“New Investment: Micron Technology (Nasdaq: MU)
If there’s one investment mistake I’ve made multiple times in my career it’s accepting lower quality businesses available at attractive prices. I hope I’m not repeating this mistake again with Micron.
Acquiring a low-quality business at a great price usually does not lead to great investment returns, but neither does a great business at a low-quality price. You have to have both to earn superior returns.
I view Micron’s share price is quite attractive, but I also believe the business is transitioning from being mediocre to rather good. If that happens, attractive returns should follow.
Historically, Micron has not been kind to shareholders, and its shares are currently priced to reflect this. However, I believe that the nature of the DRAM industry has structurally changed for the better.
From a high-level, what makes Micron attractive is how essential it is to human progress. Without getting too professorial, humanity has had multiple waves of productivity gains over the past 12,000 years.
The first big improvement in productivity came from the agricultural revolution, which allowed humans to shift from hunting and gathering into high productivity farming and the division of labor.
The second wave of human productivity came from the industrial revolution, which harnessed machines to perform repetitive tasks on our behalf.
We are now in the third wave of human productivity: the information age. Like the machines of the industrial revolution, silicon chips are now automating and simplifying information-driven tasks.
Each of these waves relied on certain enabling resources; those that had them thrived, and those that did not fell behind. In the agricultural revolution, the enabling resources were arable land and nutrient-rich grains. In the industrial revolution, it was fossil fuels. And in the information revolution, it is silicon chips.
If you’re wondering why the United States government recently imposed restrictions on Chinese use of U.S. semiconductor technology (such as in the case of Huawei) – this action is like taking control of the world’s oil fields during the industrial revolution. It’s a way to keep China in check and maintain dominance of the global economy.
Micron makes memory; the chips that allow a computing system to store information (either temporarily or on a longer-term basis). Memory chips are quite different from the most commonly thought of semiconductors: logic chips, which have the job of processing data. Memory chips don’t process data, instead, their job is to feed data into processing chips, similar to a hopper in a factory, or a storage unit. Without memory, computers can’t function, and human productivity stalls.
Similar to our investment in Taiwan Semiconductor, the key to Micron is that they are part of a small cadre that controls the means of chip production. For logic chips, most companies create their own designs but outsource their production to Taiwan Semiconductor. In memory chips, production is vertically integrated; the same companies that design the chips also make the chips.
The DRAM industry has significantly consolidated over the past 10 years, and today there are only three players of consequence remaining – Micron, SK Hynix, and Samsung. SK Hynix and Samsung are both Korean, while Micron is American (underscoring Micron’s strategic value to the west).
Each of these memory makers has invested a tremendous amount of capital and expertise into their manufacturing footprint, and without many billions in investment and unincumbered intellectual property, it’s very difficult for a new competitor to emerge. Even though Micron is the third-largest player in DRAM, its capital expenditure budget for the upcoming year is expected to be $9 billion.
Even if you have the money and the intellectual property to reach a cutting edge node, there is tremendous execution required to keep up with market leaders who have decades of experience and know-how. Without that ability, a new player will likely be left behind, chronically unprofitable, and with little to show for its massive investment.
The advantages a Chinese company can bring to the memory business (low cost of labor and lower internal margin requirements) are not meaningful if the unit economics themselves aren’t competitive. In a scale chipmaking operation, the majority of the cost base isn’t due to labor but rather the remaining production costs in each chip (determined by manufacturing yield, technology process, and scale). There is a 20-30% cost difference per chip if you aren’t on-par with your competitors, and this is a 20-40% gross margin industry!
So, from a high level, I was attracted to Micron because it’s an essential business with very high barriers that keep out competitors, but it was also available at a great price.
The Investment Case For Micron
Unfortunately, the investment case for Micron is pretty complicated. Not only does Micron have a complicated business, but its cyclical nature also makes it difficult for an investor to ascribe a valuation based on its earnings power. Micron is a large-cap stock, but these barriers to investment keep out most investors and allowed us to acquire the position at an attractive price.
I much prefer a simple thesis to a complicated one. However, when I mention Micron’s complexity, it’s not that 50 things need to happen for the investment to work, instead, the complexity lies in understanding how Micron operates, what the technology is, and how the industry has changed. These are knowledge based barriers that can be broken down through research and understanding.
Of course, there’s a real chance I will be wrong about Micron, but I believe that its share price compensates us well compared to its risks. Further, my research suggests the business is improving for multiple reasons, which I outline here.
But, before I do, I won’t dance around the biggest flaw in Micron’s business that you have to understand. Despite operating in a DRAM oligopoly, Micron’s key products are commodity. That is, any of the three major players’ products are interchangeable.
Also, of Micron’s two business lines, only one is attractive in my opinion today: DRAM is attractive, NAND is not.
Micron’s business is 75% DRAM and 25% NAND, with the vast majority of their profits coming from the DRAM business. Think of DRAM as the short-term memory of your computer, and NAND as longterm data storage.
Getting back to the investment thesis, the performance of any commodity-driven company is ultimately determined by the conditions of supply and demand. If demand sustainably outstrips supply, you have a favorable operating environment, and if supply sustainably outstrips demand you have an unfavorable one.
