1.23.24
Investor Presentation
The fourth quarter was again a revenue record for Hynix. Sales were up 12% sequentially and 75% year-over-year. Revenue for the full year was a full third higher than the previous peak. 2024 sales were KRW 66.2T. The previous high was KRW 44.6T in 2022. 2021 and 2018 were not far behind this level. The difference in 2024 is operating profit was KRW 23.5T, the highest ever. This was not far above the KRW 20.8T achieved in the 2018 cycle. That 2018 period was a super cycle of profitability. Micron achieved over 60% gross margin at that time. Still, this cycle is impressive for Hynix. Operating margin in the fourth quarter came in at 41%.
In the fourth quarter, DRAM shipments were up mid-single digits % from the prior quarter. This matches what management guided for in their Q3 release. The previous eleven quarters, newest to oldest, were down slightly, up low 20%, down mid-teens %, up low single digits %, up around 20%, up mid-30%, down around 20%, flat, down mid-single digits %, up 10% and down high-single digit %.) Over the last three years, Hynix’s DRAM bit output has grown by approximately a third. That is a CAGR of just over 10%, which is below the long-term average for the DRAM market. The reason for that is, of course, the brutal downturn of 2022-23.
In Q4, DRAM ASPs were up a healthy 10% sequentially. That is following a mid-teens % rise last quarter. Including all of 2024, the ASP changes the last eight quarters (newest to oldest) have been up 10%, up mid-teens %, up mid-teens %, up 20%+, up high teens %, up around 10%, up high single digits and down high teens %. Since the DRAM upturn began for Hynix (their last quarter of down DRAM ASPs was the first calendar quarter of 2023,) their blended DRAM ASP is up 150%. DRAM prices for Hynix have been rising for seven consecutive quarters. If history is an accurate guide to this cycle, DRAM pricing will peak in the first or second calendar quarters of 2025. What happens from here will depend on three factors. First, will AI demand strength continue? Second, will non-AI DRAM demand recover? And third, how much more supply will CXMT in China bring into the market? The second and third of those factors cannot be completely decoupled, because it is not possible to see all the demand in China that is now being serviced by CXMT that was previously fed by the big three.
NAND ASPs were negative for the first time in four quarters, declining mid-single-digits % sequentially. Starting with the second calendar quarter of 2023, Hynix’s blended NAND prices have increased by 110%. It has been a wild ride to get there. Prices were down 10% in the middle of 2023, then they more than doubled in three quarters. NAND pricing strength continued through the third calendar quarter of 2024 before peaking and turning down in Q4. Historically, this would mark the end of the upturn in NAND. While there is more elasticity in NAND prices than there is in DRAM, the market turns like a cargo ship rather than a speedboat. I expect pricing weakness in NAND to continue for most of 2025, because I don’t think the Chinese supplier YMTC is going to take their foot off the gas. Even as all the other NAND companies continue to minimally invest in NAND, supply from YMTC will continue to weaken the market. They need to keep ramping in order to reach the required scale to have a viable business.
Bit shipments for NAND were also down mid-single-digits percent in the quarter. Three months ago, when Hynix guided Q4, they expected NAND bit shipments to be up low-teens %. NAND market demand has weakened much more than management expected in just the last ninety days. The NAND bit shipments of down mid-single-digits follow the prior eleven quarters (most recent to oldest) of down mid-teens %, down low single digits %, flat, down low-single digits percent, up mid-single digits percent, up around 50%, down mid-teens %, down high single digits %, down low teens %, up high single digits % and up high teens %. In total, Hynix’s NAND bits shipments have increases approximately 11% over the last three years. In TOTAL. Not annually. Hynix is only shipping 11% more NAND bits than they did three years go. This is in a market that was widely expected to grow in the mid-teens % annually. Yet, NAND prices have rolled over and are weakening.
Gross margin in the fourth quarter was again 52%, same as in the third quarter. Gross margin for the last eight quarters, oldest to newest, are (32%), (16%), 1%, 20%, 39%, 46%, 52%, and 52%. Hynix’s gross margin has now been positive for seven quarters. As has been the case historically, their gross margin is flattening about two years into the up-cycle. Both Micron and Hynix are ramping HBM as fast as they can (as well as Samsung, who is a distant third). Micron’s gross margin curve hasn’t flattened in the last two quarters the way Hynix’s has. My guess as to why is Micron is on a steeper part of their HBM ramp curve, so their mix of HBM is increasing faster than it is for Hynix. Because Hynix is well ahead of Micron in HBM percentage, their gross margin is in the 50s while Micron’s is still yet to crest 40%. I am surprised to see the flattening of Hynix’s gross margin because their HBM percentage is still rising. I interpret this to be because their non AI memory products, especially NAND and legacy DRAM, are seeing significant price declines.
In the whole of calendar 2024, Hynix’s HBM sales increased by more than 4.5x. Their plan going forward is to increase the profitability of non-AI DRAM by selling more DDR5 and LPDDR5. The problem with that strategy is Micron and Samsung are doing the same thing, so all three large DRAM companies are rapidly shifting bit supply away from DDR4 and to DDR5. This may just move the oversupply situation from one memory type to another, unless demand outside of AI data centers improves.
In 2025, The company is forecasting DRAM bit demand to rise mid-to-high teens % year-over-year. They expect NAND bit demand to grow low-teens % year-over-year. This is the time when I remind readers that memory executives don’t know much beyond one quarter out. Their DRAM demand forecast is above the average of recent history, which means they are forecasting continued strong AI sales. I am surprised by their high NAND forecast. I think they are assuming demand elasticity kicks in to reverse the current slide. By segment, they see low-to-mid single digits % growth in both PC and mobile units. These are supported by AI device penetration rates of 30% to 40%, which I find laughably optimistic. I just don’t see the use case. They are forecasting server units to grow high-single-digits %, driven by strong demand for AI servers.
