Elon Musk Just Uttered 3 Massively Bullish Words for Micron, Sandisk, and SK Hynix

It has been a historic period for memory and storage stocks. Once relegated to a bucket of lowly valued “commodity” stocks, the agentic AI revolution has spurred an absolutely massive increase in memory and storage demand.

Not only has demand increased, but the technology has also evolved from an interchangeable commodity to a strategic enabler for AI systems. The agentic era has therefore spurred massive stock price increases for memory and storage giants Micron (MU +2.30%), SK Hynix (SKHY +0.40%), and Sandisk (SNDK +7.40%).

However, those huge gains gave way to a big pullback in July because of profit-taking, fears of more efficient models from China, short-seller skepticism, and the “blow-up” of the AI-focused hedge fund Situational Awareness. Even after a bounce back in August, these stocks remain 15% to 30% below their June highs.

Is the recent pullback a harbinger of more pain and a “bubble bursting,” or an opportunity to buy the dip? Last week, Elon Musk wrote a three-word sentence on his social media that strongly points to the latter.

“Few realize this”

Last week on X, a private technology executive noted, “Memory, not compute, is the rate limiter of the Agentic Era,” to which Elon Musk replied, “Few realize this.”

Memory, not compute, is the rate limiter of the Agentic Era.

— Peter H. Diamandis, MD (@PeterDiamandis) August 14, 2026

Few realize this

— Elon Musk (@elonmusk) August 14, 2026

In the first wave of generative artificial intelligence, the environment was dominated by simple questions or prompts directed at the AI system, which would then find the answer. That relatively simple AI application places a lot of onus on the GPU and its massively parallel-processing capabilities.

However, in the agentic AI era, in which AI is tasked with planning and executing tasks independently, the game has changed. Now the focus has pivoted to planning, thinking, data retrieval, testing, and retesting agentic outputs. That has exponentially expanded CPU-heavy “planning” tasks.

Not only that, but each “task” also requires vast amounts of storage and memory. In a recent blog post from Micron, the company wrote that every single agent instance requires:

• State and KV/context staging — keeping track of where it is in its reasoning loop.
• Tool outputs and queues — buffering results from API calls and code execution.
• Container/sandbox memory — isolated runtime environments for safe execution.
• Vector/index data — for retrieval-augmented generation and semantic search.
• OS and runtime overhead — the base cost of keeping thousands of environments alive.

Each one of these requirements entails memory to support it. Micron also notes that much of the memory for these workloads isn’t traditional “commodity” DRAM, but specialized, high-capacity, high-bandwidth DRAM.

These smarter memory architectures require more capital equipment to produce; for instance, memory makers have noted that high-bandwidth memory requires at least three times as much capital equipment per bit to produce as traditional server DRAM.

That means the supply of advanced memory required for agentic AI is becoming harder to meet, just as demand is exploding, which is why DRAM prices have boomed.

Not to be outdone, NAND flash, which stores information even when a system is turned off, though it is slower than DRAM, is also growing fast. That’s because massive KV-cache memory chains for long-context agents — basically, the prior context AI agents must “remember” to produce more tokens — require a lot of data to be offloaded to NAND-based SSDs. Both Micron and SK Hynix produce NAND along with DRAM, while Sandisk is a NAND “pure play.”

Image source: The White House.

Inference has catapulted memory to the forefront

Training even the best frontier models requires a lot of memory, but that amount is ultimately capped at the amount needed to fill a GPU. In other words, GPUs can only read so much memory at once, so training a model requires a lot of GPUs, along with a requisite amount of memory.

However, as we enter an era in which more and more consumers and enterprises use agentic AI as a daily habit, the demand for memory appears almost endless. If an AI agent operates over a long period of time, it will have to constantly read and write the KV cache ad infinitum.

That’s why researchers at Goldman Sachs just released awe-inspiring estimates of future AI token usage. By 2030, the investment bank estimates that agentic AI will consume roughly 120 quadrillion tokens per month: 24 times the token usage of early 2026.

Thus, it’s no wonder that Elon Musk highlights memory as the biggest silicon-based constraint for AI moving forward. Even as more supply comes online in 2028, it appears the demand will be there to absorb it. Thus, the current memory up-cycle may last longer than many investors realize.