5.22.25
Looking back over the last eight quarters, which is approximately the time since HBM first emerged as a meaningful product category, the total bit supply in the DRAM industry, excluding CXMT, has grown by about 50%. Over that time, Hynix has gained the most share, having increased their bit supply by a total of ~61%. Micron is up almost 50% in that period and Samsung has lagged, growing their bit supply 41%. Nanya’s market share continues to be eroded away. I don’t have any data on CXMT’s bit growth over this time. From what Micron said two calls ago, the Chinese DRAM company’s market share is around 5%. It is reasonable to assume that share was effectively zero two years ago. Thus, CXMT has added bit supply equivalent to roughly 20% of Micron’s total. Tightening the lens to the last twelve months, Hynix has grown bit output 15%. Samsung is up 4%, Micron is down 4%, and Nanya is down 26%. The overall DRAM industry has thus undersupplied the underlying end-demand growth for DRAM bits, yet pricing has still struggled. The supply difference has been made up by CXMT and inventory draw downs.
I led with the supply situation because it determines pricing. It was two full years ago – in the first calendar quarter of 2025 – that all four DRAM makers last experience blended declines in their ASPs. In the following quarter, Q2 of 2025, Hynix’s fortunes turned upward with the emergence of HBM. This was the start of the AI upturn. Unlike a normal upturn, only one company – Hynix – was participating. It wasn’t until the fourth quarter of 2023 that the rest of the DRAM industry had entered the upturn. Since then, we have seen six quarters of rising DRAM prices, with two exceptions. One, Nanya saw their blended prices drip in Q4 of 2024. They have no HBM exposure and are thus a bellwether for the non-AI DRAM market. New supply from CXMT shortened the upturn for PC and mobile DRAM. Nanya saw their blended ASPs drop 11% in the fourth quarter of 2024. Q1 of 2025 was flat to slightly down for the DRAM industry overall. The Big Three saw Micron up 5%, Hynix flat, and Samsung down 5%. Nanya was down just 1%. The companies are forecasting non-AI DRAM pricing to be bottoming now (2Q of 2025) and predict the market will get healthier through the rest of 2025. I have been highly skeptical of this, because memory markets historically have not recovered that quickly. But this cycle is different in that DRAM has bifurcated into AI and non-AI memory. Plus, CXMT has entered the marketplace, the first new DRAM company in more than twenty-five years. Their supply prematurely shortened the non-AI upturn, which caused the other manufacturers to pull wafers away from mobile, PC and consumer DRAM production. If CXMT has slowed their bit growth, then the non-AI DRAM segments could see rising prices this year.
Switching over to NAND, bit shipment growth turned positive for all four (ex-China) manufacturers starting in the second calendar quarter of 2023. In the two subsequent quarters, output from Micron, Hynix, and SanDisk (formally WD) all surged. Micron’s bits almost doubled in two quarters and the other two increased a combined 50% or more. Samsung was the laggard of the group, clocking only 8% more bits in those six months. Almost as crazy as that massive surge is the pause which has followed. Since the third quarter of 2023, NAND bit output across those four companies has actually declined. Micron’s output is the same today as it was eighteen months ago. Hynix’s output declined so much it has returned to the same level as Q1 of 2023. In the last eighteen months, SanDisk’s output is down about 5% and Samsung’s is up about 13%.
Despite this investment discipline by the NAND incumbents, NAND ASPs have been dropping for the last six months. For all but SanDisk, prices were up between 87% and 1146% for Samsung, Micron, and Hynix, over the five quarters from Q3-23 to Q3-24. SanDisk lagged over this time, seeing an increase of about 50%. Since Q3-24, NAND prices have been falling. Following the top in Q3 of 2024, ASPs are down about 20%. This has happened without a surge in bits from these four players. The reason? YMTC entered the NAND industry. Their market share by the end of 2024 was in the high single digits. That means a mass of bits about half the size of Micron’s total output has been added to the NAND market in the last two to three years. These are lower quality bits, but they are still being added to a large, connected pool of global supply. Further, the PRC forces local mobile and PC makers to purchase memory from YMTC, taking demand away from the traditional players.
The headline news for the memory industry from the first quarter of 2025 is the prediction of inventory depletions at PC and mobile customers seems to have happened. I saw both non-AI DRAM and NAND pricing roll over and start declining in the fourth quarter of 2024 and declared the downturn underway. In my experience, once prices start to drop in memory, it doesn’t reverse for at least five quarters. Well, this has been an historically short upturn for non-AI memory. Prices only rose for four to five quarters because Chinese supply came online. Maybe this caused a pullback in CapEx from the established players, such that the capacity that would normally end an upturn hasn’t been added. I have not reversed my view that the downturn hasn’t started, but I am much lower in my conviction. All the players are forecasting bit shipments to increase in both NAND and DRAM in the second calendar quarter. Most of them are saying pricing has started to rise and they believe that will continue for the rest of 2025. Nobody knows much more than one quarter out.
The swing factor here will be Chinese supply from YMTC and CXMT. If they continue their expansion seen in the last two years, the recovery in non-AI DRAM and NAND prices will be short-lived. If they pause – and there are indications from the WFE market that memory CapEx in China will come down significantly in 2025 – then market health may indeed increase. Export restrictions are having significant effect on equipment shipments. But the Chinese are new to this, which means they are inefficient. The squeeze of not having more equipment will turn their attention to increasing yield and wafer output. Even without a lot of new tools, fabs can significantly raise their wafer outs through efficiency gains. To support pricing, NAND makers are taking down wafer starts. Micron quoted they are down ~15%. All three companies are using the idle time to transition nodes faster, which increases bit output. I continue to think the NAND market is deeply structurally flawed. Reducing wafer starts just papers over this fact.
Which brings me to HBM. Samsung, Hynix, and Micron are all ramping their HBM output as fast as they can. My view – and the basis of my large short position in Micron – is that ramping of HBM will lead to oversupply in this segment by the end of calendar 2025. The DRAM industry has never not oversupplied the market. It is just a matter of when. AI is such a strong source of demand, making this harder to predict. The last seismic shift in the DRAM market was probably the emergence of mobile phones in the mid-2000s. But since that was a consumer phenomenon, the ramp up was slower.
– S. Hughes (short MU)