定制AI芯片正成为继英伟达GPU之后的第二个增长引擎，博通和美满正在争夺大型科技公司的设计订单。
美满的案例表明，当投资者已经消化了更快的增长预期时，强劲的业绩也可能令人失望。
博通周三的财报将检验定制芯片热潮以及投资者日益苛刻的预期。
一家公司可能报告创纪录的营收、上调展望并告知投资者其重要业务明年将增长一倍以上，但股价仍可能下跌 10%。
美满科技上周就经历了这种情况。
随着人工智能（AI）基础设施需求加速，这家芯片设计公司上调了长期营收预期。然而，美满股价在28八月收于216.62美元，下跌10.3%，因为投资者希望其与谷歌的新定制芯片合作能更早产生可观的收入。
这一反应为博通提供了有益的预演，博通将于周三（2九月）公布第三财季业绩。这也揭示出当今最重要的投资教训之一：优秀公司加上优秀业绩并不自动等同于优秀的股票回报。预期处于中间位置。
英伟达围绕图形处理器（GPU）构建了人工智能热潮，这是一种能够处理多种不同计算任务的灵活芯片。
但灵活性并不总是最经济的选择。
随着人工智能从训练模型转向为数百万用户反复运行模型（即推理），效率变得更加重要。谷歌、Meta、OpenAI和其他大型科技公司越来越希望使用针对自身工作负载设计的处理器。
可以把这想象为瑞士军刀和工厂机器之间的区别。英伟达销售的是一把异常强大的瑞士军刀。而定制硅片则是为重复高效执行单一任务而制造的机器。
这为博通和美满等公司创造了一个不断扩大的市场。它们帮助客户将自身的芯片构想转化为产品，并提供连接数据中心内数千个处理器的大量网络技术。
竞争日趋激烈。谷歌最近扩展了与美满在定制处理器、网络和内存技术方面的合作。该协议在2033财年最终可能带来高达120十亿美元的订单。重要的是，谷歌似乎是在美满之外增加博通，而非简单替代博通。
与此同时，博通正在帮助Meta开发其Meta训练与推理加速器（MTIA），最新一代预计将于9月开始生产。博通还为OpenAI开发了首款定制推理处理器Jalapeño，预计将于2026年底前开始部署。
关键点不在于定制芯片取代英伟达。大型科技公司越来越倾向于使用多种类型的处理器。人工智能半导体市场正同时变得更大和更加专业化。
季度营收达到2.74亿美元，管理层将2027财年营收展望上调至约12亿美元，2028财年预期上调至约18亿美元。定制芯片营收预计明年将翻倍以上增长。
管理层表示，谷歌的最大贡献将在2028财年到来。投资者今年早些时候已推动迈威尔股价大幅上涨，部分原因是他们预期定制硅片业务将大幅扩张。
因此，业绩虽好，但预期的变化还不够好。
相比之下，一天前Salesforce和CrowdStrike的表现更为突出。
Salesforce销售帮助企业管理客户和销售流程的软件，其上调了年度展望，并显示AI产品需求改善。其股价收于252.05美元，上涨22.6%。
网络安全公司CrowdStrike报告了创纪录的新增年度经常性收入（ARR），这是衡量新订阅业务的指标。其股价收于227.96美元，上涨20.5%。
为何反应如此不同？Salesforce和CrowdStrike在财报公布前，市场担忧AI可能颠覆软件业，强劲的业绩缓解了这种担忧。而迈威尔在财报公布前已背负巨大乐观预期，强劲的业绩只是印证了现有故事，而非锦上添花。
这就是一个财报周中的预期博弈。
博通周三将面临异常高的业绩门槛。
上个季度，其AI半导体营收达到10.8亿美元，同比增长143%。管理层预计本季度约为16亿美元，同比增长超过200%。总营收指引约为29.4亿美元。
这些数字通常看起来非同寻常。但问题在于，对于领先的AI公司而言，非同寻常已成为常态。
博通在6月份已领教过这一点。营收增长48%，但仍略低于市场预期。其16亿美元的AI预测也略低于分析师的期望。业务正在蓬勃发展，但预期门槛更高。
因此，投资者需超越博通是否“超预期”的表面，更需要关注管理层对下一阶段的表述：定制处理器放量速度、新客户是否带来可观收入，以及AI网络是否与计算需求同步增长。
博通的优势在于它可能实现“双重收费”。它既帮助设计专用处理器，又销售使大型处理器集群协同工作所需的网络和连接产品。
定制硅片仍依赖于少数大型客户。一个芯片项目的延迟、基础设施支出的减少，或客户将更多业务转给其他供应商，都可能影响数十亿美元的未来营收。
竞争也在加剧。迈威尔赢得谷歌订单表明，客户可能刻意使用多家供应商。这扩大了整体机会，但也使市场份额变得更难预测。
最后，预期至关重要。当投资者已经假设未来几年将实现非凡增长时，即使小幅延迟也可能引发股价大幅波动，而长期基本面未必显著恶化。
将业绩与预期分开。关注投资者在财报公布前的预期，而不只是营收是否增长。
关注时间表，而非交易头条。一项大合同的营收若在下个季度实现，与三年后才实现，其意义截然不同。
关注客户多元化。更多定制芯片客户可能增强增长，而供应商多元化可能加剧竞争压力。
将股价反应视为信息，而非最终裁决。股价下跌可能反映预期较低，而非业务基本面恶化。
因此，博通于2年9月发布的财报将不仅仅考验一家公司。它将检验定制硅是否正成为继英伟达GPU之后AI基础设施的又一持久支柱，以及投资者已在多大程度上将这一机遇计入预期。
美满电子已展现出乐观市场的奇怪算术：营收上升、预测上调，股价却仍可能下跌。Salesforce和CrowdStrike则展示了相反情况，即良好业绩消除重大担忧后股价上涨。
对投资者而言，这一区别至关重要。财报告诉你业务表现如何，而股价反应则告诉你这一表现与投资者已付费买入的故事相比如何。
