Micron vs. SK hynix: One Stock Rules AI Memory. The Other Has More Room to Run.

Two companies are quietly fighting for control of the most important component in AI - and Micron (MU) just beat the reigning champ, SK hynix (SKHY), on margins. Both supply the memory NVIDIA’s (NVDA) chips can't run without.

But to understand why this fight matters so much right now, we have to start with why the memory business was never supposed to be this exciting.

Memory chips have always been essential to almost every electronic device - from smartphones and laptops, all the way to servers and gaming consoles. But unlike processors, the memory business has largely been cyclical. 

Take the typical 8-gigabyte Dynamic Random Access Memory (DRAM) - the ones that are used in many of our modern devices. They are essentially the same “memory stocks” used across different manufacturers and models. 

There may be slight differences in timing and speed, which is important for PC gamers, but at the end of the day, they serve the same purpose. That’s why DRAM prices have largely been a matter of supply and demand rather than any real brand differentiation. 

When demand outpaced supply, prices surged, and memory producers reaped massive profits. But once production caught up and more supply came online, prices fell, inventories piled up, and earnings fell sharply. 

That cycle has been going on since computers and phones became common worldwide. But today, the AI revolution is changing that narrative. 

Data centers need vastly more memory than your typical office building - and no, your 16GB DRAM sticks with the nice heatspreaders aren’t going to cut it. Data centers need specialized memory, and they need it yesterday. Memory from a relatively common computer component has become a potential bottleneck for the entire AI infrastructure buildout. 

But what kind of memory does AI need? Who makes it? And of course, the million-dollar question: Which companies stand to benefit from the new AI memory supercycle, and which one will come out on top? 

Let’s talk about it. 

How AI Became the Game-Changer for Memory

When we think about AI hardware, GPUs are usually top of mind. But a GPU’s performance depends on more than just raw compute. Modern AI models contain billions, or even trillions, of parameters. 

When an AI model is trained, it needs to handle large volumes of data, including words, images, and videos. During training, the AI model needs to process all this data to adjust its parameters until it learns the underlying patterns. 

Then comes inference. This happens when you ask a chatbot a question, generate an image, or handle any other request. 

Essentially, this is when the model applies what it has already learned to generate outputs that are (hopefully) helpful to the end user. Although a single request is less demanding than training, once millions of users run the model simultaneously, it’s easy to see how memory requirements can add up quickly. This feeds a never-ending flywheel.  

As GPUs have become more powerful and, by extension, can handle more work, the amount of data they consume has increased. This is why memory capacity has become just as essential as compute power. If memory can’t supply data fast enough, the GPU sits idle instead of doing its job, collectively wasting millions of dollars’ worth of computing power. 

This is where the memory bottleneck comes in. 

What HBM Is, and Why AI Can't Run Without It

Traditional DRAM was never designed for the bandwidth demands of today’s AI workloads - and certainly not at scale. That’s where a different type of memory comes in: High-bandwidth memory (HBM).

HBM used to be less popular than it is now. It was used mainly in specialized applications like graphics development and high-performance computing. But now, it has grown into one of the most important layers of the AI tech stack. 

HBM solves the scaling problem by taking a completely new approach. It stacks several memory dies vertically and places them close to the GPU. With this approach, the data can travel along a much shorter electrical path, allowing it to move much more quickly while using less energy. 

For AI training and inference, these new advantages are game-changers. An advanced AI chip can’t deliver peak performance on traditional DRAM. But with HBM, AI models have much larger memory capacity, lower power usage, and more efficient performance for demanding workloads. 

Today, the latest generations of AI chips from NVIDIA and AMD (AMD) rely on increasing amounts of HBM. In fact, according to a recent report by Morgan Stanley, a newly launched AI chip uses 7.2 times as much HBM as previous generations, while a full AI system uses around 65 times as much. 

This amount will only keep growing as these companies produce more sophisticated AI chips, which will require ever more memory capacity. 

AI models are also becoming larger and more complex. 

Labs like Anthropic and OpenAI are releasing more sophisticated models, capable of generating images and videos with higher resolution, supporting millions of simultaneous users, and handling longer context windows, which is how much data a model can remember at any one time. 

Beyond those, AI superclusters are also rapidly scaling to support frontier AI models. Previously, AI models ran on only a few thousand chips, but now the most advanced models are trained on clusters with tens of thousands of chips. Many hyperscalers are likely planning even larger clusters with hundreds of thousands of chips as they roll out more AI products and services. 

