跳到主要內容

TSMC Is Building an "EMIB-Like" Package — Why Advanced Packaging Is the New Battleground for AI Chips

One-line takeaway: As transistor scaling slows, stitching multiple chips together with high-speed bridges has become the key to AI performance — and TSMC's "EMIB-like" packaging push with substrate maker Kinsus is a direct challenge to Intel's proprietary packaging moat.
TSMC enters Intel's territory — the silicon bridge war begins

The hottest topic in semiconductors right now isn't which chip got smaller — it's a far more physical question: if chips can't keep shrinking, how do they keep getting faster? The answer is advanced packaging. And this time, TSMC is taking aim at Intel's signature EMIB technology.

What is "EMIB-like" packaging, and why is TSMC chasing Intel's tech?

EMIB (Embedded Multi-die Interconnect Bridge) is Intel's flagship packaging approach. The idea: embed a tiny silicon bridge into the substrate, connecting only the chip regions that need high-speed communication — no need to replace the entire substrate with an expensive silicon interposer.

For years, EMIB has been the counterweight to TSMC's dominant CoWoS. Now TSMC is reportedly developing an "EMIB-like" packaging technology with Kinsus, a major ABF substrate maker, to support larger multi-die AI chips and reduce reliance on capacity-constrained CoWoS — stepping directly into Intel's moat.

The hidden bottleneck in AI chip shipping — advanced packaging is what actually gates supply

CoWoS, InFO, EMIB: three advanced packaging approaches, explained

  • CoWoS (TSMC): Logic chips and HBM memory sit side-by-side on a silicon interposer. The interposer is large and expensive, but virtually every AI accelerator depends on it — capacity has been "maxed out" for a long time.
  • InFO (TSMC): Skips the silicon interposer for a cheaper fan-out substrate. Lower cost, used for applications that don't need HBM.
  • EMIB (Intel): No full interposer — small silicon bridges are embedded only where needed. Flexible and relatively cost-controlled, but the yield and supply-chain ecosystem are Intel-led.

In plain terms: CoWoS is like tiling the whole floor with marble; EMIB is like laying premium tiles only on the corridors — cheaper, but technically harder. TSMC's "EMIB-like" version aims to deliver EMIB's flexibility using its own substrate and supply-chain muscle, without paying Intel's patent toll.

EMIB-T supply-chain economics — Intel is building the next packaging ecosystem

Why this move ripples through the global AI supply chain

The market reacted immediately: TSMC's ADR jumped more than 7% on the news, and analysts flagged multiple packaging and substrate suppliers as beneficiaries. Three layers of logic explain it:

1. Capacity bottleneck: AI chip demand has CoWoS running at full capacity, with customers queuing for months. A second "EMIB-like" technology route is another shipping lane.

2. Patents and ecosystem: Intel's EMIB ecosystem has attracted interest from Nvidia, and Google reportedly ordered over 3 million custom AI processors from Intel for 2028. If TSMC can offer an equivalent with its own technology, customers stay inside its ecosystem.

3. Supply-chain spread: Advanced packaging is no longer just a foundry game — substrate, packaging, and materials players have all become strategic links. ABF substrates are now described as "strategic materials."

Investor perspective: real beneficiary vs. trend-chasing hype

When a technology trend heats up, markets always produce a wave of "theme stocks." Three questions help separate fact from fiction: Is this company actually in the advanced-packaging supply chain? Are orders confirmed or merely "possible"? Does the technology barrier create a real moat? For substrate and packaging firms, profits ultimately depend on yield ramp and customer qualification — not on headlines.

The advanced packaging arms race has only just begun. TSMC challenging Intel, Chinese players pushing glass substrates — the global supply-chain map is being redrawn. Whoever can ramp volume production and lock in big customers wins the next round of the AI chip war.

FAQ

Q1: What's the real difference between CoWoS and EMIB?

CoWoS uses a full silicon interposer to connect chips and HBM — large and expensive. EMIB embeds small silicon bridges only where high-speed links are needed — cheaper but technically complex. Both are multi-die approaches that break the single-chip area limit.

Q2: Why does advanced packaging matter more as Moore's Law slows?

Transistor scaling gets exponentially more expensive. Manufacturers instead package multiple mature-process chips together to boost overall performance — advanced packaging is the new path to performance growth.

Q3: Who is Kinsus, and why did ABF substrates suddenly become strategic?

Kinsus is a major ABF substrate maker. Advanced packaging and high-end chips both need ABF substrates as their "foundation" — layer count and precision directly determine packaging yield, turning substrates into a critical AI supply-chain link.

