Elon Musk says memory could be AI’s next big bottleneck. Tesla and SpaceX CEO Elon Musk recently agreed with the view that memory, rather than computing power, could become the main limit for the next phase of artificial intelligence. Musk responded to technology executive Peter H. Diamandis on X on August 14, saying, “Few realize this.”
The comment is important for memory and storage stocks. Companies such as SK Hynix, Micron and Sandisk have benefited heavily from the rapid growth of AI. Memory was once seen as a basic commodity business, but the rise of agentic AI is turning memory and storage into a key part of AI infrastructure, according to The Motley Fool.
Memory and storage stocks have already seen huge gains. The strong demand for AI-related memory has pushed shares of Micron, SK Hynix and Sandisk sharply higher as investors bet that AI companies will need much more memory and storage. But the rally faced a major setback in July. Memory stocks suffered a big pullback as investors took profits and became worried about several risks. These included the possibility of more efficient AI models from China, doubts raised by short sellers and the collapse of AI-focused hedge fund Situational Awareness, according to The Motley Fool.
The stocks have recovered partly in August, but remain below their June highs. Even after the recent bounce, Micron, SK Hynix and Sandisk were still around 15% to 30% below their June peaks, according to The Motley Fool. Musk’s comment could support the bullish case for memory stocks. His view suggests that the recent drop may not necessarily mean the AI boom is ending. Instead, the growing need for memory could keep demand strong for years.
Also read: L3Harris names Sam Mehta CEO as Christopher Kubasik steps down after board investigation
Why memory is becoming more important
The AI market is changing from simple chatbots to AI agents. In the early generative AI boom, users mostly gave AI a question or prompt and waited for an answer. These workloads depended heavily on powerful GPUs that could process huge amounts of information at the same time.
Agentic AI works differently. AI agents are designed to plan, think, retrieve information and complete tasks on their own. They can break a job into several steps, use different tools, check their work and try again when needed. That means AI agents need much more memory. Instead of simply answering one question, an AI agent may have to remember what it has already done, store information from different tools and keep track of a task over a long period.
Micron recently explained the different types of memory needed by an AI agent. According to Micron, each agent instance needs memory for several important tasks. First is state and KV/context staging. This helps an AI agent keep track of where it is in its reasoning process and what it has already done, according to Micron. Second is tool outputs and queues. AI agents can use APIs, software tools and code execution. The results from these activities have to be stored and managed.
Third is container and sandbox memory. AI agents often work inside isolated environments. These environments need memory so that tasks can run safely. Fourth is vector and index data. This helps AI systems find relevant information through retrieval-augmented generation and semantic search, according to Micron. Fifth is the operating system and runtime overhead. Running thousands of AI environments at the same time also requires memory just to keep those systems operating.
Also read: IREN stock jumps after Microsoft approves first AI data center site
SK Hynix and Micron could benefit from advanced memory demand
The memory needed for agentic AI is not just traditional DRAM. Micron said many of these workloads require specialized, high-capacity and high-bandwidth DRAM. High-bandwidth memory is more expensive and difficult to produce. Memory makers have said that producing high-bandwidth memory can require at least three times more capital equipment per bit than traditional server DRAM.
This creates a supply problem. AI companies are demanding more advanced memory at the same time that manufacturers need more equipment and investment to produce it. That is one reason DRAM prices have surged. Strong demand combined with limits on advanced-memory supply has helped push memory prices higher. SK Hynix and Micron are directly exposed to this trend. Both companies make DRAM and high-bandwidth memory products that are increasingly important for AI systems.
Sandisk could benefit from the storage side
AI’s memory needs are not limited to DRAM. AI systems also need large amounts of storage to hold information that does not have to remain in the fastest memory at all times. NAND flash is important here. NAND is slower than DRAM, but it can store information even when a device or system is turned off.
Long-context AI agents could create strong NAND demand. AI agents can build very large KV-cache chains containing information from previous interactions and tasks. Some of this information can be moved to NAND-based SSDs. This allows AI systems to store large amounts of context without keeping everything in the fastest and most expensive memory.
This gives Sandisk a different opportunity from SK Hynix and Micron. SK Hynix and Micron have exposure to both DRAM and NAND, while Sandisk is mainly focused on NAND storage.
Why inference could be bigger than training for memory demand
AI training already requires huge amounts of memory. Training large AI models needs many GPUs and large amounts of memory to process the data required to build the models. But training has a limit. Once a model is trained, the amount of memory needed for that particular training process is tied to the size of the model and the hardware being used.
Inference could create a much bigger long-term memory demand. Inference happens when people and businesses actually use AI models. If AI agents become part of daily work, they could continuously read and write information. AI agents may need to keep their memory for long periods. An agent working on a complicated task could constantly update its KV cache as it receives new information, performs actions and checks results.
That means memory demand could keep growing with AI usage. The more people and businesses use AI agents, and the longer those agents work on tasks, the more memory and storage they could require. Goldman Sachs has made a huge forecast for future AI usage. The investment bank estimates that agentic AI could consume around 120 quadrillion tokens per month by 2030.
That would be about 24 times the token usage seen in early 2026, according to Goldman Sachs. The estimate shows how quickly AI usage could expand if agents become a regular part of work and consumer technology. This could create a long-lasting opportunity for memory makers. If AI usage continues to grow this quickly, demand for DRAM, high-bandwidth memory and NAND storage could remain strong even as manufacturers add more supply.
AI memory demand could keep stocks rising
More supply may not automatically end the memory boom. The Motley Fool said new memory capacity is expected to come online around 2028, but demand from AI could grow fast enough to absorb that additional supply.
For SK Hynix, the key opportunity is advanced memory. The company is positioned to benefit from rising demand for high-bandwidth and high-capacity memory used in AI infrastructure. For Micron, the opportunity is also broad. Micron has exposure to DRAM, high-bandwidth memory and NAND, giving it exposure to several parts of the growing AI memory market. For Sandisk, the main opportunity is AI storage. Its focus on NAND means the company could benefit if AI agents increasingly move large amounts of context and data into SSD storage.
The bigger message for investors is that AI may need more than just GPUs. The AI boom has largely focused on computing power and companies making AI chips. Musk’s comment highlights another possible bottleneck: the ability to store, access and manage the huge amount of information AI agents need. This could make memory a strategic part of the AI race. As AI systems become more independent and handle longer and more complicated tasks, memory and storage could become just as important as computing power.
The recent memory-stock pullback may therefore not mean the AI story is over. If agentic AI adoption grows as expected, the demand for advanced memory and storage could remain strong for years, potentially extending the current memory upcycle beyond what many investors expect, according to The Motley Fool.