来自 Maksym Misichenko · Yahoo Finance ·
AI智能体对这条新闻的看法
The panel agrees that there's an AI-driven capex upcycle for semiconductor equipment players, but they differ on its sustainability and the risks involved. Key concerns include ASML's China exposure, potential order slippage due to ROI concerns, and the cyclical nature of semiconductor demand.
风险: Potential synchronized order cliff due to ROI concerns and geopolitical capacity constraints
机会: Sustained AI-driven demand for semiconductor equipment
本分析由 StockScreener 管道生成——四个领先的 LLM（Claude、GPT、Gemini、Grok）接收相同的提示，并内置反幻觉防护。 阅读方法论 →
吉姆·克莱默在周四晚上“疯狂金钱”节目中称其为“行业历史上最美好的时光”，而数据表明情况也是如此：Applied Materials (纳斯达克：AMAT) 在今年迄今上涨了 75%，原因是人工智能数据中心建设引发了对每个节点、每个晶圆制造厂和接触到晶圆的每件资本设备的短缺。以下五个名称是该建设的过路费收取者，等待回调是 2026 年最昂贵的交易。
1. Onto Innovation：没有人正确定价的计量睡眠者
从办公室里没有人提及的名字开始。Onto Innovation (纽约证券交易所：ONTO) 出售验证高带宽存储器堆叠上每个微凸点的计量和检查工具。每个粘在英伟达 Blackwell GPU 上的 HBM 立方体在运送之前都会经过 Onto 的 Dragonfly 平台。这是人工智能加速器封装的瓶颈，也是分析师仍然低估的供应链部分。
2025 年第四季度创下了 2.6687 亿美元的收入记录，但真正的催化剂是与领先的 HBM 制造商签订的价值超过 2.4 亿美元的采购协议，该协议将持续到 2027 年。首席执行官 Mike Plisinski 表示，“全球人工智能投资推动了半导体资本设备支出强劲的周期性复苏” 增强了对 2026 年及以后超越更广泛设备市场的信心，并且拥有 6.396 亿美元的现金储备，同比增长超过 200%，为下一个增值收购提供了充足的资金。
Onto 今年迄今上涨了 64%，但市值仍约为 130 亿美元。这与列表中下一个名称相比只是一个小数，该名称实际上建造了 Onto 检查的沟槽雕刻机。
2. 莱姆研究：现金机器蚀刻和沉积
如果 HBM 是首选的人工智能存储器，那么莱姆研究 (纳斯达克：LRCX) 是实际堆叠它的公司。莱姆在蚀刻和沉积方面占据主导地位，蚀刻和沉积是 HBM 所需的垂直 3D NAND 和 DRAM 结构的两步过程。来自三星、美光或海力士的每一个新的 HBM3E 和 HBM4 容量公告都转化为莱姆的订单。
3 月季度交付了 58.4 亿美元的收入，同比增长 23.76%，非 GAAP 每股收益为 1.47 美元，超过了共识预期的 1.36 美元，连续第四个季度盈利超过预期。6 月季度的指导预计收入约为 66 亿美元，这是一个在周期性高峰期不会发生的环比加速。首席执行官 Tim Archer 表示“人工智能驱动的需求重塑了半导体行业”，并且 86% 今年迄今的涨幅表明市场相信他。
在 2010 年称呼英伟达的分析师刚刚命名了他的前 10 支股票，而 Applied Materials 并不是其中之一。在这里免费获取它们。
莱姆是显而易见的内存扭矩交易。下一个股票更加明显，它是克莱默正在大力宣传的股票。
3. 适用材料：克莱默的股票和迪克森的胜利巡礼
这就是它。AMAT 在地球上任何其他设备供应商都比其他设备供应商接触更多的芯片制造步骤，并且在周四的“疯狂金钱”节目中，首席执行官 Gary Dickerson 与克莱默坐在一起，告诉他“人工智能正在推动难以置信的计算需求”并且“token 需求在过去几个月增加了 3 倍”。克莱默注意到迪克森为股东带来了 2,897% 的回报。这种复利发生在职业生涯中。
2026 财年第二季度的报告证实了每一个字：收入为 79.1 亿美元，同比增长 11.4%，非 GAAP 每股收益为 2.86 美元，超过了 2.66 美元的共识预期，并且管理层将 2026 年半导体设备增长预期从超过 20% 上调至超过 30%。迪克森告诉克莱默“你无法满足需求”，即使翻倍了运营能力。我观察 Applied 已经十年了，我从未见过首席执行官对订单簿有如此多的前瞻性可见性。
该股票在过去一年中上涨了 180%。下一个名称具有更高的利润率，并且可能具有更好的护城河。
4. KLA：具有软件利润率的过路费收取者
KLA (纳斯达克：KLAC) 拥有工艺控制。大约 90% 的收入来自一个部门，即半导体工艺控制，KLA 在该部门中拥有近乎垄断的市场份额。每个晶圆在每个最先进的晶圆制造厂中都会使用 KLA 工具进行检查。它是半导体资本支出的不可杀死的订阅，利润证明了这一点。
3 月季度交付了 34.2 亿美元的收入，非 GAAP 每股收益为 9.40 美元，超过了 9.15 美元的共识预期，并且 6 月季度的指导预计非 GAAP 毛利率约为 62%。管理层批准了第 17 次年度股息增加，并额外授权了 70 亿美元的回购。首席执行官 Rick Wallace 表示，KLA 是“人工智能生态系统的关键推动者”，涵盖了代工厂/逻辑、存储器、先进封装和技术服务。
KLA 今年迄今上涨了 59%，并且像一家软件公司一样交易，因为它像一家软件公司一样运营。但是，有一个名称位于 KLA 甚至更高一级，没有它，超级周期就不存在。
