Quick Hit: South Korea’s biggest tech firm, Samsung, has revealed that its Q3 operating profit is going to soar to 107.4 trillion won ($80 billion), an astronomical increase from last year’s Q3 of 782%. This is the very first time that any company in South Korea has reached such figures for a single quarter. The reason is not phones and televisions. It’s semiconductor memory, the unsung hero of AI technology. Here’s how we came to this point.
The Number That Broke the Spreadsheet
How about we give some context to that $80 billion number?
Samsung’s total operating profit in all of 2025 was 43.6 trillion won. In three months of 2026 alone, it produced more than 2.5 times that. Year-over-year growth is not even an upward curve; it is more like a vertical line. Analysts surveyed by Yonhap Infomax projected roughly 105.7 trillion won in operating profit. Even Samsung’s conservative projections were higher than any outside estimates.
This has nothing to do with the regular business cycles. It’s a structural change in the real requirements of the world economy.
And here’s a quick answer as to why that happened. AI data centers have become gluttons for whatever memory production the world can make.
The Truth: The AI Memory Bottleneck Demystified
Memory Chips: The Forgotten Bottleneck
When you hear about AI infrastructure, you immediately think of GPUs like Nvidia’s H100s, B200s, and even the Vera Rubin platform coming out soon. They’re the main stars of the show. However, every GPU becomes pretty much useless if there isn’t memory fast enough to supply them with necessary data.
Why is it so? Because of one simple fact: AI processors can be idle while waiting for data from the memory. If memory doesn’t provide data fast enough for the GPU, you just spent a lot of money on a fancy space heater .
HBM (High Bandwidth Memory) is a special type of memory when chips are mounted vertically and connected using thousands of fine connections, allowing the transfer of information in terabytes per second instead of gigabytes.
To understand the scale, take a look at Micron’s HBM4 that was released into mass production in early 2026 for Nvidia’s Vera Rubin platform. It allows over 2.8 TB/s of bandwidth, roughly 2.3 times higher bandwidth compared to the previous generation called HBM3E.
Why This Time It Is Different
The cycle has been known in the history of memory chips development for decades. The prices go up, people start producing in excess, the prices fall, and manufacturers cut back on production.
The difference is that this time, it is caused by completely different structural factors that make the demand for memory chips unique compared to those for PCs and smartphones.
Traditional servers used between 256 and 512 GB of DRAM. Inference AI servers for large language model processing may require terabytes of memory chips. Training the frontier models requires even more memory. But the thing is not that it is required by the training phase, but by the inference process.
As stated by Micron’s Chief Business Officer Sumit Sadana, an agentic AI workload that uses several steps, tools, and logical reasoning requires 5 to 30 times more tokens compared to a standard chatbot. The more tokens are required, the more memory bandwidth will be consumed. And that means that more memory chips would be required.
As Elon Musk said, memory chip production capacity increased by 20%, but the demand increased by 200%.
An Unspoken HBM Trade-Off
There’s an unspoken trade-off with HBM, which is the reduction in the capacity of traditional DRAM when it comes to making HBM.
As per Sadana from Micron, the ratio works like this: manufacturing 100 bits of HBM3E causes DDR5 yield losses of 300 bits, at a ratio of 3-to-1 . With the HBM4E, that ratio can even go up to 4-to-1 .
In simple terms, for every wafer manufactured for HBM, it means one less wafer making the DDR5 that would be used for servers, laptops, and smartphones. The AI industry’s gain becomes the loss for everybody else.
This is why there has been a case where “data center customers [are] unable to secure more than half of the memory volume they seek,” despite paying top dollar.
From Bust to Boom: How Samsung Made It to This Point
The Dark Days of 2025
One year ago, Samsung was not in such an enviable position.
At the start of 2025, Samsung was having trouble. HBM sales lagged those of competitor SK Hynix, which landed the much-sought-after spot of becoming Nvidia’s HBM vendor. Issues with Samsung’s HBM3E chips were apparently related to their high temperatures and low yields, which prevented qualification. The company’s managers apologized for their sub-par performance.
Samsung was not innovating in mobile phones. The foundry business was being outdone by TSMC. Meanwhile, the memory business traditionally Samsung’s profit center was mired in a downcycle.
2025 operating profit was 43.6 trillion won, up 33% over last year, yet still well below pre-pandemic highs.
The Turnaround: Four Quarters of Records
After reaching critical mass in terms of AI infrastructure deployment.
The fourth quarter of 2025 became the first record with 93.8 trillion won in revenue and 20.1 trillion won in operating profit. And there was only more to come.
