Memory, Part 1: From Boom-Bust Commodity to AI Bottleneck | EP 223
Mawer Investment Management Ltd.
0:00 / 0:00
Memory, Part 1: From Boom-Bust Commodity to AI Bottleneck | EP 223
792 просмотра · 1 месяц назад
Mawer Investment Management Ltd.
789 подписчиков
792 просмотра · 1 месяц назад
The memory chip shortage has become one of the defining AI stories in markets. In Part 1 of this three-part episode, Mawer equity analyst Shan Rui Yeo explains why AI has made memory (DRAM and high bandwidth memory) the key bottleneck in semiconductors. He covers how memory works, why the industry destroyed wealth for 40 years before consolidating to three players (Samsung Electronics, SK Hynix, and Micron), why inference makes memory bandwidth the limiting factor for AI, and why supply cannot catch up before 2028. With DRAM prices up 400 to 500% in a year and device makers raising prices, the conversation examines an industry whose business model is changing in real time.
0:00 - Introduction: Memory’s Breakout Year in Markets
1:46 - DRAM vs NAND Explained: Short-Term and Long-Term Memory
3:56 - The Memory Hierarchy: SRAM, HBM, DRAM, NAND
5:03 - Why Memory Chips Have Always Been Cyclical
6:56 - 40 Years of Wealth Destruction and the 2013 Consolidation
11:54 - Why AI Inference Made Memory the Key Bottleneck
15:43 - HBM Economics: The Trade Ratio and the Yield Challenge
18:07 - The Supply Crunch: Why the Shortage Widens into 2027
20:15 - Memory Prices Up 400%: What Consumers Are Already Paying
23:10 - Long-Term Agreements and a Potential Re-Rating
25:38 - Closing Thoughts
Companies Mentioned: Samsung Electronics, SK Hynix, Micron, NVIDIA, Intel, Texas Instruments, Apple, Nintendo
Key Takeaways
• Memory is a commodity with a three-to-four-year supply lag, which is why the cycle has always been difficult.
• Consolidation to three players in 2013 turned four decades of wealth destruction into at least 15% returns on capital through the cycles.
• In AI inference, memory bandwidth sets the speed of token generation, making memory the key bottleneck.
• NVIDIA’s Rubin GPU carries 384 GB of DRAM, the equivalent of 32 iPhones per GPU, or 160 million iPhones across five million GPUs.
• HBM consumes three times the wafer capacity of standard DRAM (four times with HBM4) and is forecast to absorb 30% of DRAM wafers by 2027.
• DRAM contract prices are up roughly 200% year to date and 400 to 500% year over year, and price increases are reaching phones, laptops, and consoles.
• Customers are signing three-to-five-year agreements with prepayments, which could support a re-rating of memory companies.
Host: Rob Campbell, CFA, Institutional Portfolio Manager
Guest: Shan Rui Yeo, CFA, Equity Analyst
This episode is available for download anywhere you get your podcasts.
Founded in 1974, Mawer Investment Management Ltd. (pronounced "more") is a privately owned independent investment firm managing assets for institutional and individual investors. Mawer employs over 250 people in Canada, U.S., and Singapore.
Visit us at:
/ @mawerinvestment
https://www.mawer.com
/ mawer-investment-management
/ mawerinvestmentmanagement
#ArtOfBoring #MawerInvestmentManagement #MawerInvestment #Podcast #Finance #Investing #memorychips #aiinvesting #semiconductor #skhynix #nvidia #tech