7.24.24
Investor Presentation
Q2 was a record revenue quarter for Hynix, boosted by rising ASPs for both DRAM and NAND, and a substantial increase in DRAM bit shipments. In the quarter, DRAM bit shipments rose low 20% sequentially (previous nine quarters, newest to oldest, were 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 ten quarters in total, Hynix’s DRAM bit growth has been almost flat. I estimate it to be up a cumulative 3% over this two-and-a-half-year period. Said another way, Hynix’s DRAM bit supply is the same today as it was at the beginning of calendar 2022, when the last upturn was ending. For NAND, bit growth was down low single digits % sequentially. This follows the previous nine quarters (most recent to oldest) 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 %.) Over the last full year, NAND bit shipments from Hynix are down slightly. The lack of bit supply growth in both DRAM and NAND is why pricing has been rising. Normal bit growth is at least 10% annually for DRAM and mid-teens percent for NAND. Hopefully this lack of bit growth continues for several more quarters as prices need to rise further and hold for memory companies to earn enough to pay down the debts accumulated in the last downturn, and to fund further R&D investment. In Q2, DRAM ASPs rose mid-teens % sequentially. This follows the prior five quarters (newest to oldest) of up 20%+, up high teens %, up around 10%, up high single digits and down high teens %. Because of Hynix’s leadership in HBM, their DRAM pricing started to increase in Q2-23, thus this quarter is the fifth consecutive quarter of rising DRAM ASPs for Hynix. Off of the bottom in Q1-23, DRAM prices for the company have almost doubled (up an estimated 97% off the bottom in.) For NAND, ASPs rose mid-to-high teens sequentially. This follows (newest to oldest) a rise of 30%+, 40%+, slightly down, and down 10%. Cumulatively in the last four quarters NAND prices for the company are up 110%.
Gross margin in the second quarter was up another 700 bps, from 39% to 46%. In Q2 of 2023, gross margin was negative 16%. Since then, it has expanded by 6200 basis points. That is the fastest rise I can recall seeing over a single year for a memory company. Gross margin for the last six quarters, oldest to newest, are (32%), (16%), 1%, 20%, 39%, and 46%. Note I have not made any adjustments here for any inventory write-downs the company has done. Micron’s gross margin bottomed in their fiscal Q4 of 2023, about a quarter later than Hynix’s. That is the period from June to August of 2023. Hynix reached their gross margin nadir in the January to March period of 2023, bottoming at negative 32%. From bottom to today, the company’s gross margin is up 7800 basis points. Micron’s lowest gross margin (if unadjusted for inventory write-downs) was negative 32.7%, occurring in the December 2022 to February 2023 period. Adjusted for inventory write-downs, Micron’s GM bottomed in the June to August quarter of 2023. Based on the gross margin curves of each company, and assuming a similar COGS structure between the two, Micron is 1.5 to 2 quarters behind Hynix on the gross margin improvement slope, having clocked in at around 27% a month ago, and guided for 33.5% GM in the quarter ending in early September of this year.
In the third quarter, Hynix plans to grow DRAM bit shipments by a low single digits % sequentially. Last quarter, they forecasted to increase DRAM bit shipments by mid-teens % and they actually did low 20%. For NAND, they are planning a mid-single-digit % decreases sequentially. Last quarter, they forecasted flat and were actually down low single digits %. They commented for NAND that enterprise SSDs are the only area of demand strength. Other products have soft end market demand and “relatively high customer inventory.” The company is forecasting continued revenue growth in NAND from optimized product mix and higher pricing. They didn’t say higher ASPs in DRAM, though it is implied in their comment about higher HBM volume shipments.
Their capital expenditures in 2024 will be higher than earlier plans, “but will be executed within OCF.” That is not comforting, though they did say they will “maintain prudent investment decision based on end-demand and profitability” and they will “pursue investment efficiency and financial stability.”
Their market outlook for the second half of 2024 is that strong AI hardware demand will continue, supplemented by a gradual recovery in conventional DRAM, led by the introduction of AI-enabled PCs and mobile devices. I hope this is true but I am skeptical. Recovery in PC demand has been weaker than expected this year. They are hanging their hopes for PC demand recover on the introduction of memory-intensive AI enabled machines. Mobile has been somewhat better, with a moderate recovery in demand seen in the first half of 2024. As with PCs, the company is hoping that AI-enabled mobile phone introduction later in 2024 and in 2025 will drive a recovery in memory demand from this segment. In servers, AI demand is doing most of the work. General purpose server demand is gradually improving and they are hoping for the replacement cycle to help. Hopes for continued strength in the DRAM market in the second half of calendar 2024 are mostly pinned on AI strength continuing, along with the emergence of AI-enabled PCs and mobile devices. The NAND recovery is coming from the traditional lack of bit growth allowing demand to catch up with supply.
Analyst Call
Here are highlights from the earnings conference call:
Prepared Remarks:
· Demand for AI memory remained strong and demand for non-AI memory increased in the quarter as customers attempted to build inventory ahead of potential price increases. This pull between customers wanting more inventory and memory manufacturers holding supply back to support pricing was a theme in Micron’s call also.
· DRAM bit shipment growth of more than 20% in the quarter were driven by HBM and by increasing demand from the conventional server segment.
· Prices have risen across all DRAM products for three consecutive quarters.
· NAND demand in most products is weak, yet prices continue to rise for lack of growth in bit supply.