So, yes, the AI infrastructure buildout sorely needs HBM memory and the companies that supply it. But there’s a caveat - and it’s a big one. 

HBM is much more difficult to manufacture than DRAM. It requires more complex production processes and advanced packaging. That means billions of dollars in investments in fabrication plants, build processes, and supply chains, not to mention the technical knowledge to manufacture such chips. 

It also takes time to ramp up the production capacity needed to produce more HBM. This has led to a tight supply situation, giving the memory manufacturers better leverage in customer negotiations. So it’s not surprising that the handful of companies with the expertise and manufacturing capability to make HBM at scale are reaping the benefits. 

Furthermore, companies that can secure design deals for their AI memory tend to lock in years of sustained demand, giving them greater pricing power and healthier margins. 

And even if the number of AI chips shipped to customers remains unchanged, memory demand could continue to surge simply because each new chip generation contains more HBM than the last. In other words, HBM demand can grow rapidly even without any growth in chip shipments.

And that creates a pretty compelling setup for HBM manufacturers.

Demand is growing, supply is difficult to scale, and each new generation of AI hardware consumes more memory. That combination gives qualified suppliers something the memory industry hasn't always had: greater pricing power, better margins, and much greater visibility into future demand.

So the opportunity is clearly there. But which companies actually have a shot of grabbing it? 

The Global Memory Oligopoly  

Unlike other tech markets with dozens of players, the memory industry has grown significantly concentrated. The global DRAM market is essentially an oligopoly dominated by only three players: Micron Technology (MU), SK hynix (SKHY), and Samsung Electronics (SSNLF). Of the three, Samsung is the only one that’s trading over the counter (OTC) in the U.S. 

Together, these three companies account for nearly 90% of global DRAM revenue, which continues to surge to new highs.

Such massive industry growth has been fueled by better memory prices and surging demand as AI data centers add more HBM and DRAM content to their infrastructure. 

Translation: Not only do these companies own the biggest slice of the pie, but that pie is also growing fast. 

Again, HBM production has a ridiculously high barrier of entry, so not every company can just muscle their way into some market share. 

And within that concentrated HBM industry, SK hynix currently leads with a 58% market share. Meanwhile, Micron and Samsung are tied at around 21%. 

And you know what? That lead didn’t happen by accident. 

SK hynix: How It Built a 58% Lead in AI Memory

SK hynix started investing in HBM well before the current AI boom, giving the company a significant head start when demand suddenly took off.

Now, when I said "early," you might be thinking maybe 2015 or something. That’s around the time when compute demand was starting to pick up, and the first inklings of AI technology were making their rounds in the market. 

But no. SK hynix started developing HBM in 2009 for high-performance GPU computing and released the first iteration of the chip in 2013. 

And you know what’s funny? When the company’s first HBM chip hit the market… nobody really cared. At that point, there was no need for the “unnecessarily” high speeds and large capacity that HBM offered. That made sense; who in their right mind would invest in high-speed memory that their CPU and GPU can’t even utilize to its full extent? 

So the technology was arguably a decade ahead of its time. Vice President Myeong-jae Park said that back then, the HBM design department was “described as somewhere off the beaten path.” That made developing HBM's second generation challenging - because, again, why improve something that no one wanted? 

But technology has never been shaped by people who only look at what is. The real innovators always look at what could be.

And that's exactly how SK hynix approached HBM. There was no guarantee that the technology would ever pay off, or that the market would actually scale. 

“Nevertheless,” Park continued, “we firmly believed that HBM was an opportunity to showcase SK hynix’s unique technological capabilities, and that once we developed the best products, services to utilize them would naturally emerge in the market.” 

And because of that belief, the company was the first to commercialize multiple generations of HBM, helping it secure the top spot as the go-to supplier for AI memory when the boom came around. 

More recently, SK hynix completed the world’s first HBM4 development and readied a mass-production system in 2025. 

Compared with the previous generation, HBM4 delivers twice the bandwidth and 40% greater power efficiency. The company expects this new memory standard to enhance AI performance by up to 69%, easing the data bottleneck and reducing data center power costs. 

The result? HBM now accounts for a growing share of SK hynix's product mix, and because it commands far higher margins than ordinary DRAM, the shift has lifted the company's profitability. It’s no exaggeration to say that the company’s HBM business is now its most valuable asset. 