Q4: Will "EMIB-like" packaging replace CoWoS?

Not in the short or medium term — CoWoS remains the mainstream for AI accelerators, with demand far exceeding supply. The EMIB-like route is a second path for capacity and cost flexibility, giving TSMC a fuller menu for different customer needs.

Q5: How should investors approach this trend?

Look for orders actually landing in financials, then judge technology barriers. Advanced packaging is a long-term structural trend, but yield ramps take time — short-term rallies often run ahead of fundamentals, so check valuations before chasing.

留言

這個網誌中的熱門文章

Why Is NVIDIA Spending Billions to Buy Up America's "Dark Fiber"?

One-line conclusion: NVIDIA is reportedly spending $5–10 billion to acquire long-haul "dark fiber" networks across the United States, signaling that the AI infrastructure race is shifting from raw compute power to the networks that connect it. NVIDIA is reportedly acquiring long-haul "dark fiber" networks across the United States, with total capacity estimated at 7.6 Pbps and a price tag between $5 billion and $10 billion. The news sent optical communications stocks surging globally: Taiwan's optical module makers jumped on July 22, and three more hit the daily limit on July 23. Many now read this as the moment the AI arms race moved from "who has more GPUs" to "who owns the network." What Is Dark Fiber, and Why Buy Instead of Lease? Dark fiber refers to fiber-optic cable that has already been laid but has no transmission equipment installed and carries no optical signal . The fiber cores sit "dark" and dormant, waiting to...

What Is x402? How HTTP 402 Is Making Peer-to-Peer Web Payments a Reality

One-line conclusion: x402 is an open payment standard launched by Coinbase in 2025 that resurrects the long-dormant HTTP 402 "Payment Required" status code, enabling websites to complete peer-to-peer cryptocurrency payments directly within the HTTP request-response cycle — no accounts, no intermediaries, no payment gateways required. HTTP 402: The Status Code That Waited 25 Years for Its Moment If you know your HTTP status codes, you're probably familiar with 200 (OK), 404 (Not Found), and 500 (Server Error). But did you know there's a 402 status code? Back in the 1990s, the HTTP/1.0 specification reserved 402 Payment Required for digital cash and micropayment scenarios — a way for servers to tell clients: "Pay up before you access this resource." For 25 years, it remained virtually unused. The reason? Infrastructure wasn't ready. Payments were slow, expensive, and required accounts, sessions, and intermediaries. Now, that forgotten status c...

Grok 45

title: "Grok 4.5 正式登場!1.5T 參數、比 GPT-5.5 便宜 60%,xAI 如何改寫 AI 戰場格局?" date: 2026-07-09 labels: tech, AI, Grok, xAI, LLM, benchmark, coding 一句話結論: Grok 4.5 用 GPT-5.5 四成的價格,做到九成的實力——AI 戰場格局正在被改寫。 2026 年 7 月 8 日,xAI(SpaceXAI)正式發布 Grok 4.5——這是該公司首個專門為 編碼與 AI 代理(Agent) 訓練的大型語言模型。Elon Musk 在 X 平台發文後短短數小時內獲得超過 4.7 萬個讚、520 萬次觀看,話題迅速引爆科技圈。 但這次 Grok 4.5 的真正殺手鐧,不是 benchmark 分數有多高,而是 同等級智慧下,成本只有競爭對手的一小部分 。根據 Artificial Analysis 獨立評測,Grok 4.5 在 Intelligence Index 排名第四(僅次於 Fable 5、GPT-5.5 和 Opus 4.8),但每任務成本僅 $0.31——是 Opus 4.8 的五分之一、GPT-5.5 的三分之一。 這篇文章將深入拆解 Grok 4.5 的技術規格、benchmark 數據的真實意義,以及它對開發者和投資者意味著什麼。 Grok 4.5 是什麼?為什麼重要? Grok 4.5 是 xAI(SpaceXAI)推出的最新旗艦模型,基於 1.5 兆參數的 V9 架構 (是前代 Grok 4.3 的三倍大),並與 Cursor(知名的 AI 編碼助手)聯合訓練。這是 xAI 第一次明確將模型定位為 「編碼與 Agent」專用 ,而非通用對話模型。 關鍵規格一覽: 參數規模: 1.5 兆(1.5T),V9 架構 上下文窗口: 500K tokens(較 Grok 4.3 的 1M 有所縮減,但保留了可配置推理和視覺輸入) 定價: $2 / 百萬輸入 tokens,$6 / 百萬輸出 tokens 推理速度: 約 80 tokens/秒 每任務成本: $0.31(Intelligence Index)、$2.49(Coding Agent I...