亦持續推進。

7月28日7 月 28 日

The strategies differ, but the objective is the same:

策略雖有不同，但目標一致：

Reduce dependence on any single country, company, or technology platform.

降低對任何單一國家、公司或技術平台的依賴。

The future semiconductor industry may no longer be organized around one globally optimized supply chain.

未來的半導體產業可能不再圍繞單一全球最佳化的供應鏈來組織。

Instead, it may evolve into several regional ecosystems that are more expensive, but more resilient and strategically autonomous.

取而代之的是，它可能演變成數個區域性生態系統，這些系統雖然成本較高，卻更具韌性與戰略自主性。

7月28日7 月 28 日

Why Does It Matter?

為何這至關重要？

When all of this week’s developments are viewed together, it becomes clear that the AI semiconductor industry is moving from its first phase into its second.

當本週所有發展綜觀來看，很明顯人工智慧半導體產業正從第一階段邁向第二階段。

The central question in the first phase was:

第一階段的核心問題是：

Who can secure the GPUs?

誰能取得 GPU？

The central question in the second phase is:

第二階段的核心問題是：

Who can build an entire AI factory and keep it operating reliably at scale?

誰能夠打造整座 AI 工廠，並在大規模運作下維持其穩定性？

GPUs remain the core of AI computing, but they are only one component within a much larger system.

GPU 仍是 AI 運算的核心，但它僅是更大系統中的一個組件。

The constraints on AI expansion are increasingly shifting toward multiple supply-chain bottlenecks:

AI 擴張的限制正日益轉向多重的供應鏈瓶頸：

HBM and server memoryHBM 與伺服器記憶體

CoWoS and advanced packaging

CoWoS 與先進封裝

Wafer fabrication and equipment

晶圓製造與設備

Testing time and production yield

測試時間與生產良率

High-speed networking and switches

高速網路與交換器

Optical interconnect光學互連

Power delivery電源供應

Cooling散熱

Electricity supply電力供應

These components are deeply interconnected.

這些元件緊密相連。

A GPU without HBM cannot ship.

沒有 HBM 的 GPU 無法出貨。

A chip without advanced packaging cannot integrate the GPU and memory.

沒有先進封裝的晶片，就無法整合 GPU 與記憶體。

A server without sufficient power and cooling cannot operate reliably.

若伺服器缺乏足夠的電力與散熱，便無法可靠運作。

A compute cluster without high-speed networking and optical interconnect cannot scale efficiently.

若運算叢集缺少高速網路與光學互連，就無法有效擴展。

AI-factory competition is therefore no longer about the performance of a single chip.

因此，AI 工廠的競爭已不再取決於單一晶片的效能。

It is about total system efficiency and supply-chain coordination.

這關乎整體系統效率與供應鏈協同。

7月9日7 月 9 日

Who Benefits?誰是受益者？

As AI enters its second phase, the beneficiaries are no longer limited to GPU designers.

隨著人工智慧進入第二階段，受益者已不再限於 GPU 設計廠商。

Value is increasingly spreading toward companies that control critical, difficult-to-replace bottlenecks.

價值正日益流向那些掌握關鍵且難以取代瓶頸的企業。

First Layer: Semiconductor Equipment

第一層：半導體設備

Key beneficiaries include ASML, Lam Research, Applied Materials, KLA, and Tokyo Electron.

主要受益者包括 ASML、Lam Research、Applied Materials、KLA 以及 Tokyo Electron。

As AI chips and HBM become more complex, equipment intensity increases. Even if wafer capacity does not grow proportionally, the number of process steps and the amount of equipment required per unit of capacity may continue to rise.

隨著 AI 晶片與 HBM 日益複雜，設備密集度隨之提升。即使晶圓產能未成比例成長，每單位產能所需的製程步驟數量與設備用量仍可能持續增加。

Second Layer: Memory第二層：記憶體

Key beneficiaries include SK hynix, Samsung, and Micron.

主要受益者包括 SK hynix、Samsung 和 Micron。

HBM has become a major determinant of AI-accelerator performance, power efficiency, and shipment volume. Server DRAM and enterprise SSDs are also benefiting from the expansion of AI servers and data centers.

