EP695 | 🍊

Volatile markets punish chasing strength and day trading most easily

The host believes Taiwan stocks have remained range-bound for more than a month, unlike the previous environment in which stocks would rise soon after being bought. Stocks that hit their limit-up price or showed strong momentum the previous day may open sharply lower the next day, leaving short-term chasers to absorb the selling pressure. Even if the price recovers a few days later, rushing in can still mean taking an initial loss.

These repeated trading costs may not be large on any single occasion, but they can accumulate over multiple trades and erode net worth by 5% to more than 10%. In this kind of market, the host’s approach is to maintain the original strategy while reducing position sizes, or temporarily chase strength less frequently. Even if you really want to buy, you can wait for the price to pull back; missing one entry does not mean missing every opportunity, because the market will continue to offer new entry points.

The host still believes that if the broader market truly breaks to a new high and confidence is rebuilt, the trading environment for theme stocks will improve. Until a breakout is confirmed, the focus should not be on finding the hottest theme every day, but on first limiting the damage caused by volatility and waiting for the market to show its hand.

Google’s AI infrastructure is already showing a rapid payback

Comments by the head of Google$GOOGL Cloud at a Goldman Sachs conference became the central market observation of this episode. The host relayed that Google’s pace of acquiring new customers is about twice what it was a year ago, while deals worth more than $100 million have grown by more than twofold. Customers who initially committed to spend $100 often end up spending more than $150, indicating that cloud demand is not only present but also significantly exceeding initial commitments.

Each of Google’s major products now generates more than $1 billion in annual revenue, while revenue, operating profit, and market share continue to improve. More importantly, the payback period for AI servers is less than two years, while Google’s internally developed ASICs and TPUs can pay back in roughly one year. This means AI infrastructure may no longer be a business where money is invested upfront and profitability remains uncertain for years. Instead, it may generate cash returns relatively quickly after the capital is deployed.

Therefore, concerns about Google’s declining free cash flow and expanding capital expenditures should not automatically be viewed as financial deterioration. If new servers and TPUs can pay back in one or two years, management may be allocating capital to high-return projects more efficiently than it could through share buybacks. Only when the returns on these investments decline and Google resumes buybacks might that become a signal that demand and capital efficiency are weakening.

From the model race to infrastructure returns

The host distinguishes between two ways the market may view the industry: frontier models can command higher valuation multiples because investors are still imagining how they might change the future; but when a company gradually becomes a provider of computing power and cloud infrastructure, its valuation multiple may be lower even though its revenue is easier to measure through actual contracts and cash flow. This may explain why Warren Buffett would value the measurable returns of Google’s infrastructure business rather than simply betting on an abstract vision of technology.

The host does not accept Google’s claims about training efficiency, inference efficiency, and cost advantages without question, since every company emphasizes its own benchmark tests when releasing models and hardware. However, the market has repeatedly shown recently that the rankings of frontier models can change quickly: just when everyone thinks one company has opened up a clear lead, a competitor may release a new version within months that catches up or even surpasses it. It is therefore unwise to treat any one model as a long-term winner too early.

Differences in model capabilities will still exist—for example, certain models may be more popular for software development or specific tasks—but the premium charged for API access is gradually converging. As the prices customers pay for models become increasingly similar, cloud customers have more incentive to adopt hardware with better cost efficiency. This change will make dedicated chips such as TPUs more attractive and may also give Google a stronger competitive position by integrating its cloud, chips, and services.