Memory plays a critical role in the world of AI, as faster and more powerful AI chips and servers will require increasing amounts of memory. The recent surge in generative AI and AI GPUs, spurred by the success of OpenAI’s ChatGPT and development of hundreds of other large language models, are forecast to bring about a new DRAM market, underpinned by high-bandwidth memory (HBM) and DDR5. Micron is uniquely positioned to benefit from this secular AI growth tailwind, leading the industry with the fastest, highest-capacity HBM, alongside other industry-leading products for AI applications.
This analysis will touch on how the memory market is recovering off one of the worst cyclical downturns it has faced recently, how NAND and DRAM are evolving, the massive shift ahead for HBM3, and how this translates to a multifaceted growth opportunity for Micron to tap into to push towards record revenues.
Please note, our plan is to patiently wait for the right entry. This is detailed below at the Conclusion.
Memory Market Recovery Begins in 2024, Accelerates in 2025
The memory market is coming off its worst cyclical downturn in 15 years, with late 2022 seeing memory revenues fall off a cliff, with challenges persisting through much of 2023. Heading into Q4, the market is showing multiple signs of bottoming – South Korean NAND flash exports returned to growth in September, NAND and DRAM pricing is expected to return to growth this quarter, and supply bit shipments are forecast to return to double-digit growth in 2024.
The downcycle starting in Q2/Q3 of 2022 exceeded the previous cycle in late 2018 in terms of its scale and duration – memory revenues declined approximately (-26%) from ~$44 billion in Q2 to nearly $33 billion in Q3. NAND revenues declined (-24.3%) to $13.7 billion, while DRAM fared slightly worse, with revenues declining (-28.9%) to $18.2 billion.
Q4 of last year saw another ~ (-27%) sequential decline to just over $24 billion in revenues. NAND revenues slipped around (-19.7%) sequentially to ~$11 billion, while DRAM revenues fell (-34.2%) sequentially to $12.3 billion. Essentially, DRAM revenues saw nearly a (-50%) decline in just two quarters, while the broader memory market declined just over (-45%).
Driving this rapid slowdown was a rapid deterioration in NAND and DRAM pricing – this was exacerbated by excess inventory at major manufacturers Samsung and SK Hynix leading to a fire sale at extremely low prices. NAND and DRAM both saw pricing decline more than (-20%) QoQ in Q3, and nearly (-30%) QoQ in Q4 2022. The previous year, from Q3 2021 through Q1 2023, NAND prices fell (-55%), and DRAM declined (-57%).
Source: Yole Intelligence
Pricing is expected to bottom out in Q3, with Q4 projected to see the first QoQ increase in both NAND and DRAM pricing since Q3 2021. That trend is set to continue through 2024, with Yole Intelligence forecasting NAND and DRAM prices to continue rising.
Source: Bloomberg
South Korean export data further supports the recovery story for NAND and DRAM.
NAND flash exports for September rose for the first time in a year, while DRAM exports registered the smallest decline. NAND flash exports increased +5.6% for the month, compared to an (-8.9%) decline in August, while DRAM exports fell (-24.6%). In October, chip exports from the country declined just (-3.1%) YoY to $8.9 billion in October, the smallest decline since August 2022, and another data point in support of memory’s recovery from the deep trough of late 2022 and early 2023.
A Broader Look at Memory’s Rebound
Pricing and export data both are signaling a return to growth for the memory market starting in Q4 and persisting through 2024 and 2025. Production cuts beginning in early 2023 are expected to lead to an undersupply of chips over the next four to eight quarters, and combined with steadily increasing NAND and DRAM prices, the memory market is projected to jump to record levels by 2025. cuts beginning in early 2023 are expected to lead to an undersupply of chips over the next four to eight quarters, and combined with steadily increasing NAND and DRAM prices, the memory market is projected to jump to record levels by 2025.
