半导体供应链处于边缘？
2026年第二季度半导体财报季已结束，连此前落后的企业也已完成2026年第三季度首次财报电话会议。文章据此梳理半导体行业当前状况，但原文未给出具体公司、数字、型号、时间地点及供应链细节。
原文深度解读
文章以2026年第二季度半导体财报季结束为背景，讨论半导体行业及其供应链的现状。涉及主体包括存储公司（如SK Hynix、Samsung）、AI计算公司（如Nvidia、Broadcom/Google/Anthropic）、晶圆代工厂、OSAT、PCB及PCB材料、被动/机电/连接器（PEMCO）等。关键数字包括：2026年Q2半导体市场收入增加1000亿美元，环比增长31%；前五大超大规模企业资本支出增长330亿美元至1810亿美元；存储占半导体市场收入份额从25%升至53%，占行业运营利润65%；AI数据中心占半导体销售额56%以上；存储资本支出占总资本支出45%，低于上一轮周期峰值52%；PCB材料原材料库存98天等。文章核心与电子元器件供应链的关系在于：分析半导体价值链中定价权、库存结构、投资纪律和需求变化如何沿供应链传导，并指出价值仍集中在顶部，但设备部门开始参与增长。
- 2026年Q2半导体市场收入增加1000亿美元，环比增长31%，比上一轮半导体周期整体收入增长高800个基点；若周期现在结束，自上次峰值以来增长为3倍，即31%的复合年增长率。
- 存储占半导体市场收入份额在两年多内从25%升至2026年Q2的53%，预计2026年Q3进一步上升；存储占行业运营利润65%，是AI计算公司（包括Nvidia）的两倍多，而其余半导体公司仅占不到7%的运营利润。
- 前五大超大规模企业2026年Q2资本支出增长330亿美元至1810亿美元，环比增长22%；其中750亿美元增长来自存储，大部分必须来自超大规模企业资本支出预算。
- 约一半当前存储产能已签订长期合同并固定产能承诺，因此回归2022-23年存储崩盘的可能性不大；HBM每比特所需产能从标准DRAM的2-3倍升至3-4倍，且随密度和速度提升还会更糟。
- 存储公司资本支出占总资本支出45%，低于上一轮存储周期峰值时的52%；其他半导体公司仅占投资13%；材料公司进入“纪律投资模式”，资本支出持续下滑。
- 供应链库存方面：PCB材料原材料库存为98天，可能构成问题；半导体材料和封装库存状况良好但增长低；WIP库存下降表明非AI半导体产品的裸片库存仍在消耗；成品库存已连续三年下降。
供应链影响
- 存储产能向HBM倾斜可能持续挤压标准DRAM和NAND的可用产能，取决于HBM渗透速度和长期合同覆盖比例，标准存储供应可能维持紧张。
- 材料公司资本支出纪律可能限制未来基础材料（如覆铜板、树脂、玻璃布）的扩产速度，若半导体设备需求如收入图表所示回升，材料供应可能成为瓶颈。
- PCB材料原材料库存98天且投资下滑，可能使PCB供应链在需求突然上升时面临原材料短缺风险，具体影响取决于PCB制造商自身库存缓冲和替代采购能力。
- AI数据中心支出占半导体销售额56%以上，若前沿AI模型开发放缓导致数据中心支出降温，可能随时间影响存储价格，但长期合同可能缓冲存储公司的短期冲击。
- 供应链价值仍集中在半导体设备顶部，代工厂和OSAT开始受益但封装材料尚未完全跟进，PCB中材料比制造本身捕获更多价值，这种定价权分布可能继续影响各环节的议价能力和库存策略。
系统解读，仅作为判断线索，不构成备货、出货、涨价或投资建议。
2026年第二季度的半导体财报热潮早已过去，就连落后者现在也已召开了2026年第三季度的首次财报电话会议。是时候把握半导体行业及其供应链正在发生什么了。
头条新闻已经很清楚，大多数专家已经根据根深蒂固的信念和什么对他们钱包有利形成了观点。他们关注下一季度，并且也已经对此下定了决心。
专家心态是一种自以为更懂、从不惊讶的心态。“我早就看到了”和“我告诉过你”告诉你你在跟什么打交道。
分析师的心态则不同。即使情况显而易见，分析中仍有很多东西可以学习，但大多数半导体利益相关者没有时间做这些。他们有生意要经营，而所有这些商业分析并不会带来什么不同。
大多数时候，他们可能是对的。然而，他们都曾处于宁愿不惜一切代价避免的境地。在某个时间点下单太少，在另一个时间点下单太多。错判半导体周期代价高昂，无论上行还是下行，而相信涓滴经济学在供应链中总是有效理论可能是毁灭性的。
