Author: Compiled by Techub News
In the latest episode of the Bankless program, hosts Ejaaz A. and Josh conducted an in-depth discussion on the currently red-hot AI memory investment track. As Nvidia becomes the world's most valuable company by market cap thanks to its GPUs, a similarly critical but potentially overlooked hardware component—memory—is undergoing a structural transformation driven by AI. Based on detailed industry data and historical cycle analysis, this conversation aims to answer a core question: Is the AI memory boom a fleeting bubble, or a historic investment opportunity with sustainable growth potential?
AI Reshapes Memory: From "Miserable Industry" to "Core Bottleneck"
Over the past four decades, the memory industry has been viewed as a classic "bad business." The three major suppliers—Samsung, SK Hynix, and Micron—produce nearly identical products, trapping the industry in violent "hog cycles": price increases spur capital expenditure and capacity expansion, followed by oversupply and price collapse, repeating endlessly.
However, the emergence of AI has fundamentally reversed this situation. The hosts pointed out that memory now accounts for 50% of the total material cost of building AI models, and as Nvidia evolves from Blackwell to future Rubin and Feynman architectures, each generation of GPU will require 2x, 3x, or even 4x more memory. AI has created an almost "insatiable" demand for memory.
More critically, AI has birthed a new memory structure: High Bandwidth Memory (HBM). Josh explained with a vivid analogy: traditional memory sticks are like single-story warehouses, while HBM is a "skyscraper warehouse" that stacks memory dies vertically in 8 to 12 layers (soon to reach 16 layers). This design allows it to be placed right next to the GPU, supplying model weights and data at extremely high speeds, directly determining the speed and scale of model operation.
The manufacturing process for HBM is extremely complex and precise, severely limiting the speed of capacity expansion. Currently, only the three aforementioned giants can mass-produce HBM, and to meet AI's frantic demand, they have even halted production of some consumer-grade memory, allocating all resources to AI memory. The supply-side rigid constraint is unprecedented.
Dual Engines of Demand: From Training to Inference, From Single-Query Conversations to Agent Networks
Beyond structural changes at the hardware level, the evolution of software paradigms is adding new, stronger engines for memory demand.
First, AI's focus is shifting from training to inference. When users query ChatGPT or Claude, that process is inference. An even more cutting-edge trend is the rise of AI agents. These agents can converse with each other, collaborate on thinking to produce superior outputs, which requires massive inference computation. Ejaaz, citing analyst Ben Thompson, noted that in a future economy powered by AI agents, we might need 10 to 50 times more memory than we do today.
Second, the agent paradigm itself consumes more memory. Josh explained that every query sent to a large language model generates a "KV cache" to store context. In an agent scenario, each running agent requires its own independent instance of a KV cache. If you run 20 agents simultaneously, you need 20 memory instances. This software-driven demand growth, combined with HBM supply constraints, has significantly extended the demand window.
Currently, the lead time for obtaining HBM is already booked through the end of 2027. Ejaaz revealed that some large AI clients have even proactively offered to purchase or fund production equipment for companies like Micron in exchange for capacity guarantees for 2027-2028. This "prepaid" lock-in of capacity is extremely rare in past memory cycles, indicating that demand is real and urgent.
Investment Targets: The Big Three and a "One-Stop" ETF Tool
Faced with clear demand and constrained supply, how can investors participate? The conversation focused on the three core players and emerging financial instruments.
- SK Hynix: Considered the current "king." It is the primary, even exclusive, supplier of HBM to Nvidia, with much of its capacity already locked in by Nvidia, similar to TSMC's position with Nvidia.
- Samsung Electronics: An important second-tier player. However, due to its diversified business, it is not a pure-play memory concept stock.
- Micron Technology: The primary choice for U.S. investors. As a U.S.-based memory giant, Micron is also one of the advanced HBM suppliers, and its stock price has risen over 8x in the past year.
For Western investors, directly investing in South Korean listed stocks presents certain barriers. Consequently, an ETF product named DRAM emerged, skyrocketing to over $60 billion in assets within six weeks, which Goldman Sachs called one of the fastest-growing ETFs in history. This ETF's core holdings are precisely the three companies mentioned above (totaling about 70%), providing investors with a convenient, one-basket tool for investing in AI memory leaders.
Josh mentioned that an ETF named RAM, a 2x leveraged long memory ETF, has also appeared in the market, offering an "upgraded" option for investors with higher risk tolerance. However, he personally prefers the non-leveraged DRAM ETF.
Historical Cycles and Current Divergence: "Is It Really Different This Time?"
Despite the strong logic, investors' biggest doubt remains: memory stocks have already risen significantly—is this a bubble? Will historical cycles repeat?
Historically, the memory industry experiences a boom-bust cycle roughly every 18 months, with rallies often followed by 40%-75% corrections. The current cycle began in Q4 2024, and we are now in the critical 18-month time window. Chart-wise, prices are at resistance levels.
Ejaaz acknowledged that in the short term, the market might see a pullback due to "buy the rumor, sell the news." But he remains optimistic for the long term, with core support coming from real orders where payments have been made in advance. These orders come from cash-rich tech giants (like the "Magnificent Seven"), not from unidentified companies, which greatly enhances the credibility of the demand.
From a valuation perspective, these memory giants have forward price-to-earnings (P/E) ratios between 5 and 12 times, far lower than Tesla (40-50x), and even lower than Microsoft or Amazon (15-20x). Relative to their earnings guidance, their stock prices don't appear overly inflated. In contrast, during past bubbles, P/E ratios often reached 30-50x.
Of course, risks remain. Skeptics argue that if more efficient AI models or GPU architectures requiring less memory emerge in the future, or if new production capacity comes online leading to oversupply, the trend could reverse. Furthermore, soaring memory prices are already trickling down to consumer electronics, causing price increases or delayed launches for PCs, game consoles, etc., which could ultimately suppress consumer-side demand.
However, the two hosts' conclusions lean towards optimism. They believe that even if consumer demand weakens, enterprise demand for AI models and agents will "only increase, never decrease." In the program's final rapid-fire Q&A session, both agreed that memory stock prices will be higher 12 months from now. The AI-driven revolution in memory demand may have only just begun in terms of depth and breadth, while supply-side catch-up will take years. This supply-demand imbalance in memory, triggered by AI, might truly be writing a new story different from the past four decades.
