The bottleneck that strangled AI chip production has moved from packaging to memory, creating a structural shortage that memory-chip maker Micron says will last 'well beyond 2026.'
Back
Micron's $200B US Bet as AI Memory Demand Outstrips Supply by 2x
The bottleneck that strangled AI chip production has moved. For two years, the supply of advanced packaging from Taiwan Semiconductor Manufacturing Co. (TSMC) set the pace for the artificial intelligence buildout. That constraint is now easing, only to be replaced by a more fundamental one: high-bandwidth memory (HBM). This shift, identified by AMD CEO Lisa Su and confirmed by Micron Technology's latest market assessment, has created a structural shortage that is making investors rich while threatening to eliminate the sub-$500 PC from store shelves by 2028.
"We see this shortage continuing beyond, well beyond 2026 timeframe," Micron CEO Sanjay Mehrotra said in a Bloomberg interview on May 22, adding that the company can only meet "about 50% to about two thirds" of customer demand for HBM and DRAM chips. For an industry conditioned by boom-bust cycles, Mehrotra’s message reframes the investment case: this is a structural shortage with years of runway.
The supply-demand imbalance is already reshaping Micron’s financials. The company reported second-quarter revenue of $23.86 billion and EPS of $12.20, crushing guidance as its Cloud Memory Business Unit generated $5.284 billion in revenue at a 66 percent gross margin. The economics are driven by HBM, the specialized memory stacked vertically alongside AI accelerators from Nvidia and AMD. Producing a single gigabyte of HBM consumes roughly three times the silicon wafer capacity as standard DDR5 memory, and it commands margins exceeding 50 percent—a figure that has incentivized memory makers to reallocate production away from consumer-grade chips.
This reallocation is why the shortage is structural, not cyclical. With Micron’s entire 2026 HBM production already sold out, Mehrotra argues that "meaningful new supply in the industry doesn't really start ramping until 2028 timeframe." This gives Micron and its competitors, SK Hynix and Samsung, at least two more years of significant pricing power before new fabrication plants begin to satisfy the market’s voracious appetite for AI memory.
Bottleneck Shifts from Packaging to Memory
For the past two years, every AI chipmaker was constrained by TSMC's CoWoS (Chip-on-Wafer-on-Substrate) advanced packaging capacity. That chokepoint is now easing, with TrendForce projecting output to reach 120,000 wafers per month by the end of 2026. But as AMD’s Su noted in May, the bottleneck has simply moved. "Commodities like memory have become tighter," she told investors, confirming that HBM is the new gating factor for AI accelerator shipments.
The reason is physical. HBM’s stacked-die architecture is more resource-intensive and has lower yields than conventional DRAM. Every wafer allocated to an HBM stack for an Nvidia B300 GPU—which requires 288 gigabytes of HBM3E—is a wafer that cannot produce the DDR5 memory used in laptops, smartphones, or desktops. With Samsung, SK Hynix, and Micron controlling over 95 percent of global DRAM production, their collective pivot toward high-margin HBM has direct consequences for the consumer market.
What It Means for PC and Smartphone Buyers
The consequences are already appearing in prices. TrendForce reported that conventional DRAM contract prices rose a record 90-95 percent quarter-over-quarter in the first quarter of 2026, with a further 58-63 percent increase projected for the second quarter. Gartner estimates that by the end of 2026, combined DRAM and SSD prices will have risen 130 percent from 2025 levels.
This translates to a 17 percent increase in average PC prices and a 13 percent increase in smartphone prices, according to Gartner. The research firm projects memory will account for 23 percent of a PC's total bill of materials by year-end, up from 16 percent in 2025. The shift is so significant that Gartner now expects the sub-$500 entry-level PC segment to disappear from the market entirely by 2028.
Micron's $200 Billion Bet on US Soil
In response to the structural demand, Micron has announced a $200 billion investment to expand its US production capacity across facilities in Virginia, Idaho, and New York. The plan aims to increase the share of its manufacturing on US soil from 10 percent today to 40 percent by 2036, creating an estimated 90,000 jobs in the process. This onshoring effort, supported by the CHIPS Act, represents a long-term strategy to capture the sustained demand from AI infrastructure.
For investors, the narrative is shifting from viewing Micron as a cyclical commodity producer to an AI infrastructure player. The stock trades at just six times earnings, a valuation that CNBC's Jim Cramer argues has not priced in the company's earnings power in a supply-constrained market. While risks remain from future capacity expansion by South Korean rivals SK Hynix and Samsung, Micron's sold-out HBM production and the two-year window before new supply arrives provide a clear runway for growth.
This article is for informational purposes only and does not constitute investment advice.
[1] Micron's 50% Problem: Why Having Too Few Microchips is Making Investors Rich[2] Micron’s 50% Problem: Why Having Too Few Microchips is Making Investors Rich[3] AI Memory Shortage: AMD’s Lisa Su Identifies High-Bandwidth Memory as AI Chip Supply’s Next Cap[4] How Jim Cramer unlocked Micron's path to $1,000 stock price[5] What CEO Sanjay Mehrota Just Said Shows Why Micron Is About to Become a $1,000...
