Elon Musk has identified memory as the single biggest constraint on the expansion of artificial intelligence infrastructure, telling investors on SpaceX’s second-quarter earnings call that demand for memory chips is growing ten times faster than manufacturers can increase supply.
News Summary
- Elon Musk said on SpaceX’s Q2 2026 earnings call that memory is “the limiting factor” in AI infrastructure expansion.
- Musk stated memory supply is growing roughly 20% annually while demand is increasing by 200% or more per year.
- Micron shares have risen approximately 205% year-to-date; Sandisk has gained more than 420% year-to-date.
- Both stocks have retreated significantly from June highs amid concerns about the sustainability of AI spending.
- Micron expects new production capacity by mid-2027; Sandisk reported Q4 revenue up 51% sequentially, with two-thirds of growth driven by higher prices.
“The limiting factor currently is memory,” Musk said during the August 4 call. He noted that memory output is increasing by roughly 20 percent annually, while demand is rising by 200 percent per year or more. “So if you’ve got demand increasing much faster than supply, then economics 101 would suggest that the price increases. It does not decrease.”
The remarks from one of the world’s largest buyers of AI compute capacity have sharpened the focus on memory chip makers Micron and Sandisk, both of which have posted triple-digit stock gains this year on the back of soaring prices and insatiable demand from hyperscalers. Yet both stocks have also sold off sharply from their June peaks as investors debate whether the supply crunch can sustain its current intensity.
The Memory Bottleneck
SpaceX reported $7.8 billion in second-quarter revenue on the call, up 92 percent from a year earlier, with its AI segment generating $2.6 billion. The company spent approximately $15.8 billion on AI compute infrastructure during the quarter alone. Despite that massive outlay, Musk pointed to memory availability , not capital, power, or GPUs , as the chokepoint holding back further expansion.
The imbalance Musk described has already reshaped the memory market. High-bandwidth memory, the specialized DRAM used in AI accelerators, is fully sold out through much of 2026. Conventional DRAM contract prices surged 90 to 95 percent in early 2026, according to market researcher TrendForce. Apple Chief Executive Tim Cook warned earlier this year that memory price increases have become unavoidable even for the world’s largest technology buyer.
Musk also announced that SpaceX will build its AI infrastructure exclusively on Nvidia’s Vera Rubin architecture, calling it “the best AI computer.” The decision underscores how tightly coupled memory supply has become to the broader AI build-out: even as Nvidia and others rush to produce more GPUs, each accelerator requires multiple high-bandwidth memory modules that only a handful of companies can manufacture.
Micron’s Capacity Challenge
Micron, one of three major suppliers of high-bandwidth memory alongside SK Hynix and Samsung, has seen its shares rise approximately 205 percent this year. The company’s HBM supply is fully booked through 2026, with demand expected to exceed supply well into 2027.
But expanding production takes years. Micron has an advanced packaging facility coming online in Singapore this year that will strengthen its position in the AI supply chain. After that, new facilities in Boise, Idaho and Japan are expected to begin operations by mid-2027 and late 2028, respectively. The company has told investors to expect the tightness in the memory chip market to persist beyond 2027.
Micron’s stock has not been immune to the broader tech selloff. After touching a high above $1,200 in late June, shares have retreated to around $860 , a decline of roughly 28 percent. The pullback reflects investor anxiety that hyperscaler capital expenditure growth may slow, even as Micron’s own fundamentals remain constrained by how quickly it can bring new fabrication capacity online.
Sandisk’s Pricing Power
Sandisk, which specializes in NAND flash storage, has experienced an even more dramatic run. The company’s shares have gained more than 420 percent year-to-date, though they have fallen nearly 50 percent from a June intraday high of $2,354.
The company’s fiscal fourth quarter, which ended July 3, illustrated why investors had bid the stock up so aggressively. Sandisk reported revenue of $8.97 billion, up 51 percent sequentially and 372 percent from a year earlier. Non-GAAP earnings per share reached $39.25, well above analyst estimates.
Perhaps more telling than the headline numbers was the composition of growth. Management said approximately one-third of the sequential revenue increase came from higher volumes, while two-thirds came from higher prices. Data center revenue, the segment most directly tied to AI demand, jumped 103 percent from the prior quarter to $2.98 billion.
“AI is a memory centric, storage intensive problem,” Chief Executive David Goeckeler said on the earnings call. The company now has more than four years of demand visibility from customers and expects bits to remain on allocation beyond calendar year 2027.
The Supply-Demand Math
The numbers behind Musk’s observation are stark. Industry estimates suggest roughly 100 gigawatts of new AI data center capacity is expected to come online over the next four years. Yet available DRAM supply is sufficient to support only about 15 gigawatts of new capacity over the next two years.
That gap explains why memory makers have been able to raise prices so aggressively. It also explains why long-term supply agreements have become increasingly common. Sandisk has signed 10 new business-model agreements with customers, locking in demand at elevated price levels. Micron’s HBM capacity is similarly pre-sold.
Whether that pricing power persists depends on whether demand continues to outpace supply. Some analysts have warned that capacity expansions by SK Hynix and others could eventually lead to oversupply. But Musk’s commentary suggests that from the perspective of a major buyer, the shortage is not resolving anytime soon.
What Comes Next
For memory manufacturers, the immediate challenge is execution. New fabrication facilities require years to plan, build, and ramp. Micron’s Boise expansion and Sandisk’s BiCS8 and BiCS10 technology ramps are underway, but meaningful volume increases remain months or years away.
For AI infrastructure builders, the memory shortage means continued cost pressure. SpaceX alone spent $15.8 billion on AI infrastructure in a single quarter. If memory prices keep rising, those capital budgets will buy fewer GPUs and less total compute capacity than planners initially assumed.
The broader technology industry is watching closely. Memory has historically been a cyclical business prone to boom-and-bust price swings. The current cycle, however, is being driven by a structural shift in demand rather than a temporary spike. Whether that makes this cycle different from previous ones will become clearer as new capacity begins to come online in 2027 and 2028.