在同一个周期内被雇用、解雇又重新雇用。问题在于，所有这些事件事后看来都容易理解，甚至更容易被归类为黑天鹅事件。
这使管理层免于任何管理不善的责任，但黑天鹅很罕见。我不是天鹅专家，但我在冬季季度每天散步时听到数千只天鹅的声音，我的分析尚未产生一只黑天鹅。
然而，灰天鹅有很多。丑小鸭都很年轻，大多数会变成白色，但你永远不知道。这正是商业分析变得有价值的地方。
我们都知道这是存储的季度。价格持续暴涨，存储公司钱多得不知道该怎么花。韩国Tinder上最受欢迎的头像是穿着破旧SK Hynix T恤的极客。穿Samsung T恤风险更大，因为那可能是移动部门的一个可怜人。
但知道发生了什么并不等于理解它。如果你问你最喜欢的LLM发生了什么，你会得到一大盘没有蛋白质的词语沙拉。这只是在洞察的针堆上又添了干草，并不能让你更接近理解情况。
如果你再折磨LLM一下，你会发现它拥有大量知识但理解甚少。它难以将点连接起来形成高层视图。
尽管LLM确实在变得更好，我仍然相信我在分析半导体供应链方面有工作要做，但我会让你来判断。
这项分析应放在当前全球背景下看待，即社会经济冲击的可能性增加。听起来可能反直觉，特朗普政府的关税政策在很大程度上有助于强化半导体供应链，在不可预测性中提高了其稳健性。
然而，我们正面临美国攻击伊朗可能带来的潜在冲击。最明显的是石油冲击，由于沙特管道遭袭、霍尔木兹海峡关闭以及胡塞武装向曼德海峡推进等最新事态发展。
2026年第二季度半导体市场收入
这个季度是前所未有的，但“前所未有”已不再前所未有。市场再次处于快速、不受控制的转型中，涟漪将波及整个供应链。
该季度为半导体市场收入增加了1000亿美元，与6年前整个市场一样多。环比增长率为31%，比上一轮整个半导体周期的收入增长高800个基点。
如果仍然存在“半导体周期”，并且它现在结束（在正常情况下会如此），那么自上次峰值以来，该周期的增长将是3倍，即31%的复合年增长率。
当前季度代表着较迄今为止平均周期增长的显著加速。
所有图片 © Claus Aasholm
我们现在已经过了半导体市场在AI计算、存储和所有其他半导体之间均匀分布的拐点几个季度，AI存储时代正处于全盛期。在1000亿美元的增长中，750亿美元来自存储，其中大部分必须来自超大规模企业的资本支出预算。
前五大超大规模企业的资本支出在2026年第二季度增长了330亿美元至1810亿美元，因此可以合理假设22%的环比资本支出增长大部分不会用于额外产能，而是用于为已在进行的项目支付存储税。
从这个角度看，大型AI实验室呼吁放缓前沿AI模型开发可能并不令人意外。他们的论点是AI变得太好太快（营销还是真实？）。可以肯定的是，AI数据中心支出降温将随时间影响存储价格。
当前存储产能的近一半现在处于长期合同下，有固定产能承诺，因此回归2022-23年存储崩盘的可能性不大。
虽然存储天鹅不是白色，但也肯定不是黑色。在上一轮存储行业萧条留下伤痕之后，当时大型存储公司在负毛利率下满负荷运营，它们已向自己承诺绝不让这种情况再次发生。
向高带宽存储器（HBM）的转变是远离商品市场、走向更可持续商业模式的转变，但问题是HBM占用大量产能。
一个HBM比特最初需要标准DRAM比特2-3倍的产能，现在需要3-4倍，而且随着HBM变得更密集、更快，这只会变得更糟。
标准DRAM产能已被蚕食用于生产HBM，导致整体存储产能下降。
与此同时，存储公司在资本支出投资上格外谨慎。冒着听起来像专家而非分析师的风险，我在2024年5月的一篇文章中指出了这一点。
虽然我和任何人一样喜欢自己是对的，但这并不重要。重要的是迹象已经存在，存储天鹅并不像一些专家所说的那样是黑色的。
存储时代现在正在重塑整个行业的构成，如下所示。
在短短两年多时间里，存储占半导体市场的份额从25%上升到2026年第二季度的53%，即使前沿实验室休息，2026年第三季度它还会进一步增加。积压订单已经到位。
这应结合销货成本（COGS）更温和的增长来看，后者更接近行业实际生产份额的代理指标。同期COGS增长“仅”增加了700个基点，达到行业COGS的40%。
凭借40%的COGS，存储公司现在占行业运营利润的65%，是AI计算公司（包括Nvidia）的两倍多。
一年多前，在AI计算时代，AI计算公司占据三分之二的运营利润，虽然运营利润增加了，但不足以抵消向存储公司的财富再分配。
其余半导体公司占总运营利润不到7%。