Custom AI chips are becoming a second growth engine alongside Nvidia’s GPUs, with Broadcom and Marvell competing for Big Tech designs.
Marvell shows how strong results can disappoint when investors have already priced in even faster growth.
Broadcom’s Wednesday earnings will test both the custom-chip boom and investors’ increasingly demanding expectations.
A company can report record revenue, raise its outlook and tell investors that an important business will more than double next year. Its shares can still fall 10%.
Marvell Technology managed exactly that last week.
The chip designer raised its longer-term revenue expectations as demand for artificial intelligence (AI) infrastructure accelerated. Yet Marvell closed at 216.62 USD on 28 August, down 10.3%, because investors wanted its new Google custom-chip partnership to generate meaningful revenue sooner.
That reaction offers a useful preview for Broadcom, which reports fiscal third-quarter results on Wednesday, 2 September. It also captures one of today’s most important investing lessons: great company plus great results does not automatically equal a great stock return. Expectations sit in the middle.
Nvidia built the AI boom around the graphics processing unit (GPU), a flexible chip capable of handling many different computing tasks.
But flexibility is not always the cheapest option.
As AI moves from training models towards repeatedly running them for millions of users, known as inference, efficiency matters more. Google, Meta, OpenAI and other large technology companies increasingly want processors designed around their own workloads.
Think of the difference between a Swiss Army knife and a factory machine. Nvidia sells an exceptionally capable Swiss Army knife. Custom silicon is the machine built to perform one job repeatedly and efficiently.
This creates an expanding market for companies such as Broadcom and Marvell. They help customers turn their own chip ideas into products and also supply much of the networking technology connecting thousands of processors inside data centres.
The race is becoming more competitive. Google recently expanded its relationship with Marvell across custom processors, networking and memory technologies. The agreement could eventually generate up to 120 billion USD of orders through fiscal 2033. Importantly, Google appears to be adding Marvell alongside Broadcom rather than simply replacing Broadcom.
Broadcom, meanwhile, is helping Meta develop its Meta Training and Inference Accelerator (MTIA), with production of the latest generation expected to start in September. It has also developed OpenAI’s first custom inference processor, Jalapeño, which is expected to begin deployment by the end of 2026.
The important point is not that custom chips replace Nvidia. Big technology companies increasingly use several types of processors. The AI semiconductor market is becoming larger and more specialised at the same time.
Quarterly revenue reached 2.74 billion USD, while management lifted its fiscal 2027 revenue outlook to around 12 billion USD and fiscal 2028 expectations to around 18 billion USD. Custom-chip revenue is expected to more than double next year.
Management indicated that the largest contribution from Google would arrive from fiscal 2029. Investors had already pushed Marvell shares sharply higher this year partly because they expected custom silicon to become a much larger business.
The result was therefore good, but the change in expectations was not good enough.
Compare that with Salesforce and CrowdStrike one day earlier.
Salesforce, which sells software that helps companies manage customers and sales, raised its annual outlook and showed better demand for its AI products. Its shares closed at 252.05 USD, up 22.6%.
CrowdStrike, a cybersecurity company, reported record net new annual recurring revenue (ARR), a measure of new subscription business. Its shares closed at 227.96 USD, up 20.5%.
Why such different reactions? Salesforce and CrowdStrike entered earnings carrying fears that AI could disrupt software. Strong results reduced that fear. Marvell entered earnings carrying enormous optimism. Strong results confirmed much of the existing story rather than improving it.
That is the expectations game in one earnings week.
Broadcom faces an unusually high hurdle on Wednesday.
Last quarter, its AI semiconductor revenue reached 10.8 billion USD, up 143% from a year earlier. Management expects roughly 16 billion USD this quarter, representing growth of more than 200%. Total revenue guidance stands around 29.4 billion USD.
Those numbers would normally look extraordinary. The complication is that extraordinary has become normal for leading AI companies.
Broadcom learned this in June. Revenue grew 48%, but still came slightly below market expectations. Its 16 billion USD AI forecast also fell just short of what analysts had hoped for. The business was booming. The expectation bar was simply higher.
Investors therefore need to look beyond whether Broadcom “beats”. More important will be what management says about the next stage: how quickly custom processors ramp, whether new customers add meaningful revenue, and whether AI networking grows alongside compute.
Broadcom’s advantage is that it can potentially collect revenue twice. It helps design specialised processors, then sells networking and connectivity products needed to make large clusters of those processors work together.
Custom silicon remains tied to a relatively small number of enormous customers. A delayed chip programme, lower infrastructure spending or a customer moving more work to another supplier can shift billions of future revenue.
Competition is also increasing. Marvell’s Google win shows that customers may deliberately use several suppliers. That expands the overall opportunity, but makes market share less predictable.
Finally, expectations matter. When investors already assume years of exceptional growth, even small delays can create large share-price reactions without meaning the long-term business has suddenly deteriorated.
Separate results from expectations. Ask what investors appeared to expect before earnings, not simply whether revenue grew.
Follow timelines, not deal headlines. A large contract matters differently if revenue arrives next quarter rather than three years later.
Watch customer diversification. More custom-chip customers can strengthen growth, while supplier diversification can increase competitive pressure.
Treat share reactions as information, not verdicts. A falling stock can reflect lower expectations rather than worsening business fundamentals.
Broadcom’s earnings on 2 September will therefore test more than one company. They will test whether custom silicon is becoming a durable second pillar of AI infrastructure alongside Nvidia’s GPUs, and how much of that opportunity investors already assume.
Marvell has already shown the strange arithmetic of an optimistic market: revenue can rise, forecasts can rise and the share price can still fall. Salesforce and CrowdStrike showed the reverse when good results removed a major fear.
For investors, that distinction matters. Earnings tell you how the business performed. The share-price reaction tells you how that performance compared with the story investors had already paid for.
正在载入文档……