Overall, the memory market is projected to register a (-41%) YoY decline to approximately $84 billion in revenues in 2023, down from ~$144 billion in 2022. DRAM revenues are projected to fall (-47%) YoY from $79.7 billion in 2022 to ~$42 billion in 2023; NAND revenues are expected to fall (-37%) YoY from $58.7 billion to $37 billion this year.
2024 is forecast to see a sharp rebound in the market, with revenues rising approximately +55% YoY to more than $130 billion, according to Yole Intelligence. The increase in prices are driving the revenue jump with DRAM growing by “as much as 87 percent” in 2024 while NAND flash memory is projected to “bounce back to grow by about 60 percent,” according to Gartner.
2025 is projected to see the market reach record revenues of above $200 billion, or over +55% YoY for a second year straight, boosted by increased prices, undersupply – especially in DDR5 and other DRAM sub-segments — and AI mega-trends.
According to Lam Research, AI servers use 8X DRAM and 3X NAND compared to an enterprise class server. For DRAM, that AI server and data center demand is expected to push the markets to new highs, from that ~$42B size in 2023 to ~$96B in 2028, according to Yole Intelligence.
Overview: HBM and DDR5’s Role in AI
High-bandwidth memory (HBM) – more specifically the next-gen HBM3/HBM3e – and DDR5 play a mission-critical role in AI, and such a role will lead the market to stunning growth: market leader SK Hynix projects the HBM market will grow at an 82% CAGR through 2027.
High bandwidth memory (HBM) offers higher bandwidth, capacity, performance, and lower power by vertically stacking up to twelve DRAM memory chips to shorten how far data has to travel, while also allowing for smaller form factors. Stacked memory chips are connected through something called “through silicon vias” or TSVs. HBM is increasingly being used to power machine learning, high performance data centers, and more recently, generative AI models. For a different perspective into HBM from an equipment leader, read our analysis on Lam Research here.here.
DDR5 DRAM, or double data rate 5, aimed to double bandwidth and data transfer speeds at a lower latency and power consumption than its predecessor, DDR4. DDR5 memory chips can be mounted on circuit boards to create memory modules, for use in servers or PCs. DDR5’s increased bandwidth allows for faster processing for memory-intensive applications, such as generative AI and training LLMs.
Demand for high-capacity memory is being driven by the sudden rise in generative AI and LLMs, which both require significant amounts of computing power and substantial amounts of DRAM to meet elevated performance requirements. SK Hynix’s head of DRAM marketing Park Myung-soo explained that “an AI server requires 500-gigabyte (GB) or larger High Bandwidth Memory (HBM) chips and at least 2-terabyte (TB) DDR5 chips.” As such, HBM3 and DDR5 are projected to see a rapid shift to become the dominant architectures over the next few years.
In 2022, HBM2e was the dominant HBM architecture, accounting for ~70% share of HBM shipments, with the emerging HBM3 taking just 8%. However, HBM3 will see a surge in demand in 2023 and 2024, rising to 39% share this year and approaching 60% share in 2024 to become the dominant architecture. This will be driven by the massive demand for Nvidia’s H100 GPU and AMD’s upcoming MI300, which are underpinned by HBM3.
As a result, HBM3 revenues are forecast to rise as much as +127% YoY to $8.9 billion, according to TrendForce, while SK Hynix estimated that this AI chip boom will lead to the HBM3 market expanding at an +82% CAGR through 2027.
Similar to HBM3, DDR5 is expected to quickly become the mainstream DDR architecture, driven by demand for faster compute for AI. DDR5 was expected to take more than 25% share of bit shipments in 2022, before rising to take more than 55% share in 2023. By 2026, DDR5 is projected to hold almost 95% share of bit shipments as the memory market completes its transition over to DDR5 from DDR4.
Source: Tom’s Hardware
Micron is well positioned to benefit from this massive shift to HBM3 and DDR5, as it is first-to-market with its 128GB capacity DDR5 RDIMMs (registered memory modules) and its eight-high 24GB HBM3e cube.