这种转变是激进的，从下面按公司类型划分的员工人数图表中可以明显看出。
值得注意的是，AI计算公司通常没有生产员工，而存储公司有内部制造，其他公司则采取混合方式；这并不能掩盖一个事实：我近百万同事中的三分之二只能分享行业运营利润的不到7%。这是史诗般的财富分配。
我没有忘记评论之前的资本支出图表，因为它需要额外一张图表来完成分析。
虽然存储资本支出几乎完全掌握在自己手中，但另外两组的部分或全部硅制造外包给代工厂。
代工厂资本支出添加如下。
有趣的是，尽管据称存储公司进行了大规模投资，但它们占总资本支出的份额为45%，远低于上一轮存储周期峰值时记录的52%。
要么它们仍然在资本支出投资上保持“纪律”，要么它们尚未达到晶圆厂建设进度的设备阶段。毫无疑问，存储公司目前正在享受其稀缺性驱动的定价杠杆。
此外，其他半导体公司在这里也处于赤字状态，仅占投资的13%。
半导体市场
已经通过普遍共识确立，数据中心正在驱动当前市场。幸运的是，数据也显示了这一点，所以这次我将避免陷入麻烦。
十年前曾经平衡良好的产品类别，消费类（我将移动归类为消费产品）和计算类分别以34%和40%的份额为最大类别，现在已成为计算市场。
右图显示了按主要应用划分的相同市场视图，计算市场占三分之二，数据中心收入份额为56%。
虽然之前也有其他强劲的产品需求周期，但份额在不到三年内增长近3倍是前所未有的。
这再次表明更广泛的半导体行业如何被迫陷入困境，而大公司主导着竞技场。
这不仅对公司本身及其65万名员工是个问题，还将扰乱供应链，并可能对社会产生更大影响。
许多拥有“其他”设备的“其他”公司提供的应用对我们的健康、安全、保障、营养、职业、出行和总体福祉很重要，但它们在当前市场转型中未被优先考虑。
由于半导体增长一直由少数公司的利润主导，对供应链的影响很小，但这种情况即将改变。
供应链分析
半导体供应网络很复杂，公司通常参与多个市场。虽然我知道我可以用复杂性迷惑我的受众，但我也非常能够迷惑自己，这更糟糕。
我通常使用颜色代码来帮助管理复杂性，并将在以下分析中这样做。
半导体市场最纯粹的形式是设备市场，尽管Nvidia的大部分收入来自完全组装的服务器。在本分析中，我使用“一般”做法将其归类为设备收入，但当更有用时，我也可以将其视为服务器/芯片收入。
虽然PCB和PCB材料不直接属于半导体供应链，但它们是更大的半导体供应网络的平行分支，将被包括在内。
被动、机电和连接器归入PEMCO标题下。
供应网络收入矩阵如下所示：
这张图表清楚地展示了从顶部半导体设备到下方供应网络其余部分的定价权瀑布。价值和增长被困在顶部。
虽然代工厂从参与计算AI增长中受益，但它们在结构上被排除在存储繁荣之外，因为存储公司有自己的制造，HBM基础裸片除外。
还可以看到，价值现在正蔓延到OSAT层，但封装材料尚未完全跟进。在PCB中，情况相反。材料比PCB制造本身捕获更多价值。
这些增长信号与定价权和潜在短缺密切相关，采购团队应予以监控。
从周期角度看，分包、PCB和PCB材料的峰值比直接半导体供应链峰值早几个季度。
换句话说，上一次下跌显然是需求驱动的，而不是产能过剩驱动的价格下跌。
虽然我们尚未确定半导体周期是否仍然存在，但信号可能来自分包市场和供应链。
自2019年初以来的相对增长显示了半导体供应网络随时间的变化。
虽然大多数部门对2020-22年新冠/芯片短缺周期的反应相似，但一些领域显示出结构性变化。虽然半导体行业当前的AI/存储繁荣目前是事件驱动的，但代工厂增长是结构性的，即使由单一公司驱动。
从噪声到信号的转变也可以在PCB材料和OSAT中看到。
很明显，大多数供应链现在的收入水平高于早期AI计算和AI存储时代，有必要确定库存短缺是否可能给半导体公司带来即时痛苦。
库存结构与水平
虽然我监控绝对库存水平，但除非结合业务活动水平来看，它们真的没有意义。库存天数是一个更重要的指标，可以揭示业务的结构性和即时变化。
从部门概览来看，该数字不受个别市场份额变化的影响。
在下面的图表中，按部门划分的库存天数在同一轴刻度上：
虽然大多数部门随时间保持稳定，但有几个异常值突出。新冠/短缺周期非常明显，连同超调结果。代工厂库存也是如此，在2026年第二季度也显示出明显的峰值。
另一个结构性变化是PCB材料的下降，需要进一步检查。
通过检查下面的库存结构可以揭示更多细节：