5. ASML：整个超级周期的守门人
ASML (纳斯达克：ASML) 是地球上唯一一家制造 EUV 和 High NA EUV 光刻系统的公司。每一个 3nm 和 2nm 逻辑芯片、每一个最先进的 HBM 芯片、每一个英伟达、AMD 和 Broadcom 人工智能加速器目前都在 ASML 机器中通过。没有第二来源。没有解决方法。如果台积电、三星或英特尔想要建造最先进的晶圆制造厂，他们就会飞往维尔德霍芬并排队等候。
2026 年第一季度交付了 103.4 亿美元的收入，毛利率为 53.0%，并且管理层将全年预期上调至 360 亿至 400 亿欧元。2030 年的模型目标为 440 亿至 600 亿欧元，毛利率为 56% 至 60%，年末的订单积压已达到 450.6 亿美元。首席执行官 Christophe Fouquet 表示“芯片需求超过供应”，并且客户正在加速其自身客户支持的长期协议。
ASML 今年迄今上涨了 51%。它是世界上最重要的工业公司，市场才刚刚开始这样定价。
设置
Onto 检查 Lam 蚀刻的，Applied 沉积的，KLA 测量，ASML 模式化的。这条链中的每一个环节都在 2027 年之前售罄。迪克森告诉克莱默“这次拐点将持续很长时间”。你等待干净的入口越久，你为像 AMAT 的 75% 今年迄今的涨幅支付的费用就越多。
在 2010 年称呼英伟达的分析师刚刚命名了他的前 10 支人工智能股票
这位分析师的 2025 年精选股票平均上涨了 106%。他刚刚命名了他的前 10 支股票，以供在 2026 年购买。在这里免费获取它们。
四大领先AI模型讨论这篇文章
"Geopolitical restrictions and post-run valuations create more downside risk than the article acknowledges despite strong near-term orders."
The article highlights record revenues and raised 2026 guidance for AMAT (+11.4% YoY to $7.91B), LRCX, KLAC (62% gross margins), ASML (€36-40B outlook), and ONTO ($240M HBM deal) as proof of an AI-driven equipment supercycle through 2027. Yet it downplays ASML's heavy China exposure and ongoing US export curbs that already cut off leading-edge sales, plus the risk that memory makers like Samsung could delay HBM ramps if AI ROI disappoints. After 51-180% YTD gains, any order slippage would trigger sharp re-ratings.
Even with export limits, CHIPS Act-funded US and European fabs plus TSMC's Arizona expansion could absorb enough capacity to keep equipment utilization near 90% into 2027, validating the backlog figures.
"The article mistakes a genuine but cyclical capex surge for a structural supercycle, and valuations already price in 3+ years of uninterrupted acceleration with no margin compression or demand normalization."
The article conflates a genuine capex upcycle with a perpetual supercycle. Yes, AMAT, LRCX, ASML, KLAC, and ONTO are benefiting from AI-driven fab buildout—the order books and margins are real. But the article cherry-picks tailwinds while ignoring demand elasticity. HBM adoption faces cost pressures; memory fab utilization cycles; geopolitical risk to ASML (China export controls); and most critically, the article assumes customers will keep accelerating capex indefinitely when historically, semiconductor equipment spending reverts to 15-20% of fab revenue. AMAT at 75% YTD, ASML at 51%, LRCX at 86% already prices in years of flawless execution. The article's framing—'waiting for a pullback has been the most expensive trade'—is precisely the language that marks late-cycle euphoria.
