8.21.25
DRAM pricing bottomed over three quarters back in calendar 2023. For Hynix, the bottom was the first quarters. For Samsung, it was the second quarter. For Micron, it was the third quarter. Since those three troughs, Hynix’s ASPs are up 150%. Samsungs are up 110%. Micron’s are up 90%. Subscale Nanya has seen their blended DRAM ASPs rise less than 10%. For the Big Three, the expansion in pricing has been steady. In the first half of calendar 2025, all three big DRAM companies have seen their blended prices flatten. Normally, this would mark the top of the cycle. Except that these two quarters of weakness are believed by the memory maker management teams to be inventory digestion at customers. They believe the rest of the calendar year will be “constructive,” in the words of Micron’s executives. The other leadership teams echo that sentiment.
DRAM bit growth was stunning this quarter, in the teens sequentially for Samsung and over 20% for Hynix and Micron. Snapping the line at the first calendar quarter of 2023, the DRAM bit supply from the big four has grown at a mid-teens percent CAGR for the last nine quarters. That excludes supply growth from CXMT. Mid-teens is the long-term average for demand growth, but this upturn has not been average. Micron believes industry bit demand will grow high-teens for the rest of calendar 2025. This bit growth is lower than I thought it would be. I think the reason is the shifting of bits over to HBM and the trade ratio associated with that. Considering the factors of both pricing and bit shipments in DRAM, the market seems to be close to balanced. The surge in bit shipments this quarter didn’t also have a surge in pricing. Customers were willing to buy more bits but not pay more for them. This could be inventory restocking ahead of potential tariffs. I think the DRAM market is close to balanced right now, and shifting into undersupply. Normally, that indicates the market is going to rise for at least a year, but this downturn hasn’t been normal, because of CXMT and the emergence of HBM.
On that topic, this quarter is the first one with a sign that the supply of HBM has caught up with demand. Specifically, Samsung said exactly that. When a market peaks and turns over, the signs are bright-light-clear only in hindsight. I want to see this as the first indicator that the HBM market has peaked, because of my short position. I am acutely aware of that bias. Micron commented that HBM pricing is “stable.” Neither Micron nor Hynix talked about how they have negotiated pricing and volumes for the rest of the year and for 2026 as they did over the last two years. This absence of commentary may mean the HBM market isn’t as strong as it was a couple quarters ago. Or it may not mean anything. The HBM market is forecasted to reach $35B in total sales in 2025. Micron is today shipping HBM at an annual run rate of $7B. Hynix plans to double their HBM revenue in 2025 over 2024. Samsung is planning to increase year-over-year by 2.5x. More on HBM below.
Unlike DRAM, NAND is not a healthy market. Because the NAND space hasn’t been bifurcated into the HBM and non-HBM markets, it is also easier to understand. The bottom of the downturn for NAND was sometime in the third calendar quarter of 2023. From there, pricing rose for four to five quarters and about doubling off the bottom. Since then, ASPs have declined for the last three quarters. Upturns usually last two years. For NAND, the most recent upturn was only one year in duration. The additional supply brought online by YMTC is responsible for that. Today, I believe the NAND market is oversupplied, but not severely so. The industry shipped a massive surge of bits in the most recent quarter, yet prices still dropped. Customers are willing to buy but won’t pay more to do so. They are still in charge.
What does the future hold? Here is what I think. I believe all the capacity that has been added in HBM – either by switching non-HBM wafer starts over or by adding new wafers – has finally caught up with supply. Besides Samsung saying so explicitly, I have read reports that it is easier to secure HBM cubes than it was six months ago. The absence of positive HBM commentary from Micron and Hynix also makes me think the companies are starting to see the signs of weakness. Within their factories, I think yields are rising faster than planned, meaning more bit supply. So much so that DRAM companies have slowed their delivery schedule of new HBM equipment. I think Micron will give weaker than expected HBM guidance in their December call because of oversupply.
So what happens to the non-AI DRAM market? This one I have less confidence in my prediction. This quarter’s surge in bit shipments didn’t come with much higher pricing, so I don’t think the market is that far from balanced. It is undersupplied, but not by a lot. CXMT is still out there, adding bits to the market. And the Big Three DRAM makers are adding capacity to the leading edge, a lot of which is non-HBM wafers. And when HBM does go into oversupply, that trade ratio of 3:1 will run in reverse. That is, as HBM wafers are switched back over to non-HBM DRAM products, they will do so by creating three wafers for every one HBM wafer. That will help the supply-demand in the HBM market but will be damaging to the rest of the DRAM market. That will take a full cycle to turn over, at least. If the oversupply is clear enough by Thanksgiving that Micron, Hynix, and Samsung start switching wafers away from HBM, those starts won’t become bit supply until sometime in March of 2026. What I see right now in both the AI and non-AI DRAM markets is undersupply, but not far from balance.
NAND going forward will not become the tragedy I saw it headed for a quarter ago. What I see now is a market that can absorb a lot of bits, which is good for suppliers. But they have derated their fabs to bring down bit growth. I see they have decided to move their tech nodes forward faster during this derate time. I believe they are doing that to keep or widen their lead with YMTC, who is still ramping their fabs and growing their bits. The increase we have seen recently in NAND WFE purchases is opportunistic and defensive. The four leading NAND manufacturers are stepping on the node migration gas. This will help them defensively against YMTC, but it also puts more bits into a market that is already on the edge of oversupply. I don’t think NAND prices will decline sharply from here, but I do think it will be bumpy for the rest of 2025.
– S. Hughes (short MU)