EP702 | 🐉

AI Usage Surges, and Companies Hit the Cost Ceiling First

The host continued his observations from the previous episode about small IC companies, pointing out that AI tools can indeed greatly improve engineers’ productivity, but companies cannot focus only on efficiency while ignoring the bill. MediaTek2454.TW was recently reported to have employees using AI extensively, with related monthly expenses potentially reaching NT$100–200 million. The company therefore tried building its own data center, but then encountered problems such as insufficient model performance, excessive server load, and slower response times, naturally prompting pushback from engineers.

This creates a dilemma between users and management. Engineers want to use the strongest models because complex programming, chip design, and research tasks can genuinely produce better results with advanced models; management, however, must answer to shareholders and the income statement. Drawing on his own experience using AI, the host explained that prolonged reasoning by an advanced model can make a single query cost several thousand New Taiwan dollars. If that level of usage is scaled across an entire company, costs can quickly spiral out of control.

The more likely solution is to deploy cutting-edge models on a company’s own servers or train dedicated models for specific tasks. This approach requires high upfront investment, but over the long term it can reduce ongoing cloud Token fees. Open-source models are suitable for more error-tolerant tasks such as text and graphic design, while high-value tasks such as chip layout and design still require stronger, more specialized models. The host’s core judgment is that whether AI can become widespread will ultimately come down to whether the cost per unit of output makes economic sense.

The Market Hits New Highs; Shorting Too Early Can Lead to a Squeeze

The host believes the market clearly strengthened this week, with overnight futures and the NQ both reaching new highs, while the broader market has yet to show signs of weakening. Before the broader market confirms a reversal, shorting too early or trying to predict the top can make someone a market focal point, as Michael Burry became during past corrections, while also exposing them to a powerful short squeeze when prices recover.

He noted that retail investors often change their views with the market: when prices fall, they look for bearish opinion leaders; after prices rise, they turn around and criticize the same people. A more prudent approach is to first observe whether the broader market has genuinely turned before deciding whether to trade against the trend, rather than interpreting every fluctuation as evidence that the trend has ended.

Passive components and power components also began to catch up this week, suggesting that different sectors are once again showing signs of broad-based strength. The host said that if the broader market is performing well while one’s holdings remain stagnant, the problem may lie in stock selection rather than in the market as a whole. AI-related components remain the clearest supply-shortage and price-increase theme; non-AI products are more like secondary beneficiaries based on the assumption that demand will be crowded out, and therefore carry relatively higher risk.

Short-Term Share-Price Volatility Does Not Mean the Industry Story Has Failed

The host rejects equating a few months of share-price performance directly with a change in industry fundamentals. The market may drive an individual stock sharply lower or higher based on a minor piece of news, but in hindsight, the news may not have affected actual supply and demand nearly as much as the share price suggested. Taking the impact of HDD capacity expansion on SSD-related stocks as an example, even a large percentage increase in capacity may not be enough to change the overall industry supply-demand balance if the original base was very low.

Investors should therefore first identify the industry trend and then assess for themselves whether short-term share prices have become overheated, rather than deciding that the original industry story was false every time prices fall sharply. The host also noted that the market sometimes magnifies small changes at a company into a major narrative, so there is no need to rush to develop a complete explanation for every short-term fluctuation.

His approach is to retain his view of the industry direction while acknowledging that an individual stock’s price can temporarily detach from fundamentals. This distinction helps prevent a short-term pullback from being mistaken for the disappearance of long-term demand, while also preventing investors from ignoring valuation and holding risk simply because the theme remains intact.

The Real Impact of Optical Communications Restrictions Is Limiting China’s Domestic Production

A report released by Morgan Stanley on October 1 discussed possible new U.S. restrictions on Chinese optical communications components, but the host emphasized that this still leans more toward rumor than an officially announced final rule from the U.S. Department of Commerce or the FCC. The report suggested that if 65% of the BOM cost of an optical transceiver module comes from the United States, the product may still be allowed to continue shipping; restrictions on 800G or 1.6T optical modules have also not clearly taken effect at this stage.