Perhaps the clearest payoff of all that early groundwork is SK hynix's relationship with NVIDIA. NVIDIA's GPUs sit at the center of AI training, and HBM is one of the most critical components inside them. So, years of successful product qualification have made the company the preferred supplier for many of NVIDIA’s flagship AI platforms. 

Decades of scaling manufacturing capacity, combined with the recent AI windfall, have also left SK hynix in a far stronger financial position. The company is now leaning on both its scale and its balance sheet to continue expanding and investing in the next generations of HBM.

In fact, Moody’s even noted the improvement and upgraded SK hynix’s rating from Baa1 to A3, citing its stable outlook and solid earnings power. That rating change moved SK hynix from the medium investment-grade range to the upper-medium range. For a company operating in an industry historically defined by boom-and-bust cycles, that’s a significant step up.

Part of that rating increase may have come from SK hynix's recent earnings: triple-digit year-over-year revenue growth, quarter after quarter, driven by strong AI memory demand. DRAM product sales have also been outpacing the broader memory market. 

Best yet? That growth hasn't come at the expense of margins. SK hynix’s recent quarter saw a gross margin of 83% and an operating margin of 76%, driven by higher DRAM and NAND prices and ongoing cost improvements. That’s a level of profitability that’s rarely seen in the memory business. 

Rare, but not unique… and certainly not the highest. We’ll get to that later. 

On top of record margins, SK hynix has locked in a string of Long-Term Agreements (LTAs) with major customers, with more under negotiation. These agreements include built-in fulfillment mechanisms, giving the company durable revenue visibility that extends well beyond any single quarter.

So all this points to the same conclusion: SK hynix’s early bet on HBM has paid off, turning into a high-margin business and market leadership in what’s turning out to be one of the most critical components in the AI infrastructure stack.

The Risks That Come With the Memory Stock Crown

But of course, being an industry leader also means having a target on your back. Every new HBM generation raises the bar for performance, yield, and packaging, and any stumble in qualification or manufacturing could allow hungry competitors to close the gap. 

With all the attention SK hynix is getting, it’s no wonder investors are now scrutinizing every nook and cranny of the company’s execution plan. Those consecutive record results also contribute to ever-growing market expectations. And when a market darling falls short of expectations, the fallout could be catastrophic. 

There's also a concentration problem. A large chunk of SK hynix's HBM demand traces back to NVIDIA and a handful of hyperscalers. That concentration provides strong near-term visibility, sure, and the massive increase in AI spending will do wonders for SK hynix’s income statement. But it also means the company is exposed if AI spending cools or if customers decide to diversify their suppliers.

Geography adds another layer of risk. SK hynix operates manufacturing facilities in China, leaving it more exposed to geopolitical friction than some rivals. Recent reporting suggests the company is already looking into alternative equipment suppliers as U.S. export restrictions tighten. But replacing established suppliers and reconfiguring a mature manufacturing operation take time and money and carry their own set of risks. 

Then there's the sheer scale of its own ambition. SK hynix's aggressive investment in HBM capacity, packaging, and DRAM production is meant to defend its lead, but it's also a big bet that could backfire at any moment. If AI demand slows or supply catches up faster than expected, the returns on all that spending could fall short. 

And even if everything did work out on the AI and HBM front, DRAM and NAND demand are still highly cyclical, and SK hynix is still earning meaningful revenue from these segments. If consumer demand weakens, these segments could pull down the overall numbers. 

Lastly, with a market this rife with opportunities, you just know that competitors aren’t sitting still while all this is happening. 

In fact, one of them is closing the gap faster than anyone expected. 

Micron: The Challenger Closing the HBM Gap on SK hynix

Micron has quickly emerged as SK hynix’s strongest challenger. 

Now, while HBM is a core part of its growth narrative, Micron’s strongest claim to the crown is actually its more balanced portfolio. Sure, HBM gets the headlines, but the company also holds notable positions in conventional DRAM, enterprise solid-state drives (SSDs), and data center storage. That means the company has multi-pronged exposure to the AI market and can capture demand where it shows up, not just when hyperscalers need more memory. 

The company's real differentiator, though, is discipline. Rather than chasing volume growth as memory makers often did in past cycles, Micron has deliberately prioritized manufacturing efficiency and steered investment toward its highest-value products. Right now, those are HBM, server DRAM, and enterprise SSDs. 

