Micron CEO Warns AI Is Only in the 'First Innings' as Memory Supply Tightens, With DRAM and NAND Demand Set to Exceed 50% of Industry TAM

Micron Technology just reported record quarterly revenue of $13.64 billion, a 56 percent jump from the $8.71 billion it earned at the same time last year. By most measures, it should be a moment of unqualified celebration for the memory and storage giant. Instead, CEO Sanjay Mehrotra used the earnings call to deliver a sobering message: the global memory shortage is going to get worse before it gets better, and AI is the reason why.

"We are in very early innings," Mehrotra told analysts and reporters, referring to the AI industry's trajectory. "You just saw at GTC how much advances are being made in AI. And memory is a strategic asset. You need more memory, you need faster performance memory in order for AI to be able to deliver its full capabilities." The phrase "strategic asset" is deliberate. Memory has historically been treated as a commodity, a component that gets cheaper and more abundant over time. The AI era is upending that assumption.

The numbers tell the story. AI data centers require extraordinary amounts of memory in multiple forms. Training large language models demands high-bandwidth memory stacked directly onto GPUs. Running inference on those models at scale requires massive amounts of DRAM. Edge AI deployment in smartphones, laptops, and autonomous vehicles depends on LPDDR, the power-efficient memory variant that is now among the tightest supply segments. Micron says that DRAM and NAND demand from AI-related applications is on track to exceed 50 percent of the industry's total addressable market, a threshold that would have been unthinkable just two years ago.

The supply response has been substantial but insufficient. Micron aims to increase DRAM and NAND shipments by 20 percent next year. A new manufacturing facility in Idaho is scheduled to begin production in 2027, and a second plant in New York is planned for 2030. Even so, Mehrotra acknowledged that Micron can currently meet only half to two-thirds of customer demand. "Despite significant efforts, we are disappointed to be unable to meet demand from other customers across all market segments," he said during the call.

The decision to shutter its consumer-facing Crucial brand earlier this year crystallized the trade-off. Consumer RAM and SSD products, while profitable in aggregate, cannot compete with the margins available on high-bandwidth memory contracts from hyperscale cloud providers. When OpenAI, Meta, Microsoft, and Google are willing to pay premium prices for HBM with guaranteed volume, the economics of diverting wafer capacity to consumer products become difficult to justify. HBM uses roughly three times the silicon wafers compared to standard DRAM, which means every HBM chip produced directly reduces the supply of conventional memory available for PCs, smartphones, and other consumer devices.

The ripple effects are already visible. DDR5 RAM kit prices have climbed steadily over the past several months, and Tom's Hardware reports that price increases are expected to spread to smartphones and other devices. Micron's own earnings report warned that "memory supply constraints" could affect PC shipments next year, a remarkable statement from a company whose business depends on selling memory for those PCs.

The shortage also has strategic implications beyond consumer electronics. Memory is a concentrated market, with Samsung, SK Hynix, and Micron controlling the vast majority of global supply. All three are prioritizing AI-related products, which means that non-AI industries face a structural supply disadvantage. Automotive manufacturers, industrial IoT companies, and networking equipment builders all compete for the same pool of commodity DRAM and NAND that is being squeezed by AI demand. The result is a bifurcated market in which AI customers get priority allocation and everyone else waits.

The inference inflection that Mehrotra described is particularly significant. The AI industry is transitioning from a training-heavy phase, in which enormous but relatively contained clusters of GPUs build models, to an inference-heavy phase, in which those models are deployed at scale to serve millions of users simultaneously. Inference requires more total compute and more total memory than training because it happens continuously and must serve real-time responses. As agentic AI systems that can take autonomous actions become more prevalent, the memory requirements will scale further, since each agentic workflow requires its own memory context maintained in real time.

Upcoming hardware launches will intensify the pressure. Nvidia's Vera Rubin platform and AMD's MI400 accelerator both feature HBM4, the next generation of high-bandwidth memory that offers both higher bandwidth and greater capacity per stack. Demand for these platforms is already outstripping supply projections. Meanwhile, agentic AI workloads are pushing server CPUs to support up to 400 GB of memory, four times the current standard, further straining DRAM supply.

What This Means For You: If you are planning to build or upgrade a PC, buy RAM sooner rather than later. Prices are rising and availability is tightening as memory manufacturers prioritize lucrative AI data center contracts over consumer products. The same applies to smartphones and laptops, where manufacturers may absorb cost increases initially but will pass them on over time. For investors, the memory shortage represents both a risk and an opportunity. Companies like Micron, Samsung, and SK Hynix are positioned to benefit from sustained pricing power, but the concentration of supply creates fragility for downstream industries. For businesses building AI products, factor memory availability and cost into your deployment plans, as these constraints are structural and will persist through at least 2027. The era of abundant, cheap memory that defined the personal computing revolution is being replaced by an era in which memory is a strategic resource allocated by priority rather than price alone.