Will be back to the US next week.
Talked to a few PMs during my trip in the last two weeks. I’d like to share some interesting findings:
Sentiment: Overall very panicked and confused.
Positioning:
1. Shanghai HFs are extremely long on memory names, partly due to the TRS restriction I mentioned a few weeks ago. They are quickly selling CN semis across the board and buying DRAM ETFs to lower margin pressure.
2. HK HFs are mixed and divided; bulls and bears are balanced.
3. US LOs are waiting—not about stock prices but still don’t believe LTAs can turn memory names from cyclical to growth valuations.
4. Macro HFs sold all of their memory/semi positions.
Interesting points from both sides:
Bulls:
1. Supply chain checks suggest prices and margins will remain higher for longer.
2. AI capex in 27/28 will grow much higher than most expected.
3. OAI/ATH’s ARR shows no signs of plateauing. Codex’s WAU is growing crazily right now.
4. Kimi K3’s tech report literally suggests every semi component is a bottleneck 😂
5. No macro risks; CPI/PPI much cooler.
6. Even if memory spot prices can’t go higher, these memory companies can do large buybacks or dividends to support the stock price.
7. OAI/ATH is hiring a large team of solution engineers to quickly build vertical solutions (finance/healthcare/legal) to find the next S-curve beyond coding.
8. Even if memory names fade, the entire semi sector will remain intact.
Bears:
1. Higher memory prices are not sustainable and will be demand destructive. Clients will push back on further price hikes, and even the government may step in to force massive production expansion.
2. Memory prices will peak in Q4.
3. No doubt on 27 capex, but 28 capex budgets are clouded.
4. Even if fundamentals are solid, the position is extremely crowded and Korean retailers’ leverage is a huge concern.
5. Even though CSPs’ capex planning is aggressive for 27, a large portion of deliveries will be delayed due to power supply constraints.
6. No intermediate story of fast AI application adoption like AI coding.
7. Higher prices and more LTAs during the upcycle can’t prove memory names aren’t cyclical anymore. We need to see resilience during the downcycle to prove that, like memory names keeping the lower bound of margins in a downturn.
8. Memory is the core of AI: if traded memory dies, then everything dies.
9. Memory has no difference vs. other commodities such as silver, oil, or lithium carbonate. When they peak, their narrative and fundamentals all look fucking good 😄