5.21.24
I’ll start with NAND because the trend over the last two years is clear for that memory. The last five quarters, starting with the first quarter of calendar 2023, have been a nuclear winter in NAND. From the fourth quarter of calendar 2020 to the fourth quarter of calendar 2022 (inclusive, so nine total quarters), spending to Applied, Lam, KLA, and TEL on NAND WFE averaged $2.94B per quarters. In the last five quarters, since CQ1 of 2023, this average was $928M per quarter. NAND WFE is down almost 70%. Pricing has recovered strongly in the last two quarters (up almost 70%). Lead time for equipment is more than six months, so if the memory companies are responding to the recent surge in NAND pricing with new equipment purchases, we won’t see it in WFE company revenue until the end of calendar 2024, at the earliest. DRAM is not nearly as clear as NAND. For the same four companies (ASML doesn’t break out sales between DRAM and NAND), WFE spending on DRAM dropped by one-third in the second calendar of 2022. This spending trended down through the first quarter of 2023 when it reached a low of $1.25B, down from a peak of $3.10B. In the second through fourth quarters of 2023, DRAM WFE spending shot back up, reaching a quarterly record of almost $4B in Q4 of 2023 before coming down some (to $3.38B) in the first calendar quarter of 2024. Outside of HBM, DRAM pricing was terrible through at least the third calendar quarter of 2023 and ASPs for non-HBM products are still not high enough for DRAM makers to justify new fab investment. This conclusion is based on their public statements. From WFE company commentary, this surge in capital expenditures on DRAM equipment is to support HBM demand and for sales to indigenous Chinese customers. The latter is rolling off in Q2 and Q3 of 2024, after which we will be left with a cleaner look at HBM equipment demand. In DRAM, I think this upcycle is actually two upcycles. The first, based on surging HBM demand, started late in the first quarter of 2023. This is the AI DRAM cycle and was led by Hynix. The second cycle is for the non-AI segments of DRAM; PCs, mobile, non-AI servers, automotive, and industrial. This cycle seems to have bottomed in the third quarter of calendar 2023. Thus, the two cycles are offset by about half a year. If this is accurate, a two-year cycle will see the DRAM peak for AI-memory in the first calendar quarter of 2025 and the rest of DRAM will peak in the third calendar quarter of that year. Of course, the two are not independent. Right now, Samsung, Hynix, and Micron are all adding advanced packaging capacity as fast as they can (hence the surge in DRAM WFE seen the last three quarters). DRAM wafers are being taken out of the supply pool for non-AI segments and added to making HBM. The trade ratio is 3-to-1, according to the memory makers, so adding HBM bits is asymmetrically lowering the supply of non-HBM DRAM. I think the 3-to-1 trade ratio is exaggerated, but it is still a significant decline in overall DRAM bit supply with a shift to more HBM in the mix. Right now, my view is the HBM DRAM upturn will run into the late spring or early summer of 2025. I think analysts and other memory industry observers are sleeping on the severity of the coming shortage in non-AI DRAM, which I believe will be serious and will extend at least until the end of calendar 2024. The image that comes to mind is the way the ocean recedes when a large wave is coming. The massive demand for HBM wafers is drawing DRAM supply out of all the other segments. The summary of commentary from WFE companies on memory is this. DRAM investment will increase in the last quarter of calendar 2024 and will continue into 2025. Investment in new NAND equipment will not increase from the historically low levels of today until early in 2025. All the DRAM makers are rushing to move as many bits from PC, mobile, and non-AI servers into HBM as they can. This rush will continue to draw bits out of all other segments. I think this will lead to a growing imbalance between supply and demand in non-HBM DRAM until at least early 2025. My question is when the HBM bit supply will catch up with demand. Micron and other DRAM companies are saying they are sold out for most of 2025, but I strongly suspect these are not take-or-pay contracts with pricing locked today. The terms are probably more favorable for memory makers than what they have been able to negotiate historically, given the strength of AI demand, but I have seen this movie before. I know it ends in severe oversupply as bit growth races past demand. As of today, I think that will happen late in the first quarter or early in the second quarter of calendar 2025. That is to say, the HBM cycle will peak in the March-April of 2025 timeframe. But there is nearly a year between now and then, so much more data to be had from the market.
-S. Hughes (cyclical long MU)