库存类型显示，虽然PCB材料下降确实是结构性的，但它也仅影响原材料库存。在早期覆铜板、树脂和玻璃布材料短缺期间，PCB材料公司充当了PCB供应链中的减震器，但随着PCB材料的质量和复杂性提高，基础材料变得更难获得足够数量。98天的原材料库存可能构成问题，即使PCB制造商本身已相应增加了原材料库存。
半导体材料和封装库存状况良好，但也应结合低增长来看。如果活动如收入图表所示回升，稳定如常的库存可能不够。
尽管半导体库存全年呈下降趋势，事情可能很快变得非常有趣。原材料库存随着价格上涨而上升，关键产品被更频繁地锁定。存储也可能受到轻微影响，尽管它们几乎不会在库存中停留几秒钟，如果有的话。
WIP库存下降表明非AI半导体产品的裸片库存仍在消耗，没有证据表明生产普遍增加。成品库存现已连续三年下降，反映了上一轮周期的大规模供应过剩。
众所周知，AI供应链的几个专业领域供应短缺，但除此之外，整体供应链库存状况良好，只有少数例外。
更令人担忧的是不同部门的投资选择。
下面的投资网格显示了按部门和季度划分的研发和资本支出占收入的百分比。虽然每个部门需要相当不同的投资水平，但每个部门的时间轨迹（8个季度）讲述了正在发生的事情。
围绕一个最佳点，代工厂和OSAT公司正在固定运营模式中根据收入调整投资。PCB公司一直处于资本支出投资模式，同时将研发利润率固定在4%。
我已将存储和AI公司从半导体图中排除以使其有意义，即使如此，收入增长也使相对投资下降。
最有趣的观察是三个基础材料类别的资本支出故事。从鉴于收入和有限增长肯定不现实的资本支出水平开始，下滑持续，上一季度也出现下跌。
像存储公司一样，材料公司正在进入“纪律投资模式”，并不害怕制造一点稀缺。这可能对半导体行业产生重大影响，因为收入增长现在开始在整个供应链中传播。
需求情况
虽然分析的前一部分侧重于半导体行业及其供应链的结构性发展，现在是时候看看2026年第二季度报告中的最新发展，包括等式的需求侧。
下面的图表显示了供应链中的产品流动和上一季度的收入变化。
半导体设备公司的强劲收入增长在细节中显而易见，而且毫不意外，它是由集成设备制造商中的存储公司驱动的。最好的猜测是短缺将持续到2028年，除非需求崩溃；即使如此，长期合同将使存储公司维持下去。
HBM战略旨在为未来每一代捕获更多DRAM和NAND产能，这使情况更糟。
材料收入增长强劲，结合本十年历史上疲软的增长来看。较低的投资水平和较高的收入增长相结合，可能对材料客户产生负面影响。
在设备层面，OSAT和PCB现在表现出一段时间未见的强劲增长。虽然PCB公司正在增加投资，OSAT公司正以收入增长的速度投资。
在需求侧，分包市场正在增长，受服务器炙热需求驱动。智能手机市场停滞，而PC市场出人意料地有韧性。尽管所有增长都是由更高的存储价格驱动，但销售没有被当前存储价格杀死（目前？）仍然令人印象深刻。
投入成本的图表，如销货成本增长所反映，证实了这一点。
这也显示了半导体行业价值的急剧增加，特别是存储IDM。还可以看到，其他部门的大多数增长与其收入增长相似，表明是成本转嫁而非有机增长。
虽然供应链迄今能够处理销货成本的增长，但仅本季度的COGS增长就是去年速率的两倍。
结论
我们进入AI存储时代还不到四分之三，看不到尽头。尽管AI模型可能放缓，存储公司已获得长期订单和价格协议，Nvidia的业务依然强劲，而Broadcom/Google/Anthropic已公布重大投资计划。半导体故事现在只有一个章节。AI数据中心建设现在占半导体销售额的56%以上。
卖空者正在耗尽空头，因为他们的故事仍然真实且无成效。整个AI革命取决于更高效AI的成功开发，而这由增加的计算能力实现。如果这一点改变，一切都会改变。
由于AI革命在很大程度上扩大了半导体利润率，半导体供应链没有被邀请参加派对，其收入仅温和增长。虽然设备部门已经增长了一段时间，材料市场正在从长期冬眠中醒来。
半导体供应网络中的价值仍然被困在顶部，但设备部门现在正在参与。
除了PCB材料，库存
The Q2-2026 semiconductor earnings rush is long gone, and even the laggards have now reported in their first Q3-2026 earnings calls. It is time to get a grip on what is going on in the semiconductor industry and its supply chain.
