The AI trade is getting more selective and shifting. Semiconductors are less favored right now, with attention moving back to software.
Strong Earnings, Rough Macro
Company earnings in AI have been very strong. Dell (DELL) beat, topping even the whisper number. Nvidia (NVDA) posted a great quarter and beat. SEC also beat. The underlying results are solid.
The problem sits in the macro picture, full of headwinds. Oil is over 90. The 30-year yield is high. Interest rates are rising worldwide. That makes it hard for these companies to get credit for how durable and strong their earnings are. It works like a fast race that keeps getting interrupted - the 2011 Canadian Grand Prix, the longest race on record at about 4.5 hours, ran the fastest cars ever going in, but the safety car came out six times. The AI trade faces the same kind of repeated stops.
Memory and the Misread Headline
One clear opportunity is memory, an area with large positions held. There are still many wrong ideas about memory, especially high bandwidth memory (HBM).
Nvidia (NVDA) made an architecture choice on its Rubin Ultra to cut the memory stack height from 12 to 8. The headline read as 33% less high bandwidth memory and less memory sold. That reading is wrong. The total number of bits, the actual amount of memory, stays the same. The bits are fixed, and Nvidia can sell more GPUs. For SK Hynix, the effect is relatively unchanged.
The part people miss: stacking memory 12 high is very hard. Dropping to 8 high gives better yield. So the fixed bit count stays about the same, but memory makers may end up with better yield and a small price increase on the bits they sell.
Networking, Data Movement, and Integrity
People often treat the slices of the AI trade as zero sum. Some architecture changes do create winners and losers, and memory and networking fall into that bucket.
Memory matters, but when there is less memory locally to store data, you have to move the data instead. The next phase depends on moving that data as fast as possible and with the highest integrity. This matters because the slowest link in a GPU cluster holds back the whole task during data movement. That is expensive - the equipment is already bought and the compute costs a lot. You want the GPUs working together as smoothly as possible, so data integrity becomes very important once you reduce the memory count.
Some interesting names fit here. SEC sits in the analog integrated circuit (IC) component space that helps move data fast while keeping it high integrity.
