Is AI compute demand durable? Emerging markets equity portfolio manager Wen Quan Cheong returns from a research trip across Taiwan and South Korea and examines AI chip demand, the semiconductor supply chain, and the companies with the strongest structural tailwinds. He breaks compute demand into training, alignment, and inference, explains why margins can widen further down the AI supply chain, and shares what TSMC, SK Hynix, and Samsung revealed about whether this capital cycle is different from the dot-com boom. He also explains how the team trimmed into a crowded AI trade, and makes the case for a non-AI idea: a port operator in the Philippines.
Key Takeaways
Slowing the pace of AI is partly a safety and alignment story, and alignment work is itself compute-hungry, which could offset a slower training cycle.
Compute demand splits into training, alignment, and inference; inference is skyrocketing with agentic AI, and one lab's training-to-inference mix has gone from ~80-85% training to about 50/50.
AI looks under-penetrated on tokens per user and per agent, the real driver of compute demand, even as user penetration is already high.
Research trips are a deliberate edge; early groundwork let the team act quickly on clean-room contractor Acter.
Margins can widen further down the AI supply chain thanks to consolidation and price-inelastic hyperscaler demand.
TSMC and memory makers point to a stronger foundation than dot-com and multi-year agreements that lengthen the cycle, though overbuilding is still a risk.
A non-AI idea: a Philippine port operator with defensive, price-setter characteristics and a potential Brazil concession as optionality.
0:00 - Introduction: AI Concerns and Emerging Markets
1:30 - AI Safety, Pacing, and What It Means for Chip Demand
5:03 - Three Buckets of Compute: Training, Alignment, Inference
8:23 - Is AI Demand Under-Penetrated? Users vs. Tokens
9:51 - Why Research Trips Are an Investing Edge
15:20 - Acter: The Clean-Room Bottleneck Behind the Chips
16:58 - Why Margins Grow Further Down the AI Supply Chain
18:50 - TSMC and Memory Makers: Is This Cycle Different?
22:54 - Reading the Market Signal in a Crowded AI Trade
25:48 - A Non-AI Idea: A Port Operator in the Philippines
28:12 - Outro & Subscribe
Host: Rob Campbell, CFA, Mawer Institutional Portfolio Manager
Guest: Wen Quan Cheong, CFA, Mawer Emerging Markets Equity Portfolio Manager
Visit us at:
https://www.youtube.com/@MawerInvestment
https://www.mawer.com
https://www.linkedin.com/company/mawer-investment-management/
https://www.instagram.com/mawerinvestmentmanagement/