The headlines are already clear, and most experts have formed their opinion in line with the dug-in convictions and what benefits their wallets. They are focused on the next quarter and have already made up their minds about that too.
An expert mindset is one that knows better, and it is never surprised. “I saw that coming” and “I told you so” tell you what you are dealing with.
An analyst's mindset is different. Even if the situation is obvious, there is still a lot to learn from analysis, but most semiconductor stakeholders do not have time for that. They have a business to run, and all that business analysis doesn't make a difference.
And most of the time, they are probably right. They have, however, all been in situations they would have done anything to avoid. Placed too few orders at a given time and too many at another. Mistiming the semiconductor cycle is expensive, both up and down and believing trickle-down economics is always a valid theory in the supply chain can be devastating.
Hired, fired and rehired in the same cycle. The problem is that all of these events are easy to understand in hindsight and even easier to categorise as black swan events.
This absolves management of any blame for mismanagement, but black swans are rare. I am not an expert in swans, but my analysis on my daily walk by the sound that is the winter quarter for thousands of swans has yet to produce one.
There are, however, plenty of grey swans. The ugly ducklings are all young, and most of them turn white, but you never know. This is where business analysis becomes valuable.
We all know that this has been the memory quarter. Prices have continued to explode, and memory companies have more money than they know what to do with. The most popular profile picture on Tinder in Korea is a geek in a worn SK Hynix T-shirt. A Samsung T-shirt is riskier, as it could be a poor soul from the mobile division.
But knowing what is going on is not the same as understanding it. If you ask your favourite LLM what is going on, you get a large portion of word salad with no protein. This is all more hay on the insight needles and does not get you nearer to understanding the situation.
If you torture the LLM some more, you will find that it has plenty of knowledge but little understanding. It has difficulty connecting the dots to form a high-level view.
Although LLMs are certainly getting better, I still believe I have a job to do in analysing the Semiconductor supply chain, but I will let you be the judge of that.
This analysis should be seen in light of the current global context of an increased potential for socio-economic shocks. In what might sound counterintuitive, the Trump administration’s tariff policies have done a lot to harden the semiconductor supply chain, improving its robustness amid unpredictability.
We are, however, facing a potential shock from the US attack on Iran. The most obvious one is an oil shock due to the latest developments like the strike on Saudi pipelines, the Hormuz Strait closure and the Houthis’ advancements towards the Bab al-Mandab Strait.
Semiconductor Market Revenue Q2-2026
The quarter was unprecedented, but that is not unprecedented any more. The market is once again in rapid, uncontrolled transformation, and the ripples will be felt throughout the supply chain.
The quarter added another USD 100B to the semiconductor market revenue, the same as the entire market just 6 years ago. The QoQ growth rate was 31%, or 800 basis points higher than the revenue growth over the entire last semiconductor cycle.
If there still is a “semiconductor cycle” and it ends now (as it would under normal circumstances), the cycle's growth would have been 3x, or a 31% CAGR, since the last peak.
The current quarter represents a significant acceleration from the average cycle growth so far.
All images © Claus Aasholm
We are now a couple of quarters past the inflexion point where the Semiconductor market was evenly divided among AI Compute, Memory, and all other semiconductors and the AI memory era is in full bloom. Of the USD 100B growth, USD 75B is in memory, and most of it has to come from the hyperscaler CapEx budget.
The CapEx of the five top hyperscalers grew by USD 33B to USD 181B in Q2-2026, so it is fair to assume that most of the 22% QoQ CapEx growth will not be used for additional capacity but rather to pay the memory tax for projects already in progress.