Background on Micron’s Positioning in HBM3 and DDR5
Throughout 2024, Micron is seeing “accelerating AI-driven opportunities for memory and storage across multiple market segments from the data center to the edge,” and it is rolling out industry-leading HBM and DDR5 products. Micron’s 8-high 24GB HBM3e cube began sampling in late July 2023, and it plans to sample its 12-high 36GB HBM3e cube in the first quarter of 2024 – among the first in the industry to reach the market.
Micron is investing heavily to capitalize on this HBM3/3e shift, with management noting last quarter that “assembly and test capex is projected to double year over year in fiscal 2024, predominantly driven by investments to support HBM3e production.”
Micron currently has deployed the industry’s fastest and highest capacity HBM3e on the market, supporting the most advanced AI training and inference. Micron’s HBM3e can deliver faster training times and more responsive queries for LLMs — it “increases performance per watt resulting in lower time to train LLMs such as GPT-4 and beyond.”
Micron’s HBM3e provides higher memory bandwidth that exceeds 1.2TB/s and 50% more memory capacity per 8-high 24GB cube, improving the accuracy and precision while training LLMs. Micron explains that this allows for up to 50% or more queries per day while reducing training times by 30%, thus lowering total cost of ownership (TCO).
TCO is an important factor for hyperscalers when evaluating equipment, especially GPUs, as it factors in not only the acquisition cost but also the costs associated with owning and operating the equipment over the hardware life cycle. Micron’s 24GB cube touts 2.5 times performance per watt improvement over previous generations; this increased power efficiency generates “tangible cost savings” for AI data centers. Micron explains that for “an installation of 10 million GPUs, every five watts of power savings per HBM cube is estimated to save operational expenses of up to $550 million over five years.”
In terms of DDR5, Micron says it currently sits as the market leader, with the most market share in the early innings of this DDR5 shift, underpinned by the memory industry’s most developed DDR5 ecosystem. Micron has launched its high-capacity 96GB DIMMs, and at the beginning of November, Micron launched the first-to-market 128GB DIMM based on its 1β technology, which it says “delivers the fastest speed and lowest latency” of any DDR5 DIMM available.
Micron explains that the 128GB DIMM offers more than 24% improved energy efficiency, as well as 16% improved latency, which is crucial for “memory-bound workloads such as generative AI, in-memory
databases, and real-time data analytics, where high-capacity is needed, and prompt response times are critical for real-time inference.” Micron adds that the 128GB DIMM “delivers up to 28% faster performance for AI training” on models such as Meta’s Llama 2-70B.
Customers are seeming optimistic about the benefits that the 128GB DIMM offers – AMD SVP Dan McNamara said that “as AMD advances compute with our next-gen EPYC processors, Micron’s 128GB RDIMMs will likely become one of the main memory options to deliver high-capacity and bandwidth per core capabilities to address the demands of memory-intensive applications.” Intel VP Dr. Dimitrios Ziakas echoed that sentiment, saying, “Intel is evaluating this 32Gb memory offering for key DDR5 server platforms based on the resulting total cost of ownership benefits to cloud, AI and enterprise customers.”
Competition Remains Stiff for Micron
While Micron claims it is first-to-market with HBM3e and holds the most market share in DDR5, competition remains stiff, as Micron is competing against two heavyweights who control more than 85% of the market – Samsung and SK Hynix. The two South Korean firms are rapidly advancing HBM3 development, with SK Hynix already firmly established in the market as it was the preferred HBM3 supplier for Nvidia’s highly popular H100 GPU.
SK Hynix unveiled its 1TB bandwidth HBM3e memory in late Q2 this year, with mass production set to start in early 2024. SK Hynix is also reportedly eyeing development of its next-gen HBM4 cube with a plan to introduce that product to market in 2026.
Samsung is currently mass producing its 12-high 24GB HBM3 cube, ‘Icebolt’, and is sampling its HBM3e cube ‘Shinebolt’ to prospective customers. While coming to market later than Micron and SK Hynix, Shinebolt is rumored to compete with Micron on performance, with both offering 1.2 TB/s bandwidth. Samsung is also expected to unveil its fifth-gen HBM3e cube, named ‘Snowbolt’, by the end of the year, followed by a sixth-gen HBM cube next year.