Unlike Taiwan Semiconductor, memory companies have far less visibility into their end-customer demand (shorter order cycles), meanwhile, suppliers historically tried to steal share from one another by increasing their scale and thereby lowering prices. This led to multiple historical supply/demand imbalances and large gyrations in product prices, margins, and earnings.
I believe the supply/demand imbalance in DRAM has structurally changed and is now in a much-improved position. NAND will take a longer time to reach such an equilibrium, but, since DRAM is virtually the entire profit pool for Micron, I believe the business has shifted from being a mediocre one to a pretty good one with high returns on invested capital.
There are two key points to understand:
1. DRAM supply is now constrained by the limitations of physics – The physical limitations of chip design now make it difficult for DRAM makers to rapidly increase their supply without incurring inordinate costs. Counter-intuitively, this is a significant positive for the industry.
2. Demand has increasingly strong secular growth trends – Demand for Micron’s products are in a fundamentally improved position due to the increasing value of data. This drives memory and storage bit demand.
Supply-Side Improvements in DRAM
If you consider a logic chip, the underlying functional unit of that semiconductor is called a transistor. There are many transistors within a processor, and each transistor is a circuit that either allows current to pass or blocks it. This action of blocking/unblocking electricity allows a transistor to transmit data in the form of 1’s and 0’s (on and off states).
Memory chips are a bit different in that they have slightly more complexity in their underlying functional unit. Unlike a logic chip that is largely a series of transistors, memory chips are made of ‘memory cells.’ Memory cells typically have a ‘1T1C’ structure: meaning one transistor and one capacitor per cell. Think of a capacitor as a temporary storage location that maintains the charge a transistor might pass through. The capacitor is what provides ‘memory’ of 1’s and 0’s.
A capacitor has a larger volume requirement than a transistor, and if you shrink it too much it won’t be able to hold a charge, rendering it useless. Since each memory cell has both a capacitor and a transistor, the amount that you can shrink it down is less than you can for just a transistor alone. In other words, for two identical silicon wafers, you can theoretically make more logic chips than memory chips.
Today, logic chips are continuing to scale down at a rapid pace (Taiwan Semiconductor is already rolling out their 5-nanometer process with plans to roll out their 4-nanometer and 3-nanometer processes in the coming years), meanwhile, memory chip makers are theoretically capped at around 10 to 15-nanometer processes. Today, memory makers are largely already at these mature nodes with little runway remaining.
So, what does this all mean? It means that the cost to fit an incremental memory chip onto each of your wafer capacity has gone up exponentially. We’ve seen this play out over the past few years. For example, the cost to add 1% more memory chips to a wafer has increased 7x over the past 7 years! We are now at the point of significantly diminishing marginal returns.
Memory chip manufacturers only have two ways to make more chips, either by 1) improving the process technology (fitting more chips into each wafer of your existing manufacturing footprint) or 2) by adding new manufacturing lines at the current process technology.
Improving process technology is usually the far superior option of the two because by fitting more chips on every wafer you not only increase capacity, but you also lower the cost of each chip you make. This ensures an increased level of cost competitiveness in addition to expanding your supply. On the other hand, when you add additional manufacturing lines to a factory, your underlying cost structure per chip does not change, historically making it the inferior choice.
In the past, Samsung, which has the largest share in the DRAM market (around 50%), had been very aggressive in its investments in process technology. Samsung invested especially heavily at times when the memory market was weak to further drive down prices, steal share, and shake out weak competition from the market. It worked, and now only three players remain.
Samsung was able to enjoy these investments because 1) it allowed them to maintain their cost leadership in DRAM, and 2) they were cross-subsidized by other divisions within their conglomerate during periods of DRAM weakness.
Now that we’ve mostly reached process maturity (for the reasons mentioned above), incremental investment in Samsung’s process technology will no longer provide a significant cost advantage. To increase supply Samsung therefore has to employ a mix of new production lines and update remaining lines that aren’t yet on the latest nodes.
Similarly, due to the state of process maturity, any earlier cost advantage Samsung enjoyed over Micron and SK Hynix is now quite limited. The cost curve is now flat, and an incremental investment in process improvement will yield limited advantage despite the high cost of executing such a change. Process improvements now offer a much lower return on investment.
Going forward, much more supply growth will come from adding new manufacturing lines rather than trying to shrink down the manufacturing process, and this I believe is the key insight of the investment. With less rampant supply expansion in boom and bust times, this will lead to far bettermatched supply and demand as well as improved profitability for all three players. Intuitively, you’d think a higher cost of capital expenditures would be a bad thing, but in an industry where oversupply is catastrophic and players have historically been quite aggressive, it’s exactly the opposite.
I believe the prisoner’s dilemma is an economic theory that is applicable in this situation. Over the past decades, each of the memory players has been ‘cheating’ behind the other’s back to get ahead. When all players cheat, everyone loses the game, (except for customers who enjoy the low prices). What’s different now is that the laws of physics have eliminated the incentive for cheating, and this breaks the prisoner’s dilemma.
