AI Investing: The Killer-App Myth, Breaking Down Stock Prices, and the SaaS Pricing Dilemma
TL;DR: AI investors need not hunt for a killer app; instead they should watch whether technology costs fall and vertical applications emerge, and check whether long-term profits and AI pricing models actually turn into revenue, because killer apps are a product of probability and only visible in hindsight.
Key ideas
- Killer apps are a product of probability, so there is no need to find them early; a more practical approach is to watch whether new technology costs fall fast enough, whether vertical applications appear in large numbers, and whether venture capital is flowing in that direction.
- A stock price can be broken into EPS times PE: a PE re-rating must rest on better long-term EPS and competitiveness, not inflated sentiment; the Google example shows that even if short-term EPS does not fall, a lowered long-term growth outlook still pulls the PE down.
- Reasoning backward from the Taiwan supply chain to US stocks is a trap: the more capacity Taiwan adds, the more margins may fall for overseas giants, and the same behavior can be read in completely opposite ways at different times.
- Big companies building their own software is often a demonstration; a bigger market is good for cloud giants, and third parties have room if they make complex features simple and intuitive; filtering out small developers from the start because 'Microsoft and Google are doing it too' filters out opportunities.
- Whether AI turns into revenue depends on pricing: the most dangerous case is costs rising with usage while revenue does not; if SaaS switches to usage-based billing, net retention falls and it faces a dilemma, while Palantir benefits because it has no subscription baggage.
Section summaries
1:20KGI Analyst Background and the Taiwan Stock Perspective
Freddy looks back on covering consumer internet at KGI from 2016 (the FANGs, later adding Microsoft); at the time Taiwan lacked dedicated US-stock analysts, and he did research for life insurance buyers because of the Cathay Life project. Taiwan analysts gained an early edge through the semiconductor supply chain and senior-junior alumni networks, while he realized he could flip this around and directly buy US end-market giants like Apple.
▶ Watch from here6:20The Capacity-Expansion Trap and Buy-Side/Sell-Side Interaction
He points out the trap of reasoning backward from the Taiwan supply chain to US stocks: the more capacity Taiwan adds, the more margins may fall for overseas giants, and the same behavior can be read in completely opposite ways at different times (such as before and after the AI boom). He also explains how buy-side and sell-side interact, and the path of his research from consumer internet deeper into infrastructure such as enterprise software, databases, and cybersecurity.
▶ Watch from here12:35Information Upgrades and Ad-Tech Details
Before 2020 there were few US-stock analysts in Taiwan, and borrowing the narrative frameworks of star foreign analysts worked well; after that, information became widespread and institutions had to upgrade. Freddy often asks Richard, a co-founder on the front line, about digital advertising technical details such as SSP and DSP; but he admits that you forget these terms if you do not use them often, just like unfamiliar company names in Taiwan stock groups, where forgetting after reading and looking up when needed is normal.
▶ Watch from here17:35Turning News into Earnings Estimates and Breaking Down PE
It is hard for investors to turn positive news such as new products or AI announcements into future earnings estimates, and most people in the industry cannot either; the market often rallies on dream-selling stories. He breaks a stock price into EPS times PE: sentiment-driven PE inflation is unreliable, and a structural re-rating should rest on better long-term earnings and competitiveness, which is what most foreign analysts are trained to do.
▶ Watch from here21:20Early-AI Déjà Vu and the Shift to the Application Layer
The AI era resembles the early days of the mainframe and iOS: highly vertically integrated, with developers still having to deal with the underlying layers, until Deepseek appeared and the focus shifted to the application layer. He favors the application layer: the golden period of triple-digit annual growth in semiconductors and infrastructure may have passed, and once the base is raised growth will certainly slow, so watch whether costs fall and vertical applications emerge.
▶ Watch from here30:05The Big-Company Demonstration Theory and Startup Positioning
Big companies building their own software is often a demonstration; if others do it better and cheaper they are happy to see it, because the overall market expands. Investors' blind spot is filtering out small developers from the start with 'Microsoft and Google are doing it too'; third parties can take share as long as they make complex features simple and intuitive. Vendors will not reinvent the wheel, so startups should focus on acquiring customers and refining service, unless like Musk they build their own data centers.
▶ Watch from here33:50Google's Self-Disruption and Breakup
Google faces a dilemma of self-disruption: search ads account for about 60% of global digital advertising, and Google holds about 90% of that, so doing AI too well means knocking itself down. He believes Google's expansion of Cloud and acquisition of Wiz are preparation for a breakup, where a Sum of Parts valuation could unlock shareholder value; he also notes that even if short-term EPS is not cut, as long as the long-term growth outlook is below average digital-ad growth, the PE will still be lowered.