Seen in this light, it is probably not surprising that the large AI Labs have called for a slowdown in frontier AI model development. Their argument is that AI is becoming too good too fast (marketing or true?). What is certain is that a cooling in AI data centre spending will impact memory prices over time.
Close to half of the current memory capacity is now under long-term contracts with fixed capacity commitments, so a return to the 2022-23 memory crash is not plausible.
While the memory swan is not white, it is certainly not black either. After the scars of the last memory-sector depression, when the large memory companies operated at full capacity at negative gross margins, they have promised themselves never to let that happen again.
The move toward High Bandwidth Memories (HBM) was a shift away from commodity markets and toward a more sustainable business model, but the issue is that HBM sucks up a lot of capacity.
An HBM bit that initially required 2-3 times the capacity of a standard DRAM bit now requires 3-4 times as much, and this will only get worse as HBMs become denser and faster.
Standard DRAM capacity has been cannibalised to produce HBM, resulting in a decline in overall memory capacity.
At the same time, the memory companies were more than careful with their CapEx investments. Given the risk of sounding like an expert rather than an analyst, I pointed this out in a post in May 2024.
While I like to be right as much as anybody, this is beside the point. The point is that the signs were there, and the memory swan is not black as some experts suggest.
The Memory Era is now reshaping the composition of the entire industry, as shown below.
In just over two years, the memory share of the semiconductor market has risen from 25% to 53% in Q2 of 2026, and it will increase further in Q3 of 2026 even if the frontier labs take a break. The backlog is already in place.
This should be seen in light of the more modest growth in Cost of Goods Sold, which is a closer proxy to the actual share of production in the industry. The COGS growth over the same period has “only” increased by 700 basis points to 40% of the industry’s COGS.
With 40% of COGS, memory companies now account for 65% of the industry's operating profit, more than twice that of AI compute companies, including Nvidia.
A little over a year ago, in the AI compute era, AI compute companies commanded two-thirds of operating profits, and while operating profits have increased, this is not enough to offset the wealth redistribution to memory companies.
The remaining semiconductor companies account for less than 7% of total operating profits.
This transformation is radical, as evident in the headcount chart by company type below.
It is worth noting that AI compute companies generally have no production employees, while memory companies have internal manufacturing, and other companies have a mixed approach; this does not obscure the fact that two-thirds of my almost a million colleagues have to share less than 7% of the industry’s operating profits. This is a wealth distribution of epic proportions.
I did not forget to comment on the earlier CapEx chart, as it needs an additional chart to complete the analysis.
While the Memory CapEx is almost solely in their own hands, the two other groups have some or all of their silicon manufacturing outsourced to a foundry.
The Foundry CapEx is added below.
What is interesting is that despite alleged massive investments by memory companies, their share of total CapEx is 45%, well below the 52% they recorded at the peak of the last memory cycle.
Either they are still being "disciplined” with their CapEx investments, or they have not yet reached the tooling phase of the fab construction schedule. No doubt the memory companies are enjoying their scarcity-driven pricing leverage at the moment.
Also, here the other semiconductor companies are in deficit, with only 13% of the investments.
Semiconductor Markets
It is already established by popular agreement that the data centre is driving the current market. Fortunately, this is also what the data shows, so I will avoid getting into trouble this time around.
What used to be well-balanced product categories a decade ago, with consumer (I categorise mobile as a consumer product) and computing as the largest categories at 34% and 40% share, have now become a compute market.
The right-hand graph shows the same market view by major application, with the compute market at two-thirds and the datacenter revenue share at 56%.
While there have been other strong product demand cycles before this one, the almost 3x increase in share in less than three years is unprecedented.
This once again demonstrates how the broader semiconductor industry is being forced into a corner while large companies dominate the arena.
Not only is this a problem for the companies themselves and their 650.000 employees, but it will also disrupt the supply chain and could have a larger impact on society.
Many of these “other” companies with “other” devices provide applications that are important to our health, safety, security, nutrition, occupation, mobility, and general wellbeing, yet they are not prioritised in the current market transformation.
As semiconductor growth has been dominated by the profits of a few companies, the impact on the supply chain has been minimal, but that is about to change.
Supply Chain Analysis
The semiconductor supply network is complex, and companies often participate in more than one market. While I know I can confuse my audience with complexity, I am also very capable of confusing myself, which is much worse.
I generally use colour codes to help me manage complexity and will do so in the following analysis.
