Recently, while reviewing the AI industry chain, I suddenly realized a trap that many people have stepped into—blindly chasing AI themes without understanding what they are actually investing in.

Rather than saying AI is an industry, it’s better described as an entire supply chain. From downstream chip manufacturing, midstream cloud platforms, to upstream application software, each link earns different profits, and the driving logic of stock prices is completely different. I found that many people just don’t understand this structure, which is why they often stumble when it comes to AI stock recommendations.

Let’s start with the upstream. NVIDIA, this company, now basically monopolizes the AI accelerator market, with a market share between 80-90%. Just data center GPUs generate over $100 billion in revenue annually. But NVIDIA’s moat isn’t just the hardware itself; it’s the software ecosystem built over more than a decade—millions of developers are accustomed to programming on NVIDIA platforms, and switching costs are extremely high.

TSMC is also very critical. Chips from NVIDIA, Apple, AMD are basically produced using TSMC’s advanced processes. Earlier this year, TSMC announced a four-year continuous price increase for processes below 5 nanometers, with AI chips seeing a 10% rise. Customers, knowing prices will keep rising, still rush to order. This reflects a reality: the demand for AI chips is extremely hot.

In the midstream, cloud giants like Microsoft, Amazon, and Google follow a different logic. They don’t sell chips but offer computing power services and model APIs. Microsoft, due to its exclusive partnership with OpenAI, deeply integrates Copilot into Windows, Office, Teams—products with hundreds of millions of users—gradually unlocking monetization potential. Amazon, through AWS, partners with Anthropic and also supplies its own chips, Trainium, forming a complete closed loop. The growth certainty of these midstream companies is actually more stable than upstream.

What about the downstream application layer? Companies like Salesforce and ServiceNow are embedding AI capabilities into their products. Meta directly monetizes through AI-optimized advertising; the improved targeting accuracy on Facebook and Instagram, thanks to AI, directly reflects in revenue. The downstream usually lags 1-2 quarters behind upstream because it takes time for infrastructure investments to translate into application-level results.

If I were to recommend AI stocks, I’d categorize them into three risk levels.

For more stability? TSMC, Microsoft, and Amazon are good choices. These companies have solid fundamentals; AI is just one of their growth drivers. Even if the hype cools down, their core businesses can still support them.

For following mainstream capital? NVIDIA and Meta Platforms. Both are highly tied to AI, with strong growth momentum, but also higher volatility. Suitable for those who can tolerate some fluctuations and are willing to hold long-term.

For high risk and high reward? Look at second-tier AI chip manufacturers and application startups. They offer the greatest flexibility but also carry the highest risks.

In Taiwan stocks, TSMC remains the cornerstone. The 2nm process and CoWoS advanced packaging have become industry standards. TSMC holds long-term technological leadership and stable pricing power. Hon Hai, as the world’s largest electronics manufacturer, is a major server manufacturer for NVIDIA, but recently its stock has weakened mainly because the gross margin improvement didn’t meet expectations. MediaTek has laid out in edge AI and mobile platforms, with the Dimensity series now featuring integrated enhanced AI computing units.

Regarding investment risks, I think there are several points to be especially cautious about.

First is valuation. The AI sector has surged significantly over the past two years, and many companies’ stock prices already reflect years of growth expectations. Once growth slows, the correction could be substantial.

Second is competition. AMD is catching up, Google’s self-developed TPU is also advancing, and the long-term competitive landscape is still evolving. In the short term, NVIDIA’s dominance is hard to shake, but not forever.

Third is capital rotation. The market might shift from hardware to software, or completely move away from AI to other themes. Sticking rigidly to a single sector can cause missed opportunities.

Fourth is geopolitical and regulatory risks. Export controls can impact supply chains, and increasing regulation on data privacy and algorithm bias will become more stringent.

In the long run, AI’s impact on human life and production methods will be no less than the internet revolution, creating enormous economic value. But this doesn’t mean all AI-related stocks are worth holding long-term. Looking back at the internet era, Cisco’s stock soared to $82 during the 2000 dot-com bubble peak, only to fall more than 90% afterward. Even with solid fundamentals, the stock price has never returned to that high. This historical lesson reminds us that infrastructure-type companies, even if well-managed, are more suitable for phased positioning rather than holding on blindly.

A more pragmatic approach, I believe, is to adopt a phased investment mindset. Keep an eye on key factors: whether the development speed of AI technology begins to slow, whether the monetization of related applications improves as expected, and whether individual companies’ profit growth shows signs of deceleration. As long as these conditions hold, the investment value of AI stocks remains supported.

In the short term, there may still be volatility, but the long-term trend should be upward. Staggered positioning, waiting for dips, and controlling single-stock exposure are still essential basic skills.