Samsung and SK Hynix are making more money from AI memory, but the market has started marking their shares lower. In a July 9 episode of Limitless Podcast, host Josh and guest EJ said the drop looks more like a mix of "sell the news" trading and fear of a cycle peak than a break in the underlying demand story.
Samsung’s profit surge was tied to AI memory
EJ said Samsung posted $58.5 billion in profit for the second quarter of 2026, beating an analyst estimate of $55 billion and topping NVIDIA’s $53 billion in the same period. He contrasted that with Samsung’s $3.4 billion profit in Q2 2025, framing the jump as the result of one product category: HBM, or high-bandwidth memory.
The podcast summary said more than 94% of Samsung’s profit came from AI memory. In the dialogue itself, Josh used a slightly different figure and said 96% of the company’s profit came from one division. EJ also described Samsung as the world’s second-largest HBM supplier.
To show the scale of the jump, EJ said Samsung is making $650 million a day, about $27 million an hour, or roughly $7,500 a second.
Three companies dominate HBM supply
The episode said only three companies in the world can produce HBM: Samsung, SK Hynix, and Micron. Two are based in South Korea, one in the United States.
Josh used a simple analogy, calling memory the computer’s workbench and storage its filing cabinet. EJ then broke the market into three categories: DRAM for temporary working memory, NAND for flash storage in devices such as SSDs and iPhones, and HBM for AI chips.
According to EJ, HBM production is far more complex than standard DRAM manufacturing, with DRAM chips stacked 12 to 16 layers high. He said SK Hynix controls about 60% of the HBM market, putting it at the center of AI chip supply.
AI inference keeps pulling in more memory
EJ called AI a "memory black hole." His point was straightforward: every time a user sends a prompt to ChatGPT or Claude, the system has to reload the full model weights. He said each new generation of models demands 10 to 20 times more memory than the previous one, and cited 15 trillion-parameter and 20 trillion-parameter systems as examples of where that curve is heading.
He added that chatbot memory across sessions also drives flash demand, because temporary user context sits on the NAND side. In the hosts’ telling, AI is not only lifting HBM demand. It is pushing on DRAM and NAND at the same time.
Josh gave another number to illustrate the supply squeeze: 1 GB of HBM consumes wafer capacity equal to 4 GB of standard DRAM. In other words, when production moves toward AI memory, the market loses a much larger amount of phone and PC memory supply.
He tied that directly to Apple product price increases, saying the MacBook Air moved 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.
Pricing and margins stayed strong
EJ said memory pricing climbed sharply over the past six months. He described a 90% increase in Q1, another 50% to 60% in Q2, and a further 20% increase expected from Samsung in Q3.
He also compared profitability across industries. In his example, a grocery store earns $3 on every $100 sold, automakers make $7, Apple hardware makes around $30, Samsung’s gross margin stands at 52%, and SK Hynix’s reaches 72%.
The podcast used those margins to show how unusual the current setup is. It also mentioned that Samsung memory staff received year-end bonuses worth six times annual salary, while South Korea’s luxury market tripled over the past four months. Josh added that 32GB memory modules now cost two to three times what they did a year ago, and memory can account for one-third of the cost of building a PC.
Supply was described as constrained for years
EJ said the demand side depends on one basic question: whether the number of AI users keeps growing. If people eventually run multiple AI agents for work and daily life, he said, memory demand should keep rising at an exponential rate.
On supply, he argued that new fabs will not come online until 2030. Because those facilities are highly complex, he said the market cannot quickly flood itself with new supply. His view was that demand is still growing three to five times faster than supply.
The podcast also said China’s CXMT is working on similar DRAM and HBM products, but that its output is being absorbed by domestic Chinese AI labs. EJ added that Apple has looked for alternative suppliers in China but still cannot secure enough product.
He pushed back on the argument that a new model architecture could reduce memory demand. His view was the opposite: cheaper memory would unlock more use cases, more deployed agents, and more total economic activity, which would still lift aggregate demand for memory.
Stocks fell even as profits rose
Josh said Micron had risen 150% from the time the show recommended it late last year, but memory shares have since dropped more than 20% from their highs, putting the group in technical bear-market territory.
He pointed to Samsung’s 9% drop on earnings day and SK Hynix’s 15% slide as the clearest examples of the disconnect. He also said Meta’s signal that it may rein in AI capital spending may have unsettled investors.
EJ’s conclusion was blunt: this was a sell-the-news event. He said funds were already heavily positioned in these stocks through earnings season and were waiting for a better re-entry point. From his perspective, the long-term setup remains the same because memory is still essential for AI and only three companies can supply it.
A cycle debate is back in focus
EJ acknowledged that investors are not imagining the cycle risk. He referred to the 2017-2018 memory supercycle, when Micron traded at 4 to 5 times earnings and still fell 60% even though profit was rising.
His argument was that this cycle differs from the last one because smartphones drove demand back then, and the ceiling was easier to estimate. This time, AI is the main driver, and the ceiling is much harder to see.
He also said SK Hynix nearly got bought by Micron three years ago, when the industry was at a much weaker point and HBM was still an uncertain bet. Instead, SK Hynix doubled down and later became South Korea’s most valuable listed company, according to the podcast discussion.
SK Hynix’s Nasdaq ADR listing is the next test
Josh said SK Hynix is set to list on Nasdaq on July 10 in ADR form, raising about $30 billion. He described that deal as an important test of how the U.S. market will value the next phase of the memory trade.
EJ said he plans to participate. He added that he already owns Micron and a DRAM ETF, and that U.S. investors often find direct access to Korean equities less convenient than buying basket products.
The episode said the IPO was about four times oversubscribed, with institutions, pension funds, and retail investors all taking part. Josh also said Leopold Aschenbrenner joined as a seed investor and speculated that $2 billion to $3 billion of the $30 billion raise could come from him.
The hosts’ bottom line
The show’s core view was that the recent pullback says more about positioning and cycle anxiety than about a collapse in memory fundamentals. Josh said buying a month ago may look painful right now, but over a six- to 24-month window he does not think the business case has changed.
EJ said the true top would come when wafer capacity turns excessive. In his view, that does not happen before new supply arrives around 2030. Until then, he said, margins and profitability can still keep expanding.