That restraint shows up directly in the bottom line, with some of the strongest profitability metrics in the company's history. In fact, Micron had better margins than SK hynix in the recent quarter, at 84.6% gross and 80.4% operating. 

On the tech front, the company has aggressively invested in next-generation HBM and has begun volume shipments of HBM4 for the NVIDIA Vera Rubin platform. It’s also currently developing HBM4E, with high-volume production expected in 2027. This has positioned the company ahead of future demand for advanced HBM generations. 

Even better, Micron also has the distinct advantage of being the only HBM supplier based in the U.S. That gives it potential access to federal incentives and easy integration into the local supply chain of other U.S.-based hyperscalers. In fact, its HBM capacity has been sold out well in advance. 

But that kind of visibility doesn't happen by accident. Micron has built it deliberately through multi-year Strategic Customer Agreements (SCAs) with major customers, which typically run for five years. The company has locked in a meaningful share of its DRAM and NAND volume well into the future, reducing its exposure to the boom-bust swings that have historically defined the sector. And in its latest earnings call, CEO Sanjay Mehrotra announced that the company has 16 of these SCAs:

Micron delivered an exceptional fiscal Q3, with significant records in revenue, gross margin and EPS (earnings per share) — all exceeding the high end of our guidance. Demonstrating Micron’s position as a leader enabling the AI era, our data center revenue exceeded $25 billion in fiscal Q3, or an annualized run rate of over $100 billion. Our data center SSD revenue exceeded $5 billion, more than doubling sequentially. DRAM and NAND industry demand continues to significantly exceed industry supply. We expect tight conditions to persist beyond calendar 2027 as a result of AI-driven demand across all segments coupled with structural supply constraints. We are excited to announce that we have now signed 16 strategic customer agreements, or SCAs, which we expect will fundamentally transform our business model. 

So, Micron was late to the party, but it’s already making meaningful strides to correct that. And its biggest advantage is actually the simplest and most obvious: it’s the smaller player.  

SK hynix already dominates the HBM market, which means there’s only so much additional share it can capture. Micron, on the other hand, has a much larger runway. It already secured an HBM design win with NVIDIA, which is widely considered SK hynix’s most important customer. That’s a foothold no investor can ignore. So every percentage point of market share Micron gains represents a meaningful increase in revenue, especially as the overall HBM market continues to expand.

That gives the company a potentially powerful combination: a growing market, increasing production capacity, and plenty of room to take share from the market leader.

Can Micron Keep Taking HBM Share From SK hynix?

However, Micron’s position as underdog cuts both ways. Yes, execution has been exceptional so far; it’s securing key design wins with major hyperscalers, and clients are committing under long-term contracts. 

But SK hynix still holds the edge in manufacturing experience and customer relationships built up over a much longer track record. That’s a tough moat to cross, and any mistakes could affect Micron’s future runway. 

And like SK hynix, the scale of its own investment carries risk, too. Micron's spending on advanced DRAM, HBM production, and U.S. manufacturing represents large, long-duration projects that require continuous investment, underwritten by all that AI demand. Right now, if I’m being optimistic, these bets could pay off handsomely if the AI buildout continues as expected. 

But what if that demand slows down, or dries up? 

And, like every memory maker, Micron isn't immune to the industry's underlying cyclicality. HBM is a growing share of the business, but, again, DRAM and NAND are still at the mercy of supply and demand. 

Micron vs. SK hynix: Which Company Will Win The AI Memory Supercycle?

So, who wins the AI memory supercycle?

SK hynix built a decade-long head start that's translating into record profits today, and it's not giving up that lead without a fight. 

Micron, on the other hand, has proven that being the challenger doesn't automatically mean being the loser. Discipline and diversification are working wonders for its bottom line, helping it achieve some of the best margins in its - and the industry’s - history. 

And we haven't even touched on Samsung, which is still clawing its way back into the conversation with HBM4 as its equalizer.

So, at this point, my answer might surprise you. There’s no clear winner yet. 

SK hynix is in the strongest position today. Micron arguably has the most room to grow. And Samsung has the scale and resources to disrupt the market if it can close the HBM technology gap. Right now, with momentum behind AI spending, all three have the opportunity to grab that crown. 

That's also what makes this matchup so interesting. The AI memory boom is rewarding companies that can deliver more advanced memory and expand production fast enough to keep pace with demand. 

And if AI continues driving memory requirements higher with every new generation of hardware, this supercycle could have much, much further to run.