For the first quarter of calendar 2025, the company is forecasting DRAM bit growth to be down low-teens % sequentially. This is a large drop. They believe the NAND market will be even weaker, down high teens % sequentially. Capital expenditures for calendar 2025 will increase “meaningfully” due to new fab construction. They are shading this to minimize the appearance that they are adding bits. They will focus investment on leading edge and AI products. They will adjust investment based on market conditions, which is always true. The company is concerned about weakness in NAND and non-AI DRAM.
The company expects HBM revenue in 2025 to grow more than 100% compared to 2024. That is a massive absolute level of sales growth given how much HBM they are making today. In NAND, the company will “maintain profit-oriented operational decisions and flexible sales strategy.” That is leaving the door open to both reduce NAND wafer starts and to hold back NAND inventory in the face of weak market conditions.
The company included a stacked bar chart of DRAM and NAND by segment. Their fraction of PC and Mobile in DRAM is surprisingly small, a total of between 20% and 25%. Their fraction of NAND bits going into non-SSD applications is around 22%. The graph makes their SSD bit share look better than it really is, at first glance, because they don’t break out NAND in SSDs Hynix sells from bits they sell to others that then make them into SSDs. The margin for Hynix is quite different between those two sales channels.
Analyst Call
Here are highlights from the earnings conference call:
Prepared Remarks:
· Mobile and PC memory demand continues to be weaker than expected. The company called the recovery in these segments “delayed.”
· HBM was over 40% of DRAM sales in the fourth quarter. DRAM was 74% of total revenue in the period, thus, revenue from HBM in Q4 was at least 30% of total company sales.
· NAND demand was weak because of inventory overhang at customers.
· The company believes the memory market is transitioning from a commodity market to a customized, differentiated market. I think this is laughable. HBM will become a commodity supplied by the big three DRAM makers, just like every other volume DRAM product ever has.
· NAND was profitable for 2024 as a whole.
· Supply-demand conditions are expected to “gradually improve” in the second half of 2025. This is just hope.
· Inventory levels continue to be high in the supply chain. They are calling the current situation in the non-AI memory to be a “correction.” Yikes.
Analyst Q&A:
· The first question was again about Chinese supply coming into the market. The company didn’t address PRC supply directly, but said that they expect weak conditions to continue in the first half of 2025.
· The AI memory market will continue to be strong. The non-AI DRAM market will go through a “smooth correction.”
· They almost said “this time is different” when they stated that this cycle is unlike past cycles in that new investment is going into AI, not into commodity DRAM. They said the same thing about the NAND market, where they expect a “milder correction.” It’s so over.
· They plan to ramp up their HBM4 product in the latter half of this year. This supply will begin with 12-high products, transitioning to 16-high products as customer demand requires.
· CapEx in 2025 will increase year-over-year and will be focused on HBM and infrastructure. Infrastructure will increase significantly over last year. This is similar to what Micron said and has the same “out.” Investing in HBM IS investing in new capacity. There may be a trade ratio when shifting from older technologies to DDR5 and HBM, but CapEx is still adding capacity to the market, a market that is currently tipping into oversupply.
· DDR4 and LPDDR4 demand is expected to continue to shrink. Revenue share (presumably for Hynix vs. for the market as a whole) will decline from 20% in 2024 to single digits percent in 2025. That is a 50% decline, from both prices and volumes dropping.
· Management disagrees that there are concerns about slowing demand for HBM as AI shifts towards inference. They brought up the compute needs for AGI as one reason, which I think is a reach. However, I do agree that AI investment has some room yet to run.
· They will adjust NAND production as needed, in response to market conditions. In other words, they will reduce wafer starts if demand is weak enough.
· They see a performance gap between their DRAM products and those made by Chinese suppliers, specifically in the quality and performance of DDR5. They plan to use their rate of technology advancement to stay ahead of these emerging competitors. This is their best strategy, in my view. More advanced process technology, customer relationships, and scale are their three most meaningful advantages over the emerging Chinese companies.
· General DRAM wafer starts at Hynix are expected to “increase slightly” in 2025. They are targeting their general DRAM production to grow in line with market demand. I interpret this to be saying DRAM bit growth outside DDR5 and HBM will be well below 10% year-over-year.
Summary
I think this earnings report marks the beginning of the downturn for the memory market, outside of AI DRAM. While they talked little about Chinese competitors, opting instead to say weak demand and excess inventory are behind their declining bit shipments. I believe the true reason is indigenous Chinese supply coming online. I believe those suppliers will continue to ramp in 2025, exacerbating the situation. Hynix is saying all the things suppliers normally say when the upturn has ended and the market rolls over. They said the non-AI market will experience a “smooth correction” in 2025. In a first for memory, they also said that the emergence of HBM as a significant product category marks the transition for DRAM away from a commodity market. I don’t think this is the case. Rather, HBM will become another commoditized product. It will just take some time for the market to mature to that point. NAND is done in this cycle. It is in oversupply and I believe it will stay there for the rest of 2025. Of course, I don’t know any more than the memory executives do beyond one quarter. I’m just not burdened by their necessary optimism. As for DRAM, a race is on between demand for HBM and DDR5 and weakness in DDR4 and LPDDR4. If AI demand stays strong, manufacturers may be able to switch production over fast enough to keep prices for products for older DRAM nodes from completely collapsing. That is the race that Micron investors need to pay close attention to. I think the weakness in NAND is mostly already priced into the stock.
-S. Hughes (no MU position)