▶ Watch from here41:20Zuckerberg's Praise for Databricks
Freddy says that within the framework of working in Taiwan it is not appropriate to make individual stock recommendations, but he can discuss representative companies. The host mentions seeing Databricks often lately because Zuckerberg strongly praised its CEO in a speech, guessing there may be an investment or strategic alliance; Freddy researched it as a result, having previously only known Snowflake, and notes that the line between Lake and Warehouse is increasingly blurred, so database topics are getting attention but the terminology is complicated.
▶ Watch from here43:50Warehouse vs. Datalake
He explains the different roles of the Warehouse and the Datalake: when AI first took off, the biggest myth was that Snowflake would benefit, and quite a few people lost money on it. Because Snowflake is mainly SQL, while Databricks takes a Not only SQL route that can handle a wider range of data types and workloads.
▶ Watch from here46:20Turning AI into Revenue and the Pricing Dilemma
There are two reasons AI revenue has not surged: one is fake AI products, the other is that subscriptions prevent the benefits AI brings from showing up in price or usage. The key is the pricing method, such as per-token billing or add-on modules; if AI costs rise with usage while revenue does not rise in step, the growth curves mismatch, which is the most dangerous. If SaaS switches to fully usage-based billing to protect gross margin, net retention drops immediately and large customers will not accept it, creating a dilemma; Palantir having no subscription baggage becomes an advantage, and future beneficiaries are not necessarily today's SaaS companies.
▶ Watch from here53:50IPO Experience with Taiwan Internet Companies
At KGI he wrote the first research report on Eland Information, when there was no generative AI concept yet, and its pitch was an internet thermometer, predictive analytics, and its own server room, with gross margin above 90%; a few years later, especially after 2022, the narrative changed completely. Moving toward generative AI requires new capital spending, which will certainly erode the original gross margin above 90%, though whether that is good or bad is hard to judge. He also mentions that around 2020 he accompanied the underwriting department to visit Kaitien (KGI was co-underwriter, Yuanta lead), when its Dian Dian Qian was still just a side project that had barely started.
▶ Watch from here58:50Kaitien's Pivot and the KK Group Listing
After the pandemic Kaitien struggled through a pivot, turning Dian Dian Qian (a Taiwan version of DocuSign), originally a side project, into its main business, built around OCR technology, and it can now add generative AI for verification. The conversation then turns to KKBOX: he personally thinks it should have listed 10 years ago, with the planned theme being KKBOX as a steady cash cow supporting KKStream, which does white-label streaming for Japanese telecom operators; because KKStream has heavy cloud usage, it also resells AWS services. He thinks such diversified cash flow that sustains people and teams is one of the survival methods of older Taiwanese startups, but admits he did not research this part in detail.
▶ Watch from here1:07:35Practical Takeaways on AI Tools
Lately he uses AI most for personal financial planning: after he feeds in conditions, AI still makes calculation mistakes but is directionally reasonable and suggests ways of balancing he had not thought of; he asks for the basis of its arguments and has it remove unreliable sources and reorganize. As for tools, he finds Gemini more accurate at math and relatively reliable for finance, Grok better at math and real-time information, while Claude hallucinates more in direct conversation, for reasons unknown.
▶ Watch from here1:08:50Career Transition and Listening After the Crash
Lately he treats looking after his kids as exercise; in April he was resting during a career transition, which is why he can speak freely. Many people come to chat about career experience and he helps for free, which feels like counseling; after the April market crash, quite a few laid-off people came to vent, and he has AI do a preliminary categorization before helping. Finally he recommends the timely film Conclave, which he appreciates from three angles: religion, facing uncertainty, and detective films.
▶ Watch from here1:11:20Embracing Uncertainty and Asset Allocation
Conclave is about how to face uncertainty: embrace it and hold no fixed views; Freddy extends this to investing, feeling that people who think they know everything usually get into trouble. Since 2020 he has used 100% diversified asset allocation, using ETFs covering stocks, bonds, and commodities, hedging risk with low- or negatively correlated assets, keeping only a small portion of funds for active investing such as stock picking. Because he was punished by the market in 2018 and 2020, because the bias of focusing on a single industry would blow up in 2022, and because short selling often went wrong, he finally chose to give up the active route.
▶ Watch from here