HBM 已成為決定 AI 加速器效能、電源效率及出貨量的關鍵因素。伺服器 DRAM 與企業級 SSD 亦因 AI 伺服器與資料中心的擴張而受惠。

Third Layer: Advanced Packaging

第三層：先進封裝

Key beneficiaries include TSMC, Amkor, ASE, and Besi.

主要受益者包括台積電、安靠（Amkor）、日月光（ASE）和 Besi。

As chiplets, 2.5D integration, 3D integration, and hybrid bonding gain adoption, the technical barriers and capital intensity of advanced packaging are rising. Packaging may also account for a larger share of total semiconductor value.

隨著小晶片（Chiplets）、2.5D 整合、3D 整合及混合鍵合技術的採用日益普及，先進封裝的技術門檻與資本密集度正不斷攀升。封裝環節在半導體總價值中的佔比也可能隨之提高。

Fourth Layer: Testing第四層：測試

Key beneficiaries include Teradyne, Advantest, and Chroma.

主要受益者包括 Teradyne、Advantest 和 Chroma。

AI chips are expensive and structurally complex. Customers cannot afford to discover defective dies only after final packaging. Known-good-die testing, wafer-level testing, system-level testing, and burn-in will therefore become increasingly important.

AI 晶片昂貴且結構複雜，客戶無法承擔僅在最終封裝後才發現不良晶粒的風險。因此，已知良好晶粒（Known-Good-Die）測試、晶圓級測試、系統級測試以及老化測試將變得日益重要。

Fifth Layer: Power and Cooling

第五層：電源與散熱

Key beneficiaries include Delta Electronics, Vertiv, and Schneider Electric.

主要受益者包括台達電、Vertiv 與施耐德電機。

As AI-server rack power density rises, data-center power architecture is shifting from conventional low-voltage AC systems toward higher-voltage, higher-efficiency DC architectures and liquid-cooling systems.

隨著 AI 伺服器機架功率密度提升，資料中心電源架構正從傳統低壓交流系統，轉向更高電壓、更高效率的直流架構與液冷系統。

GPUs can continue to multiply.

GPU 數量可持續倍增。

But without electricity, cooling, and grid capacity, a GPU is simply an expensive piece of silicon.

但若缺乏電力、散熱與電網容量，GPU 不過是一塊昂貴的矽晶圓。

6月27日6 月 27 日

What Happens Next?接下來會發生什麼？

The market should no longer focus only on GPU shipment numbers.

市場不應再只關注 GPU 的出貨數量。

Five deeper variables will determine what happens next.

五項更深層的變數將決定後續發展。

1. Will Microsoft, Meta, Amazon, and Google Continue Increasing AI CapEx?

2. 微軟、Meta、亞馬遜和谷歌是否會持續增加 AI 資本支出？

As long as hyperscalers continue raising data-center and AI-infrastructure budgets, the growth cycle for equipment, memory, packaging, networking, and power infrastructure is unlikely to be over.

只要超大型雲端服務供應商持續提高資料中心與 AI 基礎設施的預算，設備、記憶體、封裝、網路及電力基礎設施的成長週期便不太可能結束。

However, investors will increasingly examine the relationship between capital spending and revenue growth.

然而，投資人將愈發關注資本支出與營收成長之間的關係。

1. Will HBM Remain in Short Supply?

2. HBM 是否仍將供不應求？

If Samsung, SK hynix, and Micron continue to report tight capacity, it would indicate that AI-server demand remains above supply.

若 Samsung、SK hynix 與 Micron 持續回報產能吃緊，這將顯示 AI 伺服器的需求仍高於供給。

Beyond total capacity, the market should also monitor HBM4 yields, base-die processes, packaging capacity, and customer qualification progress.

除了總產能之外，市場也應密切關注 HBM4 的良率、基礎晶粒製程、封裝產能以及客戶認證進度。

1. Can CoWoS Capacity Catch Up With Demand?

2. CoWoS 產能能否趕上需求？

If the CoWoS shortage gradually eases, the next bottleneck may shift toward optical interconnect, networking, power, cooling, or testing.

若 CoWoS 產能短缺逐漸緩解，下一個瓶頸可能會轉向光學互連、網路、電源、散熱或測試環節。

Supply-chain bottlenecks rarely disappear.

