On-device flash memory to crack GPU dominance
The guest argued that 50% of AI inference queries will shift to local devices using flash memory, undermining the consensus narrative of endless demand for massive cloud data centers.
The argument
While massive GPU clusters are required for training, the guest pointed to research showing large language models can run locally on small devices using flash memory. This shift will drive a massive new demand cycle for high-bandwidth and flash memory components rather than just centralized GPUs.
The thesis, stress-tested
✓ What validates it
- ✓Increased shipment volumes of AI-optimized smartphones and PCs with expanded local memory
- ✓Earnings growth and backlog expansion at major memory manufacturers
▸ Risks discussed
- ▸Export restrictions on US memory manufacturers like Micron
- ▸Slower-than-expected adoption of AI-capable edge hardware by consumers
Hear it yourself
"The paper said we could do large language models on small devices using flash memory. I immediately went and looked and said, okay. You've got SK Hynix, Samsung on the Korean side."
00:00 / 00:11
AFFILIATE LINK · ZORTIX MAY EARN A COMMISSION · NEVER A RECOMMENDATION TO TRADE