Recently, I’ve been looking into which AI stocks are worth paying attention to, and I found that this space is far more complex than I imagined. A lot of people blindly chase the AI concept, only to end up overwhelmed by overvaluations and big swings. Instead of doing that, it’s better to first understand the AI industry chain before making a move.

AI is not a single industry, but an entire supply chain. From upstream computing power hardware, to midstream cloud platforms, to downstream application software—each layer has different logic for making money. I spent some time sorting it out, and I found that understanding the industry structure is the key to stock selection.

Upstream is the computing power segment. NVIDIA currently dominates 80–90% of the AI accelerator market. Just data center GPUs alone can generate more than $10 billion in revenue every year. Its moat isn’t only in the hardware itself, but also in the software ecosystem built over more than a decade—developers have become accustomed to NVIDIA’s platform, and the switching costs are genuinely staggering. This is something competitors find extremely difficult to replicate.

TSMC is also worth watching. NVIDIA chips, Apple’s processors, and AMD’s server chips are almost all produced by TSMC. Starting in January this year, TSMC began raising prices for processes below 5 nanometers continuously for four years. AI chips are seeing a 10% increase. Customers know prices will keep rising for four years, yet they’re still rushing to buy—this shows just how tight supply is. JPMorgan estimates that TSMC’s U.S. dollar-denominated revenue this year could grow by 35–40%.

The midstream is made up of cloud giants like Microsoft, Amazon, and Google. Microsoft has successfully embedded AI seamlessly into global enterprise workflows, thanks to its exclusive partnership with OpenAI, its Azure platform, and Copilot integration. Its monetization ability is being continuously unlocked. Amazon, meanwhile, deeply binds Anthropic through AWS and its self-developed chip Trainium, forming a complete business closed loop.

In the downstream applications space, Meta is arguably the most direct beneficiary. Advertising on Facebook and Instagram has improved significantly in precision due to AI optimization, which is directly reflected in revenue. This is a successful example of “AI directly monetizing.”

Back in Taiwan stocks, TSMC (2330) is still the core holding. It has advanced processes like 2 nanometers. No matter who wins the AI race, all high-performance chips still have to be produced by it. Growth at this layer is relatively stable, which makes it suitable as a foundation for an investment portfolio.

Hon Hai (2317) is the leading contract manufacturer for complete systems, and NVIDIA’s servers are mainly assembled by it. But to be frank, this year’s stock performance hasn’t been ideal—mainly because the increase in gross margin has been far lower than expected. Its cooperation with NVIDIA is still at the contract manufacturing level, and the technological added value hasn’t made a qualitative leap yet.

Don’t overlook cooling and power either. As AI servers consume more and more power, liquid-cooling solutions have become a must. Companies like Qichong (3017) and Shuanghong (3324) are in a clearly defined technological turning point—so long as power consumption keeps pushing higher, their earnings leverage may have a chance to be amplified.

Honestly, if you don’t want to take on too much volatility, stable companies like Microsoft, Amazon, and TSMC are good options. AI is only one of their growth engines. Even if the hype cools down, their core business can still support the stock price.

If you want to catch mainstream capital flowing into AI trends, NVIDIA and Meta Platforms are top choices. These companies are highly tied to AI, with strong growth momentum—though their volatility is also relatively high. They’re suitable for people who can tolerate some fluctuations and are willing to hold long term.

However, it’s important to note that this year, valuations for AI concept stocks have already been clearly bid up. Risks in this sector are not small either: intensifying competition (AMD and in-house chips playing catch-up), overvaluation (the stock prices already reflect years of growth expectations), capital rotation (which may move from AI to other themes), and geopolitics (export controls impacting supply chains).

From a historical perspective, back when the dot-com bubble peaked, Cisco’s stock price surged to $82, but later fell by more than 90%. Even though it performed well afterward, the stock price still hasn’t returned to its prior high. This reminds us that for infrastructure-type companies—even with solid fundamentals—stock prices are more suited to phased positioning than simply holding forever.

A more practical approach is to adopt a phased investment mindset: build positions in batches, wait for pullbacks, control position sizes in any single stock, and continuously pay attention to key factors such as how fast AI technology is developing, how well applications can be monetized, and the rate of company profit growth. As long as these conditions still hold, the investment value of AI stocks can continue to receive support from the market.

Finally, AI is definitely a long-term trend, but short-term stock price volatility really is huge. Instead of blindly chasing themes, it’s better to first figure out which part of the industry chain you want to invest in, and then choose suitable targets based on your own risk tolerance.