I believe we’ve achieved a new equilibrium where the players in this industry are finally incentivized to behave rationally.
This thesis has been in play for a couple of years and we’ve already seen it bearing fruit. However, any improvements so far enjoyed have been masked by the cyclical memory correction that began in early 2019. Over this period each of the three memory players described their approach to expanding their supply as an effort to ‘match demand,’ and in 2020, capital spending declined significantly across the industry to match this reality.
This is a critical change for the industry because the DRAM makers have shifted their production decisions from “beating the competition” into forecasting what actual demand levels are likely to be at their given market share. This creates a much more rational supply/demand balance and superior profitability for all firms. In fact, we’ve seen improved profitability across the most recent market cycle in DRAM: all players earned reasonably good profits even at the bottom of the cycle (vs losses in prior cycles).
In short, historically the DRAM market was a painful boom/bust industry with several players overproducing to lower their cost structure and gain share from the other. These overproduction periods caused massive price and profitability swings.
Today, the industry has evolved into an oligopoly of just three players where it’s no longer in the best interest of any one player to undercut the other in a parasitic game-theory dystopia.
Going forward, I expect DRAM to be a good margin and return industry with few players and more rational behavior. Both the peak and trough earnings should improve and offer attractive returns on invested capital.
Demand Side Has Increasing Secular Growth Trends
Earlier I described memory chips as the hopper that feeds data into a computer’s processor. That context is important to understand how software changes will increase the amount of memory required by the computing industry at large.
Artificial intelligence and machine learning offer a paradigm shift for software and semiconductors. What AI allows software to do is write its own code rather than rely on a software engineer to do it. The way AI does this is by gathering massive amounts of data from users, sensors, and other sources, and then trains itself off that data to make inferences around how it should behave.
For software companies to have market supremacy in an AI world, they will need to have the most and best data to fuel their growth loop. Having the best data allows machines to improve their software and offer the best service to customers. Whoever has the best service will have the most customers, and in turn, the best dataset to further train the algorithm. This is a reinforcing cycle that repeats itself.
It’s been said before that data is the new gold, and I think memory and storage act as picks and shovels. There are dramatically increasing demands placed on the chips that process these data, and without massive investments in the datacenters that process, store, and transfer this information, a technology company will fall behind. This benefits memory and storage bit demand.
Further, it’s important to understand that machine learning algorithms are incredibly data-hungry, and they suffer diminishing marginal returns on additional data. This means that a company needs even more data every time it wants to achieve an algorithmic lead. As a rough rule, to double the efficiency of the leading deep learning algorithm, the machine requires 10x more new data. To double its performance again, the machine requires yet another 10x increase (a 100x amplification from the start), and so on.
Not only are these massive quantities of data, but the way data is parameterized and analyzed by these algorithms is also very data-intensive. Machine learning algorithms often overparameterize input data by design because it can’t function like a human brain. This means that every piece of data inputted may have hundreds or even millions of categorization variables attached to each data point, and as more data is inputted, additional variable types are identified and added. We have a true data explosion problem, and the semiconductor companies that analyze and process these data stand to benefit since they are the ultimate bottlenecks.
Historically, the bottleneck of processing data had been the processor itself, but going forward, the demands of computing are shifting the bottlenecks from just the processor to networking, memory, and storage. Micron addresses two of these.
While I’ve focused on the center of the computing world today: the powerful datacenters where many of these computations happen, the edge computing nodes that generate most of these data – PCs, laptops, smartphones, the internet of things (IoT), etc. are all creating, consuming and sharing more data than ever as well. With advances such as 5G, each of these end nodes can transmit and store more data, and this requires higher amounts of memory and storage on the edges too.
While the demand side is much harder to forecast, I believe these factors all bolster structural bit growth demand in DRAM and NAND. Higher demand in addition to tighter supply should lead to an improved supply/demand dynamic over the medium to long term.
What Could Go Wrong?
Every investment has multiple ways that a thesis might not play out, and in Micron’s case, there are several ways I can be wrong. Compared to some of our other investments, Micron is less durable given its capital intensity, cyclical underpinnings, and lower gross margins.
The three largest areas of concern for me with Micron are: 1) new entrants into the DRAM and NAND markets (likely fueled by Chinese sovereign capital) 2) new technologies emerge that break the détente in DRAM and 3) the NAND business serves as a drag on overall company performance. I’m monitoring each of these risks and believe that we are well compensated for them by the low stock price.
Putting Numbers to the Investment Case
I won’t share what I think the company is ultimately worth, but I believe there are two sides to this thesis – the first is that the earnings power of the business will continue to improve through the cycle, and the second is that with reduced cyclicality and increased understanding by investors, its valuation multiple should expand.
If Micron continues to improve it’s through cycle margins while revenues grow at mid-to-high singledigits on average, I believe we will see even higher earnings growth. Compounding this intrinsic growth with reasonable multiple expansion (Micron trades today at ~6x trailing EBITDA) this will result in a healthy IRR, all while we own a well-positioned strategic business, with limited downside if the thesis doesn’t play out as expected.”