5.18.25
Back in February, I said if we have another quarter of DRAM equipment sales in the first calendar quarter of 2025 to what was seen in the calendar fourth quarter of 2024, then the top of the DRAM cycle is close, or had been reached. I made this prediction based on the top of the cycle back in late 2021/early 2022. During that period, DRAM ASPs peaked in the period from December of 2021 to January of 2022. There were three consecutive quarters of peak DRAM WFE spending: calendar Q3 of 2021 to calendar Q1 of 2022. The DRAM downturn spanned most of calendar 2022 and started to subside in late 2023. In this DRAM upturn, there has been a complicating factor: China. CXMT – the indigenous Chinese DRAM upstart – has spent at outsized amount of investment on DRAM equipment. The new bits they have produced distorted this upturn. The PC and mobile DRAM markets have been weaker than they should be in an upcycle, because of supply pressure from CXMT. Yet since the Chinese don’t make leading edge DRAM bits, the HBM market has been healthy for more than two year. In DRAM WFE, there was a surge of spending, to record levels, in the second half of 2023 into the first quarter of 2024. This was fueled by Chinese companies buying equipment ahead of sanctions. Then there was a lull in Q2 and Q3 of 2024. Spending was still healthy but was well below the levels of the China spending surge. Now spending on DRAM equipment is back to record levels. It has been nearly $4B per quarter in each of the last two quarters. That is a similar to the highest quarter ever seen, in Q4 of 2023, when the Chinese were spending ahead of sanctions. If this cycle is anything like those of the past, we are near the peak. But there is always a complicating factor to make these predictions hard. In this case, HBM demand is that factor. It may be demand for AI compute is so strong that it soaks up all this added DRAM capacity, extending the upcycle. I don’t know if this is the case.
The two year nuclear winter in the NAND market appears to be thawing. After falling to well below $1B per quarter a full six quarters, WFE spending on NAND has recovered some in the past six months, more than doubling off the bottom, though spending in the space is still half of what it was in the last upturn. NAND prices have been weakening recently, so this increase in spending is surprising. NAND fabs are running below capacity to reduce supply. Companies are taking the opportunity with idle capacity to upgrade equipment to more advanced nodes. The tension here is, more advanced nodes have lower cost per bit, but more advanced nodes also produce more bits. This trade-off can be modulated down by taking the hit to wafer output that naturally accompanies node transitions. I think this is what the NAND companies are doing. They are moving more to the leading edge, to increase learning and yields on those nodes, while accepting the lower wafer output, to control supply.
ASML doesn’t break their memory sales out between DRAM and NAND. Sales to memory customers, DRAM + NAND, were at the second highest level ever for the company. The total memory revenue of €2.411B was 13% below the record quarter, which was in Q4 of 2024. Lithography tool sales totals are lumpier than other equipment types, but I would still feel better in my short position on Micron if ASML’s memory revenue weren’t down meaningfully in the most recent period. This puts more pressure on what we see in the second calendar quarter. If it is another record, oversupply in memory is much more likely before the end of 2025. There was a surge of bookings in Q4-24, nearly double the level of the surrounding three quarters. This looks like an anomaly now – another example of a surge of sales activity ahead of a tariff deadline. While not an exact science, bookings as a leading indicator portend memory tool sales in Q2 and beyond will be flat to down.
Forecasts for overall DRAM WFE sales in calendar 2025 range from similar to 2024, to down some. One level down, the market is bifurcated. Following a strong 2024, capital expenditures by CXMT will be down in 2025. The Big Three DRAM makers will be up this year compared to last. How much? Applied quoted these sales would be up 10% to 20%. ASML said total memory sales (they don’t split out DRAM and NAND) would be similar in 2025 to 2024. Results on the DRAM equipment industry this quarter leave me less confident in a downturn by the end of 2025. Leading edge investment continues to be strong. That is good for a coming downturn. But AI demand shows no signs of slowing, so maybe HBM oversupply will be pushed out by this strong demand. China is clearly pulling back on DRAM investment; at the same time the PC and mobile DRAM markets are finding their footing. It could be that the non-AI DRAM market recovers this year while HBM and DDR5 continue their exceptional strength. From Applied, HBM is 16% of total DRAM wafer starts this year, and is growing at a 40% annual clip. . NAND has been surprising. Investment is larger than I expected in a weak market. Fabs are running below capacity, and companies are taking this idle time to upgrade equipment and move more of their NAND wafer starts to leading nodes. This is to reduce cost. If they allow wafer starts to come down with this tech migration, they may be able to continue moving the overall market to better health while also lowering their cost. This is a hard needle to thread.
– S. Hughes (short MU)