Nvidia, AMD Battling on Memory
HBM3 and HBM3e are becoming the next battleground for memory chip manufacturers as well as AI chip developers, especially Nvidia and AMD, who are pushing the boundaries with the amount of memory bandwidth in each GPU.
AMD’s competing GPUs, the MI300 series, substantially boosted memory and bandwidth relative to the H100, utilizing Samsung’s HBM3. The MI300A is shipping with 128GB HBM3 memory while the MI300x ships with 192GB memory and 5.2 TB/s of bandwidth – that’s 1.6x more bandwidth and 2.4x more HBM3 density than Nvidia’s H100.
Nvidia is rapidly moving forward with its GPU roadmap, as it aims to launch its next-gen H200 and B100 GPUs next year followed by the X100 GPU in 2025 – each GPU will accelerate AI inference times along an exponential curve, thus creating a need for more memory and more bandwidth.
Source: Nvidia
Nvidia’s A100 shipped in two different versions with either 40GB or 80GB HBM2e memory, with the 40GB offering 1.55TB/s of bandwidth and the 80GB offering 2TB/s bandwidth, the industry’s fastest at the time in 2021.
Nvidia then upgraded from HBM2e to HBM3 DRAM, tapping SK Hynix as the supplier for its H100 GPU for 1.6X the bandwidth. Nvidia’s upcoming H200 GPU, set to launch in early 2024 as the industry’s first HBM3e-powered GPU, is expected to ship with another 1.5x bandwidth boost relative to the H100 with nearly 1.8x the memory.
It is rumored by some sources that the H200 is shipping with Micron’s HBM3e, instead of SK Hynix. rumored by some sources that the H200 is shipping with Micron’s HBM3e, instead of SK Hynix. Micron reportedly sampled its 24GB HBM3e memory with Nvidia at the end of July, with SK Hynix following in mid-August and Samsung in early October. According to sources in South Korea, Nvidia remains engaged with SK Hynix for the H200.
This raises an important point about competition in this AI chip and memory race: if Nvidia is switching this quickly from one supplier to the next based on time to market, this raises the risk that Samsung or SK Hynix could be first to market with a superior HBM4 product and take share away from Micron, especially if they undercut Micron on price.
AI Will Increase Secular Growth Opportunity for Micron
The recent surge in AI and data center growth fueled by Nvidia is expected to translate into an interesting shift in revenue mix for Micron: it foresees exposure to the more cyclical and seasonal PC and mobile end market declining from 55% share of revenue in FY21 to 38% share in FY25.
Data center and graphics are forecast to rise from 30% share of revenue in FY21 to 42% share by FY25, with AI and machine learning driving such growth; the projected surge in the HBM3/HBM3e market supports this shift. In data center, Micron is expecting NAND GB shipped per server to increase 3x and DRAM 2x by 2025, as AI servers require significantly more memory than traditional servers.
Automotive and industrial are projected to rise from 15% share to 20% share, as both end markets exhibit much faster growth rates than PC and mobile due to the rise of electric vehicles, industrial robotics, and other emerging trends which require a higher semiconductor content per unit.
This revenue mix shift is underpinned by long-term agreements, at ~75% of revenue in CY22. This offers multiple benefits: more visibility into forward revenues, less exposure to cyclical pricing trends in NAND/DRAM with pricing locked in for the contract duration, reduced impacts from supply and demand imbalances, and ultimately more stable margins.
Financials Sharply Improving
The swift decline in the broader memory market over the past eight quarters has had a significant effect on Micron’s financials. Revenues plummeted, falling (-49.5%) YoY in FY23 from $30.76 billion to $15.54 billion. Operating margin also shifted deep into the red, with Micron posting a (-37.0%) operating margin, compared to a 31.5% operating margin in FY22. Micron reported a net loss of (-$5.83 billion), or ($5.53) per share, compared to net income of $8.69 billion, or $7.75 per share, in FY22. This was the sharpest decline for revenues, operating income, and net income in Micron’s history.
