Recently, there has been increasing discussion about AI concept stocks, but most people actually don't understand what they are investing in. Instead of blindly following the trend, it's better to first understand the logic of the AI industry chain—because AI is not fundamentally an industry, but an entire supply chain, with different segments earning completely different profits.

I’ve noticed that the market’s perception of AI stocks still remains at the stage of "buy chip companies to make money," but in reality, this industry is divided very finely. Upstream are computing hardware companies (NVIDIA, TSMC), midstream are cloud platforms (Microsoft, Amazon, Google), and downstream are application software layers (Salesforce, Adobe). The driving logic of each layer is completely different, and the triggers for stock price movements are also different.

Let's start with the upstream. NVIDIA currently dominates about 80-90% of the AI accelerator market revenue share, generating over $100 billion annually just from data center GPUs. But NVIDIA’s moat isn’t just in hardware—the software ecosystem it has built over more than ten years makes millions of developers accustomed to programming on its platform, creating extremely high switching costs. This is the real hard-to-copy advantage.

TSMC is also worth paying attention to. NVIDIA’s chips, Apple’s processors, and AMD’s server chips are almost all produced using TSMC’s advanced process technology. Recently, TSMC has increased prices for processes below 5 nanometers for four consecutive years, with AI chips seeing a 10% increase. Customers know they will face price hikes for four years but still rush to order. What does this indicate? It shows how urgent the demand for AI chips is.

The logic faced by midstream giants like Microsoft, Amazon, and Google is different. They don’t sell chips directly but offer computing power services and model APIs. Microsoft, through its exclusive partnership with OpenAI, deeply integrates Copilot into Windows, Office, Teams—products used by 1 billion users worldwide—creating continuous monetization. Amazon, via AWS, partners with AI companies like Anthropic, providing cloud services and self-developed chips, forming a complete closed loop.

Interestingly, the relationship between midstream and upstream is somewhat delicate. If upstream companies’ prices rise too sharply, it increases costs for midstream companies—such as NVIDIA’s gross margin reaching 75%, prompting some cloud clients to start developing their own chips (Google TPU, Amazon Trainium) to reduce costs. This could have a long-term impact on the profit structure of midstream companies.

Meta is different. It represents the AI application layer, directly monetizing through AI-optimized advertising. The targeting accuracy of Facebook and Instagram has greatly improved thanks to AI, directly boosting revenue. Meta doesn’t need to wait for customers to adopt AI; it is itself the largest application scenario.

In Taiwan, AI concept stocks also have opportunities. TSMC, as a global leader in advanced process technology, holds an unshakable position. Hon Hai (Foxconn), as a major manufacturer of NVIDIA’s servers, has recently shown weak stock performance, mainly because its gross margin increase fell far short of market expectations. MediaTek is actively pursuing edge AI and automotive AI, collaborating with NVIDIA to develop related solutions.

If you want to participate in AI but avoid too much volatility, Microsoft, Amazon, and TSMC are relatively stable choices. These companies have strong fundamentals, with AI being just one of their growth drivers; even if the hype cools down, their core businesses can sustain them. For those looking to capture mainstream capital, NVIDIA and Meta are options—they are highly tied to AI, with strong growth but also higher volatility.

Honestly, the valuations of AI concept stocks have already risen significantly recently, and market sentiment shifts could lead to sharp corrections in some major stocks. Looking back at the internet era, Cisco, the “first internet equipment stock,” saw its stock price surge to $82 during the 2000 bubble peak, but then fell more than 90%. Even though Cisco later performed well, its stock price has yet to return to those highs. This reminds us that infrastructure-type companies, even with solid fundamentals, may be more suitable for phased positioning rather than long-term holding without adjustment.

Therefore, 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 down. As long as these conditions hold, the investment value of AI concept stocks can continue to be supported by the market. Short-term fluctuations may still occur, but the long-term trend leans toward growth.