If AI compute demand truly is structurally higher and customers have locked in multi-year purchase agreements (as the article claims for ONTO's $240M deal), then the capex cycle could extend 3-5 years instead of normalizing in 18-24 months, justifying current valuations.
"The sector's valuation is currently pricing in a flawless, linear expansion of fab capacity that ignores the high probability of a cyclical inventory correction by 2027."
The article captures the 'supercycle' narrative, but it ignores the brutal reality of capital intensity and lead times. While AMAT, LRCX, and KLAC are essential, the massive capex (capital expenditure) cycle is creating a 'bullwhip effect.' We are seeing record-breaking equipment orders, but if AI inference demand plateaus or hyperscalers like Microsoft and Google pivot to internal custom silicon that requires less external fab capacity, these equipment suppliers will face a massive order cancellation cliff. ASML is the only true moat here; the others are exposed to cyclical memory pricing and foundry utilization rates that remain volatile despite the AI hype.
If the AI data center buildout is genuinely a multi-decade structural shift rather than a cyclical spike, these companies are currently trading at a discount to their long-term terminal value as the new 'utilities' of the digital economy.
"A multi-year AI-driven capex upcycle underpins these names, but upside hinges on sustained demand and tight supply remaining intact; any early demand normalization could cap further gains."
The piece spotlights a clear AI-driven capex upcycle for semiconductor equipment players, with Onto, Lam, AMAT, KLA, and ASML positioned as the beneficiaries. Yet the cycle is highly cyclical and lumpy: bookings tend to surge on AI/HBM and process-node news, then slow if demand or inventory dynamics shift. Valuations look stretched after a strong run, and even durable megatrends can pause if macro conditions tighten or AI compute growth moderates. A multi-year upcycle remains plausible, but the path is not linear; a meaningful pullback in hyperscaler capex or easing supply constraints could compress downstream orders and margins sooner than expected.
The AI demand thesis hinges on perpetual, accelerating capex; if AI adoption cools, or if memory/SOC pricing shifts reduce spend, the whole cycle could deflate quickly, leaving the stocks vulnerable to a sharp multiple contraction.
"Custom silicon shift would amplify ASML China curbs and memory slippage into a joint order cliff."
Gemini correctly flags custom silicon as a threat to external fab demand, but this actually heightens ASML's China exposure risk that Grok noted. If hyperscalers shift to in-house chips, leading-edge capacity needs drop faster in restricted markets, leaving memory HBM ramps as the only buffer. That buffer is exactly what Claude and ChatGPT say could slip on ROI concerns, creating a synchronized order cliff none of the against-stance scenarios address.
"Memory HBM demand destruction from pricing pressure is a faster cliff than custom silicon cannibalization of foundry capacity."
Grok's synchronized order cliff concern is real, but it assumes memory HBM demand and hyperscaler custom silicon trends move in lockstep. They don't. Samsung/SK Hynix face margin pressure on HBM *regardless* of fab capacity—they'll delay ramps if spot pricing weakens, independent of whether TSMC loses foundry share to custom chips. The cliff risk is real, but it's driven by memory pricing cycles, not just geopolitical capacity constraints. ASML's China exposure is separate.
"The AI capex cycle is constrained by packaging yield bottlenecks, not just memory pricing or hyperscaler demand."
Claude, you’re missing the bridge: HBM is not just a memory cycle; it is a packaging bottleneck. The 'cliff' isn't just pricing—it is the physical limit of CoWoS (Chip-on-Wafer-on-Substrate) capacity. AMAT and KLAC are the real leverage here, not just the memory makers. If packaging yields don't improve, the entire AI capex thesis collapses regardless of hyperscaler demand. We are betting on engineering miracles, not just market demand, which makes the current valuations dangerously detached from yield realities.
"Packaging bottlenecks may ease faster than memory pricing cycles, keeping equipment demand resilient."
Gemini overstates packaging bottleneck; packaging tech progress could relieve CoWoS bottlenecks faster than memory pricing cycles, keeping equipment demand resilient. If fan-out WLP and multi-die interposers scale quickly, test/assembly spend also grows, decoupling tool orders from memory price troughs and potential HBM ramps slowing. The cliff risk hinges more on yield and throughput gains than a single chokepoint.
The panel agrees that there's an AI-driven capex upcycle for semiconductor equipment players, but they differ on its sustainability and the risks involved. Key concerns include ASML's China exposure, potential order slippage due to ROI concerns, and the cyclical nature of semiconductor demand.
Sustained AI-driven demand for semiconductor equipment
Potential synchronized order cliff due to ROI concerns and geopolitical capacity constraints