Fiscal Q4 (ending Aug 31) showed initial signs of a recovery:
- NAND revenues increased 19% sequentially to $1.2 billion, bit shipments rose >40%
- DRAM revenues increased 3% sequentially to $2.8 billion
- Total revenue increased 7% sequentially to $4.01 billion
- GAAP net loss improved 25% sequentially to ($1.43 billion)
- Operating cash flow improved 938% sequentially to $249 million, but is much lower compared to $3.78 billion in the year ago quarter
Fiscal Q1’s guide was boosted at the end of November:
- Revenue is projected to be $4.7 billion, compared to a prior view for $4.4 billion +/- $200 million. This would represent YoY growth of 14.9%.
- Gross margin is expected to be (0.5%) to 0%, compared to the prior view of (4.0%) +/- 2.0%
The surge in the HBM3 market, positive outlooks for a NAND and DRAM pricing recovery through 2024 and into 2025, and a surge in AI and data center demand are expected to fuel a rapid recovery for Micron’s top and bottom line.
Moving forward through FY24 (Sept. 2023-24, this reacceleration in NAND and DRAM, buoyed by increasing pricing, is expected to send Micron’s revenue on an eight-quarter streak with more than +20% growth – even as high as +65% as it laps easy comps. On a dollar basis, revenues are forecast to rise from $4.01 billion in Q4 of fiscal 2023 to $8.71 billion in Q1 FY26.
Essentially, Micron is on track to potentially reach record revenues just over eight quarters after that massive slump. However, EPS is forecast to be below levels seen in FY18 and FY22 – estimates peak at $2.05 in Q4 FY25, compared to $3.53 in Q4 FY18 and $2.59 in Q3 FY22. What this means is that revenues are being propelled higher by this shift to HBM3/3e (as it exhibits much higher ASPs relative to typical DRAM memory), but margins are having a tougher time recovering as rapidly due to the deep trough that NAND and DRAM prices must rebound from.
Operating cash flow is also expected to rebound quickly after plunging alongside revenues and EPS in FY23. OCF margin is estimated to rebound from 10% to more than 40% by FY25, with Micron projected to generate upwards of $13.2 billion in operating cash flow, compared to $1.56 billion in FY23.
Micron has applied for CHIPS Act funding for its New York and Idaho fabs, saying that federal funding and tax incentives were needed to develop both facilities. Micron is investing up to $115 billion over the next 20 years to build out its US production base, in an effort to boost the US’ share of production from 2% to 15% and diversify away from East Asia – Micron’s current high-end chip production is more concentrated in Japan and Taiwan. Given that construction isn’t set to begin until late next year with production commencing as early as 2026, any margin benefits from the CHIPS Act are unlikely to be recognized over the short and medium term.
Risks
China presents a real risk to Micron as it does to much of the semiconductor industry. Micron is generating nearly one-quarter of its revenues from China, and CEO Sanjay Mehrotra recently told CNBC that “about half that revenue is at risk.” Micron was the first American chip company targeted by China with a partial ban earlier this year, and the company is working to improve relations with the nation, though there is no guarantee that such a ban will be lifted.
However, there are risks to this recovery, in that it may not unfold as smoothly as projections picture. Quarterly revenues have been variable over the past few fiscal years, with multiple sequential declines present along the growth trend, so there is a risk that current projections calling for sequential growth through Q1 FY26 do not account for some of that lumpiness.
In addition, there is also a tail-end risk in that NAND and DRAM pricing does not exhibit consecutive sequential growth through 2024. This could be exaggerated if NAND pricing slips back to sequential declines in Q2, as it is forecast to see just +3-4% sequential growth in that quarter. Pricing has shown to be volatile in the past, and there is no guarantee that pricing will rebound smoothly and steadily. Should some sequential declines appear in NAND and DRAM pricing in 2024, this would likely weigh on both margins and EPS.
