Recently, I have been observing the investment logic of AI stocks and found that many people are blindly chasing themes, actually not understanding the division of labor across the entire supply chain. Instead of shooting in the dark, it's better to first understand the three-tier structure of the AI industry chain, so you can truly judge which AI stocks are worth paying attention to.

Let me start with a core understanding: AI is not a single industry, but an entire supply chain. From upstream hardware for computing power, midstream cloud platforms, to downstream application software, each layer's driving logic is completely different, and the key variables affecting stock prices also vary.

Upstream mainly includes NVIDIA, TSMC, Foxconn—companies providing GPUs and AI chips. This layer is directly driven by AI chip supply and demand, and capital expenditure (Capex) from cloud giants. I noticed that the combined Capex of the four major cloud giants by 2026 is approaching 600 to 700 billion USD, which provides strong support for upstream AI stocks. Especially TSMC’s advanced CoWoS packaging capacity has become a bottleneck in the entire AI ecosystem. TSMC will start continuous price increases for all processes below 5nm from January 2026, with AI chip prices rising by 10%. Customers know they will face price hikes for four years but still rush to place orders, indicating high demand.

Midstream consists of cloud giants like Microsoft, Amazon, and Alphabet. They do not sell chips directly but offer computing power services and model APIs. The key here is the growth rate of AI service revenue and capital return cycles. The market is increasingly concerned with "when will the investment pay off," and if analysts question the return rate, midstream stocks may face pressure. An interesting phenomenon is that some cloud customers are starting to develop their own chips (such as Google TPU, Amazon Trainium) to reduce costs, which could impact the profit structure of midstream stocks in the long term.

Downstream includes application software companies like Salesforce, ServiceNow, and Adobe. They embed AI capabilities into their products, mainly focusing on enterprise adoption speed and the premium rate for AI features. Downstream stocks usually lag 1 to 2 quarters behind upstream stocks because after chip shipments, it takes time to build infrastructure.

Regarding specific targets, NVIDIA currently accounts for about 80 to 90% of the AI accelerator market revenue share, generating over 100 billion USD annually just from data center GPUs. Its moat is not only in hardware; a software ecosystem built over more than a decade has made millions of developers accustomed to programming on NVIDIA platforms, making switching costs extremely high. TSMC is the production bottleneck for all AI chips; whether NVIDIA, Apple, or AMD chips, almost all are manufactured by TSMC. JPMorgan estimates that TSMC’s revenue growth in USD terms for 2026 and 2027 will be 35% and 30%, respectively.

Microsoft is a leading platform for enterprise AI transformation, successfully integrating AI into global enterprise workflows through Azure and Copilot. Amazon, through AWS and its self-developed chip Trainium, deeply binds into the AI ecosystem, forming a complete closed loop. Meta is a representative of AI application layers, monetizing AI directly through ad AI optimization and the Llama open-source model.

In Taiwan stocks, TSMC is naturally the first choice; process technology and advanced packaging are industry standards that cannot be replaced. Foxconn, as NVIDIA’s main server manufacturer, saw its stock price weaken early 2026, as market patience wears thin, mainly because its gross margin improvement was much lower than expected. MediaTek has also made some layouts in edge AI and mobile AI chips.

If you want to participate in AI but avoid too much volatility, companies like Microsoft, Amazon, and TSMC are solid and AI is just one of their growth drivers. To capture mainstream AI capital flows, NVIDIA and Meta Platforms have strong growth momentum but are more volatile. Those willing to accept higher risks for explosive opportunities can look at second-tier AI chip companies.

Frankly speaking, valuations for AI stocks are already significantly elevated. By early 2026, Alphabet’s market cap once surpassed 4 trillion USD, reflecting market expectations for AI-driven revenue and competitive advantages. But short-term volatility is inevitable, especially with macroeconomic changes, interest rate adjustments, capital rotation, and other factors that could cause sharp corrections.

From a long-term perspective, AI’s impact on human life and production modes will be no less than the internet revolution, creating enormous economic value. However, historical experience shows that even foundational infrastructure giants like Cisco saw their stock price fall from $82 to over $8 during the dot-com bubble, and it took more than twenty years to recover. This reminds us that infrastructure-type AI stocks, even with solid fundamentals, are more suitable for phased positioning rather than holding indefinitely.

A more pragmatic approach is to adopt a phased investment mindset. Continuously monitor whether the pace of AI technology development slows, whether application monetization improves as expected, and whether individual companies’ profit growth slows. As long as these conditions remain, the investment value of AI stocks can continue to be supported by the market. Of course, be aware of risks such as overvaluation, increased competition (AMD and in-house chips potentially eroding market share), capital rotation into other themes, and geopolitical influences on supply chains.

My current view is that AI stocks will show a "long-term bullish, short-term volatile" pattern from 2026 to 2030. The most prudent approach is to stagger investments, wait for dips, control individual stock positions, and diversify through AI ETFs to effectively reduce risk. Short-term fluctuations are certain, but the long-term trend remains toward growth.