Samsung’s profit jump puts AI memory at the center
TechFlowPost’s write-up of a July 9 Limitless Podcast episode framed Samsung as one of the biggest winners of the AI memory boom. On the show, hosts Josh and EJ said Samsung reported quarterly profit of $58.5 billion, above analyst expectations of $55 billion and ahead of NVIDIA’s $53 billion for the same period.
The discussion argued that Samsung is still widely viewed as a consumer electronics company even though its earnings mix has shifted sharply. In the podcast summary, more than 94% of profit was attributed to one division, AI memory. In the body of the conversation, Josh described that share as 96%. EJ also contrasted the latest result with Samsung’s Q2 2025 profit of $3.4 billion, saying the move from $3.4 billion to $58.5 billion in one year came down to HBM, or high-bandwidth memory.
The episode also put Samsung’s earnings pace into simpler terms: $650 million a day, $27 million an hour, and $7,500 a second.
Three companies dominate HBM supply
EJ said the market for advanced memory is concentrated in three companies: Samsung, SK Hynix, and Micron. Two are based in South Korea and one is based in the US. The show described Samsung as the world’s second-largest HBM supplier and said SK Hynix controls 60% of the HBM market.
Josh broke memory into three categories: DRAM, NAND, and HBM. DRAM was described as working memory for computers, NAND as flash storage used in SSDs and phones, and HBM as the most important memory product in the AI buildout. According to the episode, HBM manufacturing is unusually complex, with DRAM stacks reaching 12 to 16 layers, which has reinforced the existing oligopoly and widened margins.
AI inference keeps pulling more memory into the system
EJ called AI “a memory black hole.” His point was simple. Every time a user sends a prompt to ChatGPT or Claude, the system has to reload the full model weights. The podcast described those weights as the parameters created through multibillion-dollar training runs, and said the repeated reads during inference create much heavier memory demand than typical consumer or enterprise workloads.
The hosts said each new model generation requires 10 to 20 times more memory than the previous one, and referred to models with 15 trillion and 20 trillion parameters. They also said chatbot memory features and cross-session context add pressure on NAND demand, meaning DRAM, HBM, and NAND are all being pulled higher at the same time.
Josh added another supply-side figure: 1 GB of HBM uses wafer capacity equal to 4 GB of standard DRAM. In practice, that means more wafer allocation for AI memory leaves less output for phones and PCs.
Price increases are feeding through to end markets
The episode laid out a steep pricing sequence. Memory prices rose 90% in Q1, then another 50% to 60% in Q2, with Samsung set to raise prices by a further 20% in Q3, according to the show. EJ said Samsung is making more in one year than it did in the previous 40 years combined, and 19 times what it made in the same period a year earlier.
The margin comparison was one of the show’s clearest arguments. Grocery stores were said to make $3 for every $100 sold, carmakers $7, and Apple hardware about $30. Samsung’s gross margin was put at 52%, while SK Hynix was placed at 72%, meaning SK Hynix keeps $72 from every $100 of memory sold.
The hosts also pointed to knock-on effects outside chipmaking. Samsung memory staff were said to be receiving bonuses worth six times annual salary. South Korea’s luxury goods market was described as having tripled over the past four months. Josh also said 32GB memory sticks now cost two to three times more than a year ago, and that memory can account for one-third of the cost of building a PC.
The episode tied some of that inflation to Apple pricing, citing increases from $1,100 to $1,300 for the MacBook Air, from $1,700 to $2,000 for the MacBook Pro, and from $4,000 to $5,300 for the Mac Studio.
New supply, according to the hosts, will take years
On durability, EJ said the main variable is whether AI adoption keeps expanding. If more people end up using multiple AI agents across work and daily life, he said, memory demand would keep compounding.
On supply, the podcast said new fabs will not come online until 2030. EJ’s view was that demand is running three to five times faster than supply growth, which would keep the market constrained for years. The episode also mentioned China’s CXMT as a producer of similar DRAM and HBM products, but said its output is being absorbed by domestic AI labs and is not available as an alternative source for Apple.
EJ also rejected the idea that a future model architecture might sharply reduce memory needs. His argument was that lower memory costs would allow more AI use cases and more deployed agents, which in turn would lift total memory demand.
Record profits, falling stocks
The sharpest tension in the episode was the gap between company performance and market pricing. Josh said Micron had risen 150% since being recommended near the end of last year, but memory stocks have since fallen more than 20% from their highs, putting the group in technical bear-market territory.
Samsung fell 9% on the day it beat expectations and posted quarterly profit above NVIDIA’s. SK Hynix fell 15%. The hosts said part of the concern may have come from Meta signaling restraint on AI capital spending, but EJ’s core explanation was “sell the news” combined with renewed cycle fears.
He pointed to the 2017–2018 memory supercycle, when Micron’s price-to-earnings ratio fell to 4 to 5 times and the stock later dropped 60% even as profits were still rising. The difference this time, he argued, is that the previous cycle was driven by smartphones, where demand ceilings were easier to estimate. AI demand, in his view, does not yet have a visible cap.
SK Hynix’s Nasdaq ADR becomes a pricing test
The episode also focused on SK Hynix’s July 10 Nasdaq listing via ADR, which the hosts said raised about $30 billion. Josh described the deal as an important test for how US markets value the memory trade at this stage of the cycle.
EJ said he planned to buy. On the show, he said he already owns Micron and a DRAM ETF, and had been waiting for SK Hynix to trade in the US. The IPO was said to be four times oversubscribed, with institutions, pension funds, and retail investors all participating.
Josh also said Leopold Aschenbrenner took part in the offering as a seed investor, and estimated that $2 billion to $3 billion of the $30 billion raise could have come from him. The episode added that NVIDIA shares had fallen 20% since a previous discussion of his positions.
Podcast view: weak near-term tape, strong medium-term setup
The hosts ended by splitting the story into two time frames. In the near term, they said investors are dealing with valuation resets, fast prior gains, and familiar cycle anxiety. Over a 6- to 24-month horizon, EJ said the business fundamentals of the major memory makers still look intact.
Josh’s conclusion was that only three companies can supply this part of the AI stack at scale, and new entrants are unlikely to matter soon. EJ said the real top of the cycle would come when wafer capacity turns excessive. Based on the figures discussed in the episode, that is not expected before 2030, while demand is still running three to five times faster than supply.