Valuation
Micron’s fundamental backdrop is projected to see rapid top-line growth and a sharp bottom-line improvement on the backs of surging HBM3/3e demand; however, semis have rallied this year due to Nvidia leading the historic Nasdaq rally in the first six months of 2023. This has resulted in a rising of all boats, as many semis are trading far above historical multiples. The product of strong gains in the broader semiconductor industry has pulled Micron’s shares higher: the iShares Semiconductor ETF has gained +48.7% YTD, significantly outperforming the S&P 500’s +18.7% return.
Micron currently trades at a 5.43x PS ratio, its highest level in more than 20 years, in part due to FY23’s (49.5%) YoY decline in revenues while shares have gained +54.5% YTD. Even with +34.5% estimated revenue growth in FY24 to $20.9 billion, Micron still trades at a 4.05x 1-year forward PS ratio, far above its 5-year median PS ratio of 2.75x.
Micron’s forward EV/EBITDA multiple of ~13.0x also is elevated, at nearly double its 5-year median multiple. This accounts for the expected top and bottom line recovery for next year, but it will take more than a few quarters for this top-line growth to translate to a strong recovery in EBITDA and regression to the mean.
Conclusion
As Beth stated in Nvidia’s Q3 earnings preview, HBM3e is rapidly making its presence known, with Nvidia’s upcoming H200 GPU to be the first with HBM3e memory, rumored to come from Micron (still needs confirmation, but is an exciting possibility should it come to fruition).
Nvidia’s rapid GPU upgrade roadmap in response to AMD’s MI300X is a testament to the fast-moving nature of not just the AI GPU market, but also memory – HBM3e is coming to market in 2024, but may quickly be replaced by HBM4 in 2025 and future iterations of HBM memory beyond 2025 to 2026.
It is a highly concentrated market, dominated by Samsung and SK Hynix, though Micron remains an important player as it moves ahead with industry-leading first-to-market HBM3e and DDR5 solutions. Micron looks well positioned to capitalize on this AI mega-trend, with revenues from both solutions contributing in 2024. On a broader scale, AI and data center are set to transform Micron’s revenue mix to stronger and more secular end markets from its historically cyclical and seasonal concentration in PCs and smartphones — Micron estimates data center and AI to rise from 30% of revenue in FY21 to 42% in FY25, while smartphone and PC’s 55% share is estimated to decrease to just 38%.
However, Micron is valued at elevated multiples, when memory stocks typically would be on a deep discount given the steep downcycle, yet SOXX returns are > QQQ returns. This has lifted the tide of all boats, and memory stocks such as Micron are not trading where they’d normally trade, which makes a near-term buy less likely for the I/O Fund until we see a pullback. This is where the I/O Fund is unique, not only do we strive to be early in our research, such as to the importance of memory for the next generation of AI accelerators, but we are also careful with our timing.
Technical Analysis
The long-term pattern best fits as a large degree diagonal pattern. This is a 5 wave uptrend that is characterized with large corrective swings. If accurate, we are in the middle of the 4th wave correction, which would target the $48 – $35 region before completing.
If we zoom into the 2021 top to now, we can get a better look at the risk parameters that would either confirm or invalidate this setup. For one, the price action from the 2021 high best fits this pattern. Note how we only have a clear 3 wave drop into the October, 2022, low. This is followed by a messy and overlapping uptrend into the recent high. More times than not, when the following bounce after a 3 wave drop is a messy and overlapping move, it signals a bounce in a larger corrective pattern.
If we see a break down below the $62-$58 support region, then this pattern will be confirmed, as we establish lower targets to buy this stock. On the other hand, in order to invalidate the risk present in the current price structure, we need to see price break above the $84-$91 resistance zone. The higher we go into this region, the more likely that we will see a continuation of the larger uptrend.
I/O Fund Equity Analyst, Damien Robbins, contributed to this analysis. I/O Fund Portfolio Manager, Knox Ridley, contributed to this analysis.
Recommended Reading: