Samsung and SK Hynix profits surpass NVIDIA amid surge in AI memory demand

Organized & Compiled by Deep潮 TechFlow

Guest: EJ, Co-Host of the Limitless Podcast

Host: Josh, Limitless Podcast

Podcast source: Limitless Podcast (formerly the Bankless channel)

The AI Trade Everyone Is Getting Wrong

Broadcast date: July 9, 2026

Samsung's latest quarterly profit reached $58.5 billion, surpassing NVIDIA's $53 billion during the same period, with over 94% of that profit coming from a single division: AI memory. Only three companies worldwide can produce HBM (High Bandwidth Memory): two in South Korea (Samsung and SK Hynix) and one in the United States (Micron). Each AI inference requires re-reading the entire model weights, resulting in memory demands 10 to 20 times higher than those of consumer products, with requirements doubling with each new model generation. Meanwhile, 1 GB of HBM consumes the equivalent production capacity of 4 GB of standard DRAM, directly displacing memory supply for smartphones and computers—a direct consequence of which is price increases across Apple’s entire product lineup.

The contradiction lies in the fact that memory companies are reporting record profits and prices are still rising, yet their stock prices have collectively entered a bear market. Samsung fell 9% on its earnings day, and SK Hynix dropped 15%. The two hosts believe this is a combination of "buy the rumor, sell the news" and cyclical panic, with no change in fundamentals. SK Hynix listed on Nasdaq in the form of ADRs, raising approximately $30 billion with fourfold oversubscription, with Leopold Aschenbrenner participating as a seed investor.

The profit margin of memory monopolies crushes everything.

Samsung's gross margin is 52%, and SK Hynix's is 72%. For every $100 of memory sold, $72 goes directly to the balance sheet. In comparison, Apple's hardware gross margin is approximately 30%.

Samsung's memory division employees received year-end bonuses equal to six times their annual salary. Sales in South Korea's luxury market have tripled over the past four months.

AI is a black hole for memory.

AI is essentially a memory black hole, far more extreme than any consumer or enterprise technology wave we've seen.

Every time you submit a prompt, whether it's ChatGPT or Claude, it must reload the entire model weights—every single time.

Each new model requires 10 to 20 times more memory than the previous generation. For models with 15 trillion or 20 trillion parameters, this demand will only continue to rise.

The price surge isn't over yet

Q1 increased by 90%, Q2 increased by another 50 to 60%, and Q3 is expected to rise another 20%. Samsung earned more in one year than it did in the previous 40 years combined.

Profits hit a new high, but the stock price entered a bear market.

Samsung's earnings day beat expectations, with quarterly profits surpassing NVIDIA's—yet its stock fell 9%. SK Hynix dropped 15%.

Meta hints at cutting AI capital expenditures. To be honest, they haven't produced anything truly useful.

If you bought them a month ago, it might seem silly now. But six to 24 months from now, these companies will still be strong businesses.

Samsung: The Misunderstood King of AI Profits

Josh: Samsung has just become the world’s most profitable company, yet almost no one knows what they actually make their money from. I used to think of Samsung as smartphones and computers, but in reality, 96% of their profits come from one division: memory. The memory market has only three major suppliers, and two of them happen to be in South Korea. How is it that the two most profitable companies in the world are both in the same country? EJ, how did Samsung achieve this?

EJ: When I was a child, I only knew Samsung as a phone maker. Over the past year and a half, I’ve constantly heard this name tied to AI, and I’ve been curious about what the company is actually doing. Let’s look at a few numbers. Samsung’s Q2 profit was $58.5 billion, surpassing analyst expectations of $55 billion by $3.5 billion. More importantly, it outperformed NVIDIA, which posted just $53 billion during the same period. A year ago, Samsung’s Q2 2025 profit was only $3.4 billion. In just one year, it jumped from $3.4 billion to $58.5 billion—what happened in between?

The answer is HBM—High Bandwidth Memory. Samsung is currently the world’s second-largest HBM supplier. Why is HBM so valuable? Because every GPU sold by NVIDIA, every TPU manufactured by Google, and any AI chip requires massive amounts of memory. Pay attention to the word “massive”—the memory requirements for consumer devices (laptops, computers) are predictable, but AI memory demand is exponential.

Samsung earns $650 million per day, $27 million per hour, and $7,500 per second—more than the world’s most valuable company. Yet Samsung isn’t the world’s most valuable company, which is intriguing.

The Three Memory Giants and HBM Monopoly

Josh: A computer needs two things: a workspace (memory) and a filing cabinet (storage). Only three companies worldwide can manufacture these.

EJ: Yes. Three types of memory, each with its own purpose. DRAM is the DDR5 in your computer—the workspace, temporary RAM that clears when powered off. NAND is the flash storage in your SSD or iPhone—the filing cabinet, slower but cheaper. Then there’s HBM, high-bandwidth memory—that’s the foundation of everything in the AI era.

The manufacturing of HBM is extremely complex. DRAM chips are stacked 12 to 16 layers high using highly precise processes. SK Hynix holds 60% of the entire HBM market. Any AI chip produced is likely to use their memory. A monopoly among three companies has taken shape, driving profit margins sharply higher.

Josh: Here’s a straightforward number: 1 GB of HBM consumes the wafer capacity equivalent to 4 GB of standard DRAM. This means that every wafer shifted to AI memory results in the loss of four times the memory capacity for phones and computers. Apple has raised prices across its entire lineup for the first time: the MacBook Air increased from $1,100 to $1,300, the MacBook Pro from $1,700 to $2,000, and the Mac Studio from $4,000 to $5,300. This is because buyers of the Mac Studio need massive memory for local inference—and memory simply isn’t sufficient.

AI is a black hole for memory.

EJ: If I had to explain in one sentence why everyone is so bullish on these companies: AI is essentially a memory black hole, more extreme than any technology wave we've ever seen.

Every time you submit a prompt to ChatGPT, Claude, or similar models, the system must reload the entire model weights. Model weights are the parameters that companies have spent billions of dollars to train. Each inference requires reloading these weights from scratch. Moreover, each new model demands 10 to 20 times more memory than the previous generation—and as models grow to 15 trillion or 20 trillion parameters, these demands will only continue to rise.

But that’s still not the largest part. Your chatbot remembers what you said last time and retains your information across conversations—these temporary storages rely on NAND flash memory, which also demands enormous capacity. So all three types of memory demand are being pushed to their limits simultaneously.

Some say this is a bubble that will burst. Historically, this makes sense—the memory industry has always been cyclical. Three years ago, SK Hynix nearly got acquired by Micron, at the absolute bottom of the cycle. They invested in HBM, unsure if it would sell, and Micron was close to buying them. In the end, they didn’t sell; instead, they doubled down on HBM—and now they’re South Korea’s most valuable company.

Price Surge: A Stepped Increase from 90% to 20%

EJ: Let’s look at the pricing. Over the past six months, SK Hynix and Samsung have increased prices: Q1 rose by 90%. The memory in your everyday phone or computer has suddenly doubled in price because just one component increased by nearly 100%. Q2 saw another 50 to 60% increase. Samsung is set to raise prices another 20% in Q3.

Samsung earned more in one year than it did in the previous 40 years combined—19 times more than the same period last year.

In terms of gross margin, a grocery store earns $3 on every $100 of sales, an automaker earns $7, and Apple’s hardware earns $30. Samsung earns $52, and SK Hynix earns $72—$72 directly from every $100 of memory sold. Samsung’s memory employees receive year-end bonuses six times their annual salary. South Korea’s luxury market has tripled over the past four months. In Taiwan, some individuals are taking out high-interest bank loans of $60,000 to buy TSMC stock. Something wild is happening in Asia’s AI market.

Josh: This surge is too extreme. Consumers are already feeling the impact. A 32GB RAM stick is now two to three times more expensive than last year. For building a PC, one-third of the cost is just memory. These price increases have already reached the real consumer market. The question is: How long can this last? Can they keep raising prices indefinitely?

EJ: It depends on one variable: whether the number of people using AI will continue to grow. This is the only proxy metric for measuring whether memory demand will keep increasing. If you believe that in the future everyone will run multiple AI agents for both work and life, then memory demand will grow exponentially.

What about supply? New fabs won’t come online until 2030. These facilities are designed with extreme precision and cannot rapidly flood the market. Demand is outpacing supply by 3 to 5 times, creating a supply constraint for years to come. In China, CXMT is producing similar DRAM and HBM, but all of their capacity is absorbed by domestic AI labs. Apple has tried to find alternative suppliers in China, but there’s simply no inventory available.

Another perspective is: What if a new model architecture emerges that requires less memory? I think the logic is precisely the opposite. As memory becomes cheaper, more applications can run, more AI agents can be deployed, and overall economic output increases—leading to even greater demand for memory.

Contradiction: Profits hit a record high, but stock price enters a bear market.

Josh: We’ve always been bullish on memory—since Micron’s recommendation at the end of last year, it’s risen 150%. But recently, all memory stocks have dropped more than 20% from their highs—a technical bear market. Samsung beat earnings expectations on its earnings day, posting quarterly profits exceeding NVIDIA’s, yet its stock fell 9%. SK Hynix dropped 15%. Profits are hitting new highs, prices are still climbing, but the market is saying, “Slow down.”

What might be unsettling the market is Meta’s hint at cutting AI capital expenditures. But to be honest, they haven’t produced anything particularly useful yet. This is the strange moment we’re in: companies say, “We’re fine—profits are at record highs, demand is strong,” while the market responds, “Wait, prices have risen too much; there are unknown risks.” What do you think?

EJ: My judgment is simple: this is "sell the news." The quarterly earnings season has just ended, and global funds hold heavy positions in these stocks, waiting for a better price to re-enter. You can think I'm just being stubborn if you like. But if you're a long-term AI investor, memory is essential, and only these three companies can produce it—this dynamic won't change anytime soon.

People are concerned about the cycle. During the last memory supercycle from 2017 to 2018, Micron’s P/E ratio reached 4 to 5 times, then dropped 60%, even though profits continued to rise. History is indeed repeating itself, but this time is different: last time, demand was driven by smartphones, and you could predict the demand ceiling. This time, it’s driven by AI, and there’s no visible demand ceiling.

There’s another interesting historical detail: during the last cycle low, when memory prices were at their cheapest, who was aggressively driving down prices and stockpiling memory? Apple. At the time, Apple held pricing power over Samsung and SK Hynix, forcing them to supply DRAM and HBM at the lowest possible rates. Now the tables have turned—Samsung and SK Hynix are doing exactly what Apple once did. In an efficient market, this is perfectly normal.

SK Hynix lists on Nasdaq

Josh: Next up, a major test. SK Hynix is a Korean company, but it listed on Nasdaq in ADR form on July 10, raising approximately $30 billion. This is a pivotal moment for the memory industry—how the U.S. market prices SK Hynix will determine the next phase of this memory trade. EJ, will you participate?

EJ: Short answer: Yes. I’m bullish on memory. I hold Micron and a DRAM ETF (a basket of memory companies). If you’re a U.S. citizen, buying South Korean stocks directly can be inconvenient, but you can gain exposure through these basket products. I’ve been waiting for SK Hynix to list in the U.S.—I’ll buy it when it does.

This IPO is currently reported to be four times oversubscribed, with institutions, pension funds, and retail investors all competing for shares. Securing this level of demand indicates that institutions have conducted in-depth research on the memory sector and intend to hold long-term positions.

Josh: Guess who contributed $2 to $3 billion of that $30 billion? Leopold Aschenbrenner. He’s back. Last time we analyzed his positions, some questioned his logic for shorting NVIDIA. Since we recorded that episode, NVIDIA has dropped 20%. Now he’s participating as a seed investor in SK Hynix’s IPO. You can question his judgment, but he’s never been wrong.

Long-term perspective: Super cycle or bubble?

Josh: Let’s align on the conclusions. In the long term, memory demand has no ceiling, and only three companies can manufacture it—new entrants won’t pose a meaningful competitive threat in the short term. In the short term, people are fearful due to the shadow of historical cycles and the rapid, steep price increases. That makes complete sense—you might feel foolish buying a month ago. But six to 24 months from now, these companies’ fundamentals won’t have changed.

EJ: Yes, there will eventually be a cycle peak, and that will happen when wafer capacity begins to exceed demand. But supply won’t catch up until 2030, while demand is three to five times higher than supply. Until then, margins and profit rates will continue to expand.

I’m more curious about what the audience thinks. AI memory isn’t the most exciting topic, but it’s becoming the most critical component in the AI infrastructure capital expenditure race. It’s the source driving up prices across all consumer goods. Do you think we’re being too optimistic? Can profit margins continue to rise, or is this simply a bubble that will eventually burst, and are Josh and I marking the top?