Recently, I've been looking at the US AI stock layout and found that many people actually don't understand the division of labor in the AI industry chain. Instead of blindly chasing themes, it's better to first understand that different segments make different money, and the driving factors of stock prices are also completely different.

AI is not really an industry, but an entire supply chain. Upstream is computing hardware (like NVIDIA, TSMC), midstream is cloud platforms (Microsoft, Amazon, Google), and downstream is application software layers. The investment logic for each layer is very different.

I noticed that many people buying US AI stocks haven't actually clarified which segment they are buying. Upstream companies focus on chip supply and demand and prices, midstream looks at whether cloud capital expenditure can be recouped, and downstream considers whether companies are willing to pay for AI features. The price movements of these three layers are fundamentally out of sync.

Let's start with the upstream. NVIDIA currently accounts for about 80% to 90% of the revenue share in the AI accelerator market, earning over $100 billion annually from data center GPUs. Its moat is not just hardware but a software ecosystem built over more than ten years; developers are accustomed to programming on NVIDIA platforms, and switching costs are very high. TSMC is also key because NVIDIA, Apple, and AMD's AI chips are almost all produced at TSMC. TSMC's high-performance computing business grew 48% year-over-year in Q1 2026, and since January this year, it has continuously raised prices for processes below 5nm for four years, with AI chip prices increasing by 10%. Customers knowing that prices will rise for four years still scramble to buy, which shows the market's demand for chips.

Midstream giants like Microsoft, Amazon, and Google are spending heavily to build infrastructure. The four cloud giants' combined capital expenditure is projected to approach $600-700 billion by 2026. The question is: when will they see a return? Microsoft has already integrated AI services into its Windows, Office, and Teams ecosystems through Azure AI and Copilot, reaching 1 billion users worldwide. But Amazon's AWS is also tying itself to Anthropic, self-developing chips like Trainium to create a closed loop. Competition here is becoming more intense, and whoever can monetize quickly will be the winner.

The downstream application layer usually lags 1 to 2 quarters behind upstream. After chips are shipped, infrastructure needs to be built, and only then does revenue reflect in the application layer. Software companies like Salesforce and ServiceNow are now competing for enterprise IT budgets. Meta has taken a different route—it's itself the largest application scenario, using AI to optimize ad delivery for direct monetization, without waiting for customers to adopt.

Regarding specific targets in US AI stocks, NVIDIA is definitely the core. Wall Street expects an average growth of 79% in the first quarter of fiscal 2027, which is hard to imagine. TSMC, although less volatile, offers better stability and is suitable as a core holding in a portfolio. Microsoft and Amazon are solid, with AI being just one of their growth drivers; even if the hype cools, their core businesses can sustain them. Meta is a successful example of direct monetization through AI.

However, investing in US AI stocks now requires caution. Valuations have already risen significantly, and many companies' stock prices have long reflected growth expectations for the coming years. If growth slows or market sentiment shifts, the correction could be substantial. Short-term volatility may still occur, but the long-term trend leans toward growth. Competitors like AMD and self-developed chips (such as Google TPU) are catching up; although they are unlikely to shake NVIDIA's position in the short term, the long-term landscape is still evolving. Geopolitical risks, such as export controls, could also impact the supply chain.

A more pragmatic approach is phased deployment. Enter in batches, wait for pullbacks, and control individual stock positions. Focus on key signals: whether the pace of AI technology development is slowing, whether application monetization is improving as expected, and whether individual companies' profit growth is decelerating. As long as these conditions hold, the investment value of US AI stocks remains.

According to Gartner, global AI spending is expected to reach $2.53 trillion in 2026, climbing further to $3.33 trillion in 2027. This market is far from saturation. But it’s also important to realize that infrastructure companies, even with solid fundamentals, may be better suited for phased deployment rather than simply holding long-term without action. Historically, Cisco's stock price surged to $82 during the dot-com bubble, but after the bubble burst, it fell more than 90%. Even though its management has been good since, the stock price has not returned to the high point of that time. This reminds us to stay alert.