來源TechNode•較早收集於 65h
TSMC 3nm產能吃緊,預計2026下半年價格上漲15%

💡3nm成本上漲將直接影響未來AI硬體的單位經濟效益與GPU供應狀況。
⚡ 30 秒速覽
有什麼變化
TSMC第二季3nm月產能已達160,000至175,000片晶圓。
為什麼重要
代工成本上升將可能增加下一代AI加速器的物料清單成本,迫使AI硬體新創公司調整其定價或募資策略。
下一步行動
如果您計劃在2026年於3nm節點進行新AI晶片的流片(tape out),請將15%的成本漲幅納入您的硬體藍圖規劃中。
誰應關注:Founders & Product Leaders
關鍵要點
- •TSMC第二季3nm月產能已達160,000至175,000片晶圓。
- •對AI相關晶片的強勁需求持續超過供應量。
- •預計2026年下半年3nm代工服務價格將調漲15%。
🧠 深度解析
背景與延伸:來自公開資料,非原文內容。引用 31 個來源。
🔑 增強重點摘要
- •TSMC's 3nm capacity is reportedly booked through 2028, indicating a sustained period of high demand and limited supply for its most advanced nodes.
- •Major customers like Nvidia have been forced to adjust their production strategies, reportedly scaling back orders for their next-generation 3nm AI chips (Rubin Ultra) and increasing reliance on 4nm chips (Blackwell) due to TSMC's 3nm capacity constraints.
- •TSMC's 3nm process (N3) continues to utilize FinFET transistor technology, distinguishing it from Samsung's 3nm process which adopted Gate-All-Around (GAAFET) technology.
- •The high demand and TSMC's pricing power are partly attributed to its superior 3nm production yields, which are reported to exceed 90%, significantly higher than Samsung Foundry's 3nm yields, which have hovered around 50% or even 20% for its second-generation process.
- •TSMC is actively expanding its 3nm production footprint globally, with new fabrication facilities planned or under construction in Taiwan, Arizona (USA), and Kumamoto (Japan), to address the overwhelming demand for advanced chips.
📊 競品分析▸ Show
| Feature/Aspect | TSMC (3nm - N3/N3E/N3P) | Samsung Foundry (3nm - 3GAA/SF3) | Intel (3nm - Intel 3) |
|---|---|---|---|
| Transistor Technology | FinFET (last FinFET node) | Gate-All-Around (GAAFET/MBCFET) | Refined FinFET |
| Volume Production Start | Late 2022 (N3), H2 2023 (N3E) | Mid-2022 (initial production) | Planned 2023 |
| Reported Yields | >90% (for N3) | ~50% (for 3nm), ~20% (for SF3/second-gen) | Not widely reported for 3nm production |
| Performance vs. 5nm | 10-15% speed increase or 25-35% power reduction | 23% performance improvement, 45% power reduction, 16% area decrease | Not directly comparable, focuses on performance per watt |
| Logic Density vs. 5nm | 70% increase (N3) | 16% decrease in surface area (vs. 5nm) | About 33% increase (vs. 5nm) |
| Key Customers | Apple, Nvidia, AMD, Broadcom, Intel, MediaTek, Qualcomm | Google (Tensor G5), Nintendo (Switch 2 - 8nm), some AMD, Nvidia, Qualcomm for 3nm (with yield challenges) | Intel's own products, potential for GPU and FPGA chiplets from TSMC |
🛠️ 技術深入
- TSMC's 3nm process (N3) is a full-node advancement over its 5nm generation, offering significant improvements in performance, power, and area.
- Compared to the 5nm process, N3 is projected to deliver a 10-15% increase in performance or a 25-35% decrease in power consumption.
- It achieves a 70% increase in logic density, a 20% increase in SRAM cell density, and a 10% increase in analog circuitry density over N5.
- The N3 process utilizes FinFET (Fin Field-Effect Transistor) technology, which is expected to be TSMC's last node employing this architecture before transitioning to nanosheet (GAAFET) technology for its 2nm node.
- Manufacturing at 3nm requires extensive use of multipatterned Extreme Ultraviolet (EUV) lithography, with N3 using over 20 EUV layers.
- TSMC has developed several variants of its 3nm process family, including N3E (Enhanced), N3P (Performance-enhanced), N3X (optimized for High-Performance Computing with higher clock speeds), N3C (cost-effective), and N3A (automotive applications).
- The N3E variant, for instance, is slightly faster and more power-efficient than the original N3, though it may offer reduced density gains due to single-patterning some layers to lower manufacturing costs.
- TSMC's FinFlex technology allows for intermixing cells with different numbers of fins within a single chip, providing design flexibility for power, performance, and area optimization.
- Projected transistor density for TSMC's 3nm chips is nearly 300 million transistors per square millimeter.
🔮 前景展望基於引用來源的 AI 分析
The price hike will likely accelerate the development and adoption of alternative manufacturing strategies by major chip designers.
With TSMC's 3nm prices rising and capacity remaining tight, customers may invest more in co-development with other foundries or explore advanced packaging solutions to integrate chips from different nodes and suppliers.
TSMC's dominant position in advanced node manufacturing will be further solidified in the short to medium term.
Despite price increases, TSMC's superior yields and established production capabilities at the 3nm node give it a significant competitive advantage over rivals like Samsung, who continue to face yield challenges.
The intense demand for AI chips will continue to be a primary driver for TSMC's capacity expansion and technological advancements.
The ongoing AI boom is the main factor straining TSMC's 3nm capacity, prompting the company to invest heavily in new fabs and advanced packaging technologies to meet the robust multi-year pipeline of AI-related demand.
⏳ 時間線
2020-08
TSMC announced details of its 'N3' process, projecting volume production in H2 2022.
2022-06
Samsung began initial production of its 3nm process node using Gate-All-Around (GAA) transistor architecture.
2022-12
TSMC announced that volume production using its 'N3' 3nm semiconductor node was underway with good yields.
2023-H2
TSMC planned to start volume manufacturing using its refined 'N3E' 3nm process technology.
2024
TSMC's 3nm process reached full capacity, driven by demand from customers like Intel and Apple.
2026-04
TSMC's 3nm process technology contributed 25% of its wafer revenue in Q1 2026.
📎 來源 (31)
Factual claims are grounded in the sources below. Forward-looking analysis is AI-generated interpretation.
- biggo.com
- wikipedia.org
- medium.com
- design-reuse.com
- gadgetbytenepal.com
- digitimes.com
- reddit.com
- nanotechnologyworld.org
- eetimes.com
- qz.com
- techinsights.com
- phonearena.com
- samsung.com
- samsung.com
- edn.com
- cadence.com
- netdrycabinet.com
- imeciclink.com
- techpowerup.com
- techspot.com
- reddit.com
- phonearena.com
- design-reuse.com
- reddit.com
- tsmc.com
- semiwiki.com
- tsmc.com
- ariat-tech.com
- wccftech.com
- substack.com
- patentpc.com
📰
AI 週報
閱讀本週精選 AI 大事摘要 →
👉相關動態
AI 策展新聞聚合。所有內容版權歸原始發布者所有。
原始來源: TechNode ↗
每週電子報
每週一封,可隨時退訂。