供應鏈瓶頸很少會消失。

They simply move.它們只是轉移了位置。

7月13日7 月 13 日

1. Will Hybrid Bonding Enter Large-Scale Production?

2. 混合鍵合技術是否將進入大規模量產？

If more logic, memory, and chiplet customers begin adopting hybrid bonding, the next generation of 3D AI packaging will formally enter the commercial phase.

若更多邏輯晶片、記憶體及小晶片（chiplet）客戶開始採用混合鍵合技術，下一代 3D AI 封裝將正式邁入商業化階段。

Equipment orders, customer qualifications, and manufacturing yields will be the most important indicators.

設備訂單、客戶認證與製造良率將成為最重要的指標。

2025年7月11日2025 年 7 月 11 日

1. Will China Continue Increasing Self-Sufficiency in Materials, Equipment, and Memory?

2. 中國是否會持續提升材料、設備與記憶體的自給自足能力？

CXMT, semiconductor materials, domestic equipment, EDA, advanced packaging, and testing capabilities all deserve continued attention.

長鑫儲存（CXMT）、半導體材料、國產設備、EDA、先進封裝及測試能力，皆值得持續關注。

China is not building a single semiconductor product.

中國並非在打造單一半導體產品。

It is building an ecosystem spanning design, manufacturing, materials, equipment, packaging, and testing.

它正在建構一個涵蓋設計、製造、材料、設備、封裝與測試的完整生態系。

Its near-term cost and efficiency may remain inferior to those of the globalized supply chain, but its long-term strategic goal is not maximum efficiency.

其短期成本與效率或許仍遜於全球化供應鏈，但其長期戰略目標並非追求極致效率。

It is supply-chain autonomy and the ability to continue operating under external restrictions.

這是供應鏈自主以及在外部限制下持續營運的能力。

7月20日7 月 20 日

SemiVision Insight

The market still tends to understand AI through a simple equation:

市場仍傾向透過一個簡單的方程式來理解人工智慧：

NVIDIA = AI

Industry insiders already see a much broader picture.

產業界人士早已看見更宏觀的全貌。

Today’s AI competition is no longer fundamentally a competition between individual GPUs.

當今的 AI 競爭，本質上已不再是單一 GPU 之間的較量。

It is a competition between entire AI factories.

這是一場整體 AI 工廠之間的競爭。

The winners over the next five years will not be determined only by who can design the fastest chip.

未來五年的勝出者，將不只取決於誰能設計出最快的晶片。

They will be determined by who can secure stable access to memory, advanced packaging, equipment, testing, networking, optical interconnect, power, cooling, and global supply-chain resources.

而是取決於誰能確保穩定取得記憶體、先進封裝、設備、測試、網路、光學互連、電力、散熱以及全球供應鏈資源。

In the first phase, the market competed for GPUs.

在第一階段，市場競爭的焦點是 GPU。

In the second phase, it is competing for bottleneck capacity.

在第二階段，競爭轉向瓶頸產能。

In the third phase, the competition may shift toward who can convert computing, memory, networking, and electricity into the greatest volume of usable AI tokens at the lowest total cost of ownership.

在第三階段，競爭可能轉向誰能以最低的總擁有成本（TCO），將運算、記憶體、網路和電力轉化為最大量的可用 AI token。

This is the part that public information often fails to explain fully, but that industry participants increasingly understand:

這部分是公開資訊往往未能充分解釋，但產業參與者日益理解的關鍵：

GPUs have not become less important, but they are no longer the only scarce resource. The truly scarce asset is the entire AI infrastructure ecosystem required to manufacture, deploy, and operate GPUs at scale.

GPU 的重要性並未降低，但它已不再是唯一的稀缺資源。真正稀缺的資產是能夠大規模製造、部署與運作 GPU 所需的完整 AI 基礎設施生態系。

The next generation of AI winners may not be the companies that own the most GPUs.

下一代 AI 的贏家，可能並非擁有最多 GPU 的公司。

They may be the companies that control the most critical bottlenecks, coordinate the broadest supply-chain resources, and convert capital expenditure into computing output and cash flow most effectively.

它們可能是那些掌握最關鍵瓶頸、協調最廣泛供應鏈資源，並能最有效將資本支出轉化為運算產出與現金流的公司。

Thanks for reading SEMIVISION @_@! This post is public so feel free to share it.

感謝閱讀 SEMIVISION @_@！本文為公開文章，歡迎自由分享。

Share分享

Previous上一頁Next下一則