Q1 2026: 133.9 trillion won in revenue and 57.2 trillion won in operating profit all-time highs, more than Samsung managed to make in the whole of 2025. The DS (Device Solutions) segment that contains semiconductors generated 81.7 trillion won in revenue and 53.7 trillion won in operating profit. Memory alone made records in both categories.
Q3 2026: This is the number that hit the headlines. Operating profit: 107.4 trillion won. Revenue: 195 trillion won.
These are four quarters of records where each subsequent quarter was bigger than the previous one.
Technical Changes That Occurred
However, Samsung’s turnaround was not only a result of market changes but also certain specific technical steps taken by the company which put it on the track to meeting the growing demands of the AI sector:
Mass production of HBM4. Samsung has been the first company in the market that launched mass production and shipment of sixth generation HBM, HBM4 for the Nvidia’s Vera Rubin platform. It was a sharp change compared to the problems with HBM3E the company faced before.
SOCAMM2. Besides, Samsung began the mass production of SOCAMM2 (Small Outline Compression Attached Memory Module 2), memory specifically designed for AI and HPC workloads. These modules are capable of providing up to 2 TB of memory per CPU with 1.2 TB/s of bandwidth .
Next generation PCIe Gen6 SSDs. Samsung created and shipped SSDs specifically designed for AI KV cache storage requirements.
It is important to pay attention to the technical aspects of the problem since they explain how Samsung managed to meet the market demands.
The Market Share Landscape: Who Manufactures What
DRAM: The Samsung Monopoly
Samsung commanded a whopping 39.4% of the worldwide DRAM market by revenue during the second quarter of 2026, based on information from TrendForce.
Here’s the breakdown:
- Samsung: 39.4% share, generating $60.98 billion from DRAM sales (up 63.4% quarter-over-quarter)
- SK Hynix: 24.9% share, generating $38.59 billion (up 37.9%)
- Micron: 23.3% share, generating $36 billion (up 65.5%)
The total revenue generated by the DRAM market stood at approximately $155 billion in Q2 2026, representing a 60% increase quarter-over-quarter.
HBM Race: A New Storyline
Now, let me tell you where it gets more exciting. Even though Samsung tops the general DRAM race, it is not the case in the high-end HBM race, which is the most profitable one.
According to Counterpoint Research, SK Hynix had about 50% of the HBM market, Samsung got 33%, and Micron’s share dropped to 18% in Q2 2026 from 21%.
However, this HBM battlefield keeps changing fast, as Samsung’s mass HBM4 production, and Micron’s shipments of HBM4 for Vera Rubin make this market more complicated. If Micron lost its share in HBM, yet its overall DRAM share grew significantly, there is a clear reason why this happened. HBM prices skyrocketed faster than the conventional DRAM ones in this cycle.
Why? Because the production of HBM is destroying the capacity of the “ordinary” DRAMs.
The Micron Story: Bridging the Gap
The success story of Micron merits a mention. For the third quarter of 2025, Micron trailed SK Hynix in DRAM market share by 7.5 percentage points. But for the second quarter of 2026, this gap had been narrowed to 1.6 percentage points .
Micron saw a quarterly increase in its DRAM revenue of 66% compared to the growth rates of Samsung (63%) and SK Hynix (38%).
It wasn’t due to HBM. Conventional DRAM is responsible for the company’s success. Due to the increase in production of HBM, prices of DRAM have skyrocketed.
The Statistics Behind the Statistics: The Memory Market Explained
Price Increases Which Don’t Follow Traditional Economic Patterns
In order to get an understanding of why Samsung was experiencing such a spike in profitability, we need to look at what was happening with the prices of memory chips.
In Q1 2026, DRAM contract pricing increased by 90% – 95%. Meanwhile, NAND flash pricing had climbed 55% – 60%.
According to a Goldman Sachs research report released in April 2026, the DRAM price increase estimate for the year had been changed to 250% – 280% from a range of 150%. In contrast, the NAND price hike had been forecast to increase to 200% – 250% compared to 100% initially.
The Revenue Math
Let’s do the numbers to get a better idea of how things stand for Samsung.
Assume that your sales increase by a factor of two while your price doubles; you’ll double your profits. But when your production costs remain constant and you increase your volume of production significantly, you multiply your profits rather than doubling them.
For instance, the memory division, which had already the highest production capacity worldwide, saw a drastic jump in its average selling prices for both DRAM and NAND technologies. Operating profit for the DS division of 53.7 trillion won for Q1 2026 constituted 65.7% of its revenues. That is quite high for a semiconductors manufacturer.
It was estimated in Q3 2026 that the division of memory alone generated around 80 trillion won in operating profit, and another 30 trillion won from NAND.
The Consumer Electronics Problem
However, there is an opposite side to all of this. If Samsung Semiconductor generates huge profits, Samsung’s consumer electronics divisions suffer from the negative effects.
The DX (Device eXperience) division, which produces phones, TV sets, and appliances, reported record operating losses for two consecutive quarters in 2026. And why? Simply because the price of memory is one of the key components of these devices, and when it increases by 200-300%, it is difficult to remain profitable.
Memory makes up around 30% of the bill of materials for most consumer electronics products. When its cost goes up threefold, profit margins become extremely small.
It is simply an inherent trait of the market at this point. Consumers of AI products do not mind paying anything for the memory. But consumers of consumer electronics do not.
The Advanced Corner: Technical Deep Dive
For those readers who wish to understand the technology in greater depth.
HBM Architecture: Why Stacking Matters
High Bandwidth Memory (HBM) functions through stacking DRAM dies vertically with the use of TSVs microscopic wires that go through the silicon. That way, there can be thousands of connections between the layers, which provides huge parallelism.
The basic HBM4 stack, such as Micron’s 36 GB 12H stack, consists of 12 DRAM dies stacked atop a logic die. The logic die is responsible for the interface with the GPU or AI accelerators. Data is moved through the stack at the rate of more than 11 Gb/s per pin, with thousands of pins working in parallel .
The speed difference is massive. A standard DDR5 module could provide a 50-100 GB/s bandwidth. In contrast, the HBM4 provides more than 2.8 TB/s, which is 30 times faster.
Why does it matter for artificial intelligence? The thing is that large language models are memory bandwidth bound, rather than compute bound. While the GPU is capable of performing trillions of operations per second, it cannot perform them if it’s waiting for the data.
The 4-to-1 Ratio: Why HBM Starves Conventional DRAM
The revelation by Micron regarding the tradeoff ratio between HBM and DDR is among the most critical pieces of information for grasping the current situation.
For each 100 bits of HBM3E produced, there is a reduction in the production of 300 bits of DDR. For HBM4E, the ratio may rise to 4 to 1.
It is all because of the size of the wafer and the level of process complexity. The dies used for HBM are bigger than those used for the common DRAM because they have the TSV and extra logic. Moreover, their yields are lower because of the stacking involved. And the packaging of many dies requires lots of effort and investment.
When they choose to produce more HBM, they are depriving themselves of DRAM production capacity.
SOCAMM2: The New Memory Form Factor
SOCAMM2 (Small Outline Compression Attached Memory Module 2) is very different from other form factors such as DIMM.
The conventional form factor of DIMM is used for socketed mounting where signaling is quite typical. However, SOCAMM2 employs the technology of compression-attached memory where the module is mounted on the board with a retention clip rather than into a socket. The advantages are higher density, improved signal integrity at high speeds, and lower power consumption.
Micron’s SOCAMM2 memory modules vary from 48GB to 256GB. Currently, the version of 192GB is produced in large quantities for the Nvidia Vera Rubin platform . One CPU may have up to 2TB of memory and 1.2 TB/s of bandwidth thanks to SOCAMM2 modules.
Such technology becomes important for AI inference due to limited CPU memory bandwidth for tasks such as embedding lookups and KV cache management.
The Future: How Long Can This Last?
The Bull Case: Shortage Until 2030 and Beyond
The most bearish projections imply that memory shortages may stay in place for years.
According to Citi’s estimates, memory supply crunch on a global scale may last until 2031. The logic behind: AI is transitioning from one-time training to always-on learning, resulting in an ever-lasting need for memory. All three categories, HBM, server DRAM, and enterprise SSDs, will grow at the same time, and the supply expansion is limited due to the capacity allocation cycle and construction schedule .
Citi’s estimates:
- HBM bit demand: 62% increase in 2027, 69% in 2028
- HBM supply shortage: ~21% in 2027, ~36% in 2028
- Server DRAM demand: over 50% increase in 2027
- Global DRAM supply shortage: ~8.7% in 2027, ~9.7% in 2028
In August 2026, Sumit Sadana of Micron said that 2027 is expected to be “even tighter than 2026” and the company didn’t have visibility yet when the supply will meet the demand .
The Bear Case: AI Bubble Concerns
Not everyone believes this process will go on forever.
The potential existence of an AI bubble is indeed possible. Should there be any reduction in AI investments because of economic or other factors, or lack of satisfaction in AI capabilities, the demand for memory can drop drastically. Memory industry has always been extremely cyclical, and there might even be an AI crisis.
As Mike Howard of TechInsights explained: “There is indeed a lot of concern about an AI bust and it is warranted.” Nevertheless, Mike Howard and others believe that the new dynamics of the memory industry, namely long-term contracts, customer-funded capacity expansion, and controlled capacity expansion reduce the chances of an oversupply crash.
The companies producing memory have already learned their lessons in the past cycles. Instead of building too much during booms and paying for busts, they have started acting more responsibly. Long-term contracts now constitute a large part of production.
The Physical AI Wildcard
Physical AI is one area that can prolong the boom cycle.
Robotic systems and humanoid robots could be a huge new demand generator. According to Sadana of Micron, “A humanoid robot will need hundreds of gigabytes of DRAM and terabytes of solid state storage just for onboard computing and safety when there is no cloud connection”.
Sadana anticipates the robotics market to “take off towards the latter part of this decade and possibly into rapid growth in the early part of next decade”.
If all goes according to plan, demand from robotics alone could eclipse today’s data center build-out.
What It Means to Diverse Audiences
For Investors
The memory supercycle has resulted in enormous value creation at Samsung, SK Hynix, and Micron. Now the real question is whether the valuations currently reflect the ongoing boom.
The indicators to keep an eye on:
- DRAM contract prices: the first signal about memory margins
- HBM supply contracts: whose design wins will the AI accelerators have
- Capex announcements: how eagerly the manufacturers are investing in capacity expansion (moderate capex is bullish, aggressive capex is bearish)
- AI infrastructure investments: the main source of demand
Tech Professionals
There are several practical implications from the memory shortage:
- Cloud prices for memory-heavy operations will stay high
- Hardware purchases will take additional planning and time
- System design should perhaps be optimized for memory efficiency rather than maximum speed
- AI edge implementations could be limited by memory shortages and prices
Consumers
The indirect impacts can be seen now:
- Memory shortage is driving smartphone prices up
- Next, laptop and PC prices will rise
- Budget computers such as Raspberry Pi could suffer especially
It is not going to go away any time soon. Memory shortages and price increases will last until at least 2027.
The Broader Context: What This Earnings Report Really Means
The earnings report from Samsung is more than just a financial news story. It is telling us something about the future of computing. Memory has now become the bottleneck on AI deployments. Not GPU capability. power consumption. Not physical space. Memory. The silicon that holds and transfers data is now the critical resource for the most important technology transition of the decade.
For decades, memory was a commodity: inexpensive, plentiful, mundane. Those companies that supplied memory were subject to savage cycles of boom and bust. Profits were episodic rather than structural.
This has all changed. AI requirements have now turned memory into a strategic resource. Companies that manufacture the memory required by AI are the ones controlling its rollout. Companies that cannot secure the memory cannot deploy AI. Samsung’s record earnings are the expression of this reality. The company has been preparing for this moment for years.
Now, the question is how long it will last. In the bull scenario, the memory market should do well for years, even until 2031. In the bear scenario, the AI bubble will burst and memory prices will fall. However, most likely the truth is somewhere between the two extremes, but currently, memory companies are making lots of money.
Actionable Takeaways
- Know your memory hierarchy. Not all memory is equal. HBM, DDR5, LPDDR5, and NAND each play a role and have distinct supply characteristics. The artificial intelligence revolution is all about HBM and server DRAM.
- Focus on the supply/demand imbalance, not price. Price is a lagging indicator. The supply/demand imbalance the extent of demand over available supply is what tells you where prices are headed. Recent projections suggest imbalances lasting out to 2028 and beyond.
- Look for higher pricing in consumer electronics. Memory is everywhere. When buying new technology equipment, don’t expect falling prices anytime soon.
- Market competition is evolving. HBM4 mass manufacturing from Samsung and strong growth in conventional DRAM from Micron is changing the market share landscape. HBM4 dominance will be powerful in the age of AI.
- Long term contracts count. Memory suppliers are now signing long term contracts of up to five years with their customers from the previous three year period. This gives an indication of the demand levels and removes cyclicality.
Conclusion
A record Q1 with sales of $80 billion is clear evidence that memory chips and not graphics processors are now the true constraining factors in the era of AI. Having invested silently in HBM technology and advanced DRAM, Samsung became prepared for the huge wave of spending on AI infrastructure development, and the results showed themselves: four consecutive record-breaking quarters, year-over-year increase by 782%, and the first trillion-worth of 100 won by any Korean corporation.
This process is defined by certain trends, not cycles. The production of HBM consumes 3-4 times less DRAM than its classical analogs, there is an increasing need for memory bandwidth in AI inference 5-30 times higher than in conventional usage, and the supply cannot catch up with the rising demand growing at the rate of about 200% versus only 20% capacity growth yearly.
Disclaimer: The above information is provided for general information and educational purposes only and does not constitute financial advice. All the technical specifications and financial numbers used in the article have been taken from publicly available reports and company statements.
Explore Our AI Category. And if you are reading it up to here, leave a sweet comment to motivate us to write blog everyday.