The semiconductor market in its purest form is a device market, although most of Nvidia’s revenue comes from fully assembled servers. In this analysis, I use the “general” practice of categorising this as device revenue, but I can also treat it as server/chip revenue when more useful.
While PCB and PCB materials are not directly part of the Semiconductor supply chain, they are a parallel branch of the larger semiconductor supply network and will be included.
Passive, electromechanical, and Connectors are grouped under the PEMCO header.
The supply network revenue matrix is shown below:
This chart clearly demonstrates the pricing power waterfall from semiconductor devices at the top to the rest of the supply network below. The value and growth are trapped at the top.
While the foundries benefit from participation in the Compute AI growth, they are structurally excluded from the memory boom, as the memory companies have their own manufacturing, with HBM base dies excluded.
It can also be seen that value is now creeping into the OSAT layer but is not yet fully followed by the packaging materials. In PCB, the opposite is true. The materials are capturing more value than the PCB manufacturing itself.
These growth signals are closely associated with pricing power and potential shortages, and should be monitored by purchasing teams.
From a cycle perspective, the Subcontract, PCB, and PCB Materials peaks predate the direct semiconductor supply chain peak by a couple of quarters.
In other words, the last dip was clearly demand-driven rather than due to overcapacity-driven price declines.
While we have yet to determine whether the semiconductor cycle still exists, the signal will likely come from the subcontracting market and the supply chain.
The relative growth from the beginning of 2019 shows the changes in the semiconductor supply network over time.
While most sectors responded similarly to the Covid/Chip shortage cycle in 2020-22, some areas show structural changes. While the current AI/Memory boom in the semiconductor sector is event-driven for now, the foundry growth is structural even if driven by a single company.
The switch from noise to signal can also be seen in PCB material and OSAT.
It is clear that most of the supply chain is now operating at a higher revenue level than in the earlier AI Compute and AI Memory eras, and it is necessary to determine whether inventory shortfalls could cause immediate pain for semiconductor companies.
Inventory Structure and levels
While I monitor absolute inventory levels, they are really meaningless unless seen in the context of the business activity levels. Days of inventory is a much more important metric that can reveal both structural and immediate changes in business.
When viewed as a sector overview, the number is not impacted by individual changes in market shares.
In the following chart, the days of inventory by sector is on the same axis scale:
While most sectors remain stable over time, a couple of outliers stand out. The Covid/Shortage cycle is very visible, along with the overshoot result. The same is evident in the foundry inventory, which also shows a pronounced spike in Q2-26.
The other structural change is the decline in PCB materials that requires further examination.
More detail can be uncovered by examining the structure of the inventory as seen below:
The inventory type shows that while the PCB material decline is certainly structural, it is also solely affecting raw materials inventory. During the earlier shortages of materials for Copper Clad Laminate, Resins, and Glass Cloth, PCB material companies acted as shock absorbers in the PCB supply chain, but as the quality and complexity of PCB materials have increased, the basic materials have become harder to obtain in sufficient quantities. The 98-day raw materials inventory could pose a problem even though the PCB manufacturers themselves have increased their raw materials inventory in response.
The Semiconductor materials and packaging inventories are in good shape but should also be seen in the light of low growth. Should activity pick up as the revenue charts suggest, the steady-as-she-goes inventories might not be sufficient.
Despite semiconductor inventories trending down over the year, things could soon become very interesting. Raw material inventories are rising as prices increase, and critical products are being secured more frequently. Memory might also be impacted slightly, although they do not stay in inventory for many seconds, if at all.
The declining WIP inventory indicates that the die banks are still being depleted for non-AI semiconductor products, with no evidence of a general increase in production. The finished goods inventories have now trended down for three years, reflecting the massive oversupply in the last cycle.
It is already well known that several specialised areas of the AI supply chain are in short supply, but apart from that, the overall supply chain inventory is in good shape, with a few exceptions.
What is more worrying are the investment choices in the different sectors.
The Investment Grid below shows R&D and CapEx spend as a percentage of revenue by sector and quarter. While each sector requires quite different levels of investment, the time trail (8 quarters) of each sector tells a story about what is going on.
Circling around a sweet spot, the Foundry and OSAT companies are adjusting their investments to their revenues in a fixed operating model. The PCB companies have been in CapEx-investment mode while keeping the R&D margin fixed at 4%.
I have excluded memory and AI companies from the semiconductor plot to make it meaningful, and even so, the revenue growth makes relative investments go south.
The most interesting observation is the CapEx story in the three basic material categories. From what were certainly unrealistic CapEx levels given the revenue and limited growth, the slide has continued, with dips in the last quarter as well.
