8.4.24
There are now three memory markets. DRAM is split into AI (HBM and some of DDR5) and non-AI (PC, mobile, consumer, auto, industrial, traditional server). The third is all of NAND. AI server demand is causing an increase in NAND consumption for eSSDs in new servers, but it is not affecting NAND the way it is DRAM. In the second quarter, the NAND market is the hardest to understand. There continues to be minimal investment in new NAND capacity. Micron, Samsung and Hynix all say little about NAND, instead talking mostly about HBM and DDR5. WD, the pure look investors get at the NAND market, sent mixed signals. They are forecasting ASPs to flatten in the third calendar quarter and then grow again in the fourth. It typically doesn’t work this way with ASPs. After a sharp rise in pricing, once the increases flatten, they almost always top out and start going down. But there has been almost no investment in new capacity, so demand growth must be slowing for pricing to not be rising. My best guess on NAND is that demand elasticity is resulting in lower demand, caused by the sharp price rise in the last three quarters. None of the NAND manufacturers are showing any indication they want to add capacity. In DRAM, AI demand is dominating the conversation, yet AI bits are only significant in the recent results for Hynix. Micron’s results include almost no contribution from HBM in their most recent quarter. Even in their fourth fiscal quarter, which finishes at the end of August, will be almost all non-HBM. They will have a couple hundred million in HBM revenue out of a total of nearly $8B. The recovery seen in the financials from both Micron and Samsung are almost all non-HBM DRAM and NAND price increases. It is hard to recognize this through the cloud of AI effect on Micron’s stock price. For months it was hype and strong price appreciation. Lately, the stock price has been beaten down by pessimism that the AI bubble has burst and we may be near or in a recessions. Putting aside the views of equity markets, the memory markets still appear to be in recovery and not nearing the top of an upcycle. The only investments companies are making are to migrate DRAM nodes faster and over to HBM and DDR5. Yet the strength seen in the last three quarters from DRAM pricing is almost all (ex-Hynix) from non-HBM products. As more capacity shifts over to HBM and DDR5, with only tech migration adding bits, the non-HBM DRAM market should continue to be undersupplied. Only two factors threaten that. First, a pullback in demand from a global slowdown in economic activity. While we have seen a slowing in jobs growth in the US in the past three quarters, this is off of an employment situation that was at a 50+ year high. Second, if AI investment were to slow, all those HBM bits that are coming online won’t have a place to go. This is why I am highly focused on the nature of the sales contracts between memory companies and their customers for HBM. Micron executives continue to say they are sold out through most of 2025. Even if this means only FY25 (ends in August of next year), that is still more than three full quarters of product. If indeed the volume and pricing of these sales are set, then the only variable for Micron’s results in the next five quarters is how non-HBM DRAM and NAND perform. With HBM absorbing more DRAM bits and capacity additions constrained, the setup is good for conventional DRAM to continue to see price increases. If it weren’t for the large decline in Micron’s stock price and in tech/AI stocks more broadly, there would be little indication that the memory cycle is anything but healthy and in strong expansion. Is the Market just overreacting here? Or does it have some information about a broader economic slowdown that will lead to an historically short memory cycle? If this really marks the peak of the upturn, the AI DRAM cycle will have been less than 1.5 years long and the conventional DRAM upturn will have lasted about one year. The NAND upturn will have been between three and four quarters long. All cycles are different and many outcomes are possible, but it would be really extraordinary if this downturn were to be historically short, especially following an historically brutal downturn.
– S. Hughes (cyclical long MU)