· Across suppliers, DRAM bits being shifted to HBM. This is constraining non-AI bit supply. Because of this, even as growth in demand for non-AI memory has been muted, ASPs are “continually rising.”
· Demand recovery in the PC market is relatively weaker than had been expected. Customers are preparing for the launch of high-spec AI-enabled computers. The market is diverging into expensive machines to support AI and low-spec machines for cost-sensitive end customers.
· Mobile demand in the back half of the year is expected to strengthen with the release of new flagship and AI models.
· Demand for AI servers continue to be strong. Non-AI server demand is expected to increase as the replacement cycle for hardware installed 6-7 years ago kicks in.
· Bit crossover to HBM3e is expected in the third quarter. More than half of HBM bit shipments in calendar 2024 will be HBM3e products. 12-high HBM3e samples have been shipped.
Analyst Q&A:
· Need to lower server power consumption is motivating customers to replace servers with new ones using lower power components, including LP memory.
· Non-HBM server demand is expected to see mid-20% growth this year and in 2025. The answer was not clear if this is DRAM bit demand or if it is unit demand. My guess is the former, because of the magnitude of the growth.
· In response to a question about HBM investment leading to oversupply, the company believes the HBM market is structurally different than conventional DRAM, thus one cannot assume investment will automatically lead to oversupply. Not this time is different, but this product is different. Their answer suggested that the HBM contracts with customers are commitments from customers to purchase the product at the agreed time and for an agreed price. Still, this is just a strong suggestion vs. a definitive statement.
· The company expects the DRAM market to evolve from a commodity to a non-commodity business because of AI demand. This is fanciful.
· Capital investment so far this year has been to meet strong HBM demand and for cleanroom space to allow future expansion.
· In 2025, investment is expected to be above the historical average, to support increasing demand.
· If demand for conventional DRAM increases, it is possible that prices for these bits may rise above those for HBM bits. The company will balance supply between HBM and non-HBM to maximize profitability over the long-term, not the short term, considering customer relationships as one of the factors.
· The AI market is growing faster than expected.
· The company is committed to maintaining their NAND market share but is also focused on profitable investment. Because of low demand and customer inventory conditions, Q3 NAND bit shipments will decline. They expect to return to NAND bit shipment growth in the fourth quarter.
· In Q2, sales of both DRAM and NAND exceeded production volumes. They are depleting inventory to support current demand for both memory types. This was also the case in Q1.
· In 2023, eSSD demand was weak because of higher AI server demand reducing conventional SSD demand as IT budgets are limited. eSSD demand is significantly exceeding their expectations at the beginning of the year.
Summary
Demand for AI memory products has been the story of this upturn so far. The recovery in DRAM demand and pricing has been (and is) bifurcated into the AI cycle (HBM and, to a lesser degree, DDR5) and conventional products (PC, mobile, non-AI data center, auto/industrial). Hynix enjoyed a wide lead in HBM so the beginning of the AI memory cycle can be judged from their results. The bottom of the AI DRAM cycle was in calendar Q1 of 2023, five quarters ago. Since then, Hynix’s blended DRAM ASPs have doubled. Micron’s DRAM recovery is 1.5 to 2 quarters behind this, having bottomed in the June-August quarter of 2023. The story of the rest of the DRAM upturn hinges on AI in two ways. First, will HBM demand proceed in the next 4+ quarters at the levels Hynix and Micron are expecting. Both companies say they are sold out through calendar 2025 but I am skeptical that if demand from AI cloud customers pulls back, those HBM sales agreements will be honored. I hope I am wrong in this view, but the memory industry has little to no history of suppliers having enough market power to force customers into take-or-pay contracts. And second, recovery in PC and mobile demand is being pinned on AI-enabled devices emerging in these segments. I am also skeptical of this, because new products such as those typically take longer to emerge and be adopted than is initially believed. Working in favor of the memory makers is the supply situation. So far, the two of the big three (Samsung reports next week) seem to be restraining themselves from adding meaningful new supply at least in calendar 2024. With HBM demand soaking up bits and the conventional DRAM and NAND markets already undersupplied, the setup seems good for continued rising supply. Both Micron and Hynix seem to be holding back inventory to support higher prices. On NAND, the first quarter of 2024 seems to have been the start of the recovery. None of the suppliers so far have shown any interest in adding capacity, leading to a sharp rise in ASPs. I think this will continue to be a tailwind for the financial results of the big three memory makers. I will paint three scenarios for the next six to twelve months in the DRAM market. One, HBM demand holds, demand from non-AI memory segments remains tepid, and supply growth is flat. In this scenario, ASPs will continue to rise at least through the third and fourth calendar quarters of 2024, probably by double digits sequentially in both quarters. Two, HBM demand holds and non-AI memory demand growth strengthens, along with constrained supply growth. In this scenario, sequential ASP gains will be in the high-teens to 20% range in both Q3 and Q4. And in the third scenario, HBM demand pulls back, non-AI memory demand remains tepid, and supply growth is constrained. This will lead to memory makers seeing inventory growth and ASP sequential percent increases falling into the single digits or worse. I think the chances of either scenario one or scenario two are in the 70% to 80% range, thus the chances of AI demand collapse are 20-30%. Typing that out, 30% seems high. 20% feels about right. I think this pullback in AI optimism, reflected in the recent retreat of the NASDAQ, will prove to be temporary and not the downslope on the back of the peak of inflated expectations.
-S. Hughes (cyclical long MU)