Elon Musk says the biggest obstacle to scaling artificial intelligence is no longer electricity or graphics processors, but memory chips, a claim that has put renewed investor focus on Micron Technology, SK Hynix and Sandisk.
On SpaceX's second-quarter 2026 earnings call held August 4, 2026, Musk was asked about the pace of the company's AI compute buildout and answered with five words that quickly circulated among memory-industry investors: "Limiting factor currently is memory." According to Yahoo Finance, Musk framed the issue in his own supply-and-demand terms, saying memory output is increasing by roughly 20% a year while demand for it is rising by around 200% a year, or possibly more. He added that under basic economics, when demand outpaces supply that dramatically, prices rise rather than fall. Yahoo Finance noted these figures were Musk's own characterization on the call and not independently verified industry data.
Musk's comment is notable partly for what he did not blame. He told the SpaceX call that the company's "tentative target is to actually have 20 gigawatts at the power and cooling level online by the end of next year," while conceding "I don't think we're going to achieve 20 gigawatts" and predicting a figure closer to 15 gigawatts. He said SpaceX deliberately builds "far more power, cooling, and electrical equipment than we have GPUs," calling that the logical approach given the relative cost of GPUs versus the rest of the system. On graphics chips, he said SpaceX's understanding is that it will receive "a very small percentage" of Nvidia's GPU output next year. In other words, Musk described power as abundant and GPU supply as constrained but not the top constraint — memory was.
The remark echoed comments Musk made roughly two weeks earlier. On Tesla's second-quarter 2026 earnings call around July 23, 2026, Musk thanked Micron by name twice for giving Tesla what he called a "significant" memory allocation "on reasonable terms," and described current memory pricing as "the biggest price jump in anything I've ever seen," according to Yahoo Finance. Micron shares rose 3.1% intraday that day even as Tesla stock fell 14.3% following an earnings miss, underscoring that investors treated the memory comments as a distinct signal from Tesla's own results.
Musk returned to the theme again on August 14, 2026, when he replied to a post on X from technology executive Peter Diamandis, who wrote that "memory, not compute, is the rate limiter of the Agentic Era." Musk's response — "Few realize this" — was highlighted by The Motley Fool as a bullish signal for Micron, SK Hynix and Sandisk, the three companies most central to advanced memory and storage production.
Why Agentic AI Consumes So Much Memory
The Motley Fool explained that the shift from simple prompt-and-answer generative AI to "agentic" AI, in which systems plan and execute multistep tasks independently, has changed which components matter most. Citing a Micron blog post, the outlet reported that every AI agent instance requires memory for tracking its reasoning state, buffering tool outputs, maintaining isolated sandbox environments, storing vector data for retrieval, and covering general runtime overhead. Much of this workload relies not on standard commodity DRAM but on specialized, high-bandwidth memory that is more expensive to produce. Memory makers have noted that high-bandwidth memory requires at least three times as much capital equipment per bit as traditional server DRAM, according to The Motley Fool, which helps explain why supply has struggled to keep pace with demand.
NAND flash storage, used to retain data when systems power down, is also seeing rapid demand growth because long-running AI agents must offload large amounts of context, or "KV-cache" data, to NAND-based solid-state drives. Micron and SK Hynix produce both DRAM and NAND, while Sandisk operates as a NAND specialist. The Motley Fool cited Goldman Sachs estimates projecting that agentic AI could consume roughly 120 quadrillion tokens per month by 2030 — 24 times the token usage seen in early 2026 — as a reason the memory upcycle could persist even as new supply comes online in 2028.
The scale of the imbalance has been described in stark terms elsewhere in the market. Yahoo Finance reported that nearly 100 gigawatts of new AI data center capacity are expected globally within four years, compared with only about 15 gigawatts of new DRAM supply capacity over the next two years. DRAM contract prices have been projected to rise another 90% to 95% in early 2026, building on already sharp increases, and all three major HBM producers — Micron, Samsung and SK Hynix — are reportedly sold out of 2026 HBM capacity, with significant new supply not expected until 2028 or later.
Micron's stock has reflected that tightening market. Yahoo Finance reported the shares closed at $971.66 on August 14, 2026, up 240.65% year to date and up 676.79% over the prior twelve months, trading at around 6 times forward earnings. Separately, Bank of America projected that AI-driven demand could push Micron's fiscal 2030 earnings per share to $236, far above consensus estimates, according to Pluang, driven largely by growth in high-bandwidth memory. Micron chief executive Sanjay Mehrotra told investors the company expects tight supply conditions to persist beyond calendar 2027, a view Yahoo Finance said is backed by Strategic Customer Agreements disclosed in Micron's most recent quarterly filing.
Not every analyst is convinced the boom will continue at its recent pace. A Motley Fool analysis cited by Yahoo Finance pointed to TrendForce forecasts for the third quarter of 2026 showing DRAM contract price growth decelerating to just 13% to 18%, and NAND growth slowing to 10% to 15%, down sharply from increases of more than 60% and 80% in the prior quarter. Wall Street's fiscal 2027 earnings estimates for Micron have reportedly plateaued as well. Memory has historically been a cyclical, boom-and-bust business, and some analysts caution that today's low-looking valuation multiples could prove misleading if price growth cools further and margins compress.
For now, Musk's public framing — repeated across two earnings calls and a social media exchange within a three-week span — has reinforced the view among many investors that memory, not power or GPU availability, is the binding constraint on how fast AI systems from companies like SpaceX and Tesla can scale.