Like the memory companies, the materials companies are entering a “disciplined investment mode” and are not afraid of creating a little scarcity. This can have a significant impact on the semiconductor industry, as revenue growth is now beginning to propagate throughout the supply chain.
The demand situation
While the prior part of the analysis focused on the structural developments in the semiconductor industry and its supply chain, it is now time to look at the most recent developments in the Q2-2026 reporting, including the demand side of the equation.
The chart below shows the flow of products in the supply chain and the change in revenue from last quarter.
The strong revenue growth of semiconductor device companies is evident in detail, and, to no one’s surprise, it is driven by memory companies among the Integrated Device Manufacturers. The best guess is that the shortage will last until 2028 unless demand collapses; even then, long-term contracts will keep memory companies afloat.
The HBM strategy is designed to capture more DRAM and NAND capacity for every future generation, which makes things even worse.
The materials revenue growth is strong, seen in light of the historically weak growth over this decade. The combination of lower investment levels and higher revenue growth is likely to negatively impact material customers.
At the device level, OSAT and PCB are now exhibiting strong growth not seen for some time. While PCB companies are increasing investment, OSAT companies are investing at the rate of revenue growth.
On the demand side, the subcontract market is growing, driven by red-hot demand for servers. The smartphone market has stalled while the PC market is surprisingly resilient. Even though all the growth is driven by higher memory prices, it is still impressive that sales are not being killed by current memory prices (yet?).
The plot of input costs, as reflected in Cost of Goods Sold growth, confirms this.
This also shows the dramatic increase in value in the semiconductor sector, particularly for memory IDMs. It can also be seen that most growth in the other sectors is similar to their revenue growth, suggesting cost pass-through rather than organic growth.
While the supply chain has been able to handle the increase in Cost of Goods Sold up to now, COGS growth this quarter alone is at twice last year's rate.
Conclusion
We are not three-quarters of the way into the AI Memory Era, with no end in sight. Despite potential AI model slowdowns, memory companies have secured long-term orders and price agreements, and Nvidia’s business remains strong, while Broadcom/Google/Antropic has unveiled significant investment plans. The Semiconductor story now has only one chapter. The AI data centre buildout now accounts for more than 56% of semiconductor sales.
The short sellers are running out of shorts as their story remains true and unproductive. The entire AI revolution rests on the successful development of more productive AI, enabled by increased compute. If that changes, everything changes.
As the AI revolution has largely expanded semiconductor margins, the semiconductor supply chain has not been invited to the party, and its revenue has grown only modestly. While the device sector has been growing for some time, the materials markets are waking up from a long hibernation.
The value in the semiconductor supply network is still trapped at the top, but the device sectors are now participating.
Except for PCB materials, the inventory situation remains reasonable, with pockets of issues related to the AI supply chain.
As we move into a higher-growth period, the more worrying development is the lower CapEx from materials companies. This looks very similar to the development in the memory markets, where terms like “disciplined CapEx spend” were code for creating pricing power in shortage markets.
While material companies will not be as successful as memory companies at capturing value, there could be pain on the horizon for material customers, and there is not much room for expansion in materials.
The demand side of the Semiconductor supply network is also a story of the AI data centre, but most of the increase in demand revenue is based on cost increases carried forward.
There is a limit to how much pain the PC and Smartphones can take, but their shares of the total market are diminishing every quarter. From 45% a decade ago, PCs and Smartphones now account for 27% of the market.
While many areas of the supply network appear balanced, investments remain weak, leaving little room for the expansion now underway.
That is, if the AI revolution continues. A potential oil shock could disrupt the entire supply network and significantly impact demand.
Semiconductor fabs don’t run on oil, but several materials used in manufacturing come from petrochemical production, and an oil shortage can lead to scarcity of certain semiconductor materials.
But more importantly, an oil shock could collapse demand, popping the current profit bubble. This would not lower supply much, since semiconductor growth is primarily profit-driven, as shown below.
Since the last Semiconductor cycle peak 4 years ago, Gross Profit has grown 4.2x while COGS has grown 50%, and almost a quarter of that growth came from this quarter alone.
What is certain is that uncertainty is at an all-time high in the semiconductor industry, and caution is needed.
From manufacturing and design to supply chains and emerging technologies, the electronics industry never stands still. Meet the people driving it forward at the upcoming Evertiq Expo events. The next stop is Evertiq Expo Warsaw on October 22, where industry professionals gather to exchange knowledge, discover new solutions and build valuable business connections.
To stay up to date with the latest developments, sign up for the Newsletter.