Micron Says 2027 Will Be Even Tighter Than 2026 as Customers Get Less Than Half the DRAM They Want
KeyBanc Technology Leadership Forum, August 10, 2026
Micron Technology used a fireside chat at the KeyBanc Technology Leadership Forum to deliver a message that will reverberate across the memory supply chain: the shortage is getting worse, not better, and the company now believes calendar 2027 will be tighter than 2026. Executive Vice President Sumit Sadana told analyst John Vinh that aggregate demand signals from customers have increased further since the company's last earnings call, and that the growth in demand is now outpacing the growth in supply on an industry-wide basis for 2027. "We now expect that 2027 calendar year will be even tighter than 2026," Sadana said, a notably more aggressive stance than the tight-through-2027 framing Micron had previously offered.
Capex Ramps to $250 Billion as Micron Chases a Moving Target
The company has raised its U.S. manufacturing investment commitment from $200 billion to $250 billion, and is backstopping that build-out with supply chain investments including $500 million into GlobalWafers for raw wafer supply and a broader $3 billion commitment across its supply base. Despite the scale of spending, Sadana was blunt about the limits of Micron's control over the situation: "We still don't have line of sight as to when the supply is going to be able to meet demand because demand continues to escalate at a very rapid pace over time." He attributed this to a structural mismatch between AI demand growing in the digital domain and memory supply having to be built in the physical domain, a gap he argued the market has not fully priced in given how long it takes to bring a leading-edge fab online.
Strategic Customer Agreements: A New Contract Architecture With Teeth
The most concrete new disclosure was around Micron's Strategic Customer Agreements, which Sadana positioned as fundamentally different from the legacy long-term agreements the industry used historically. Unlike prior LTAs, which were unenforceable 12-month handshake arrangements, the SCAs are binding, take-or-pay commitments with no contractual outs, mostly running five years through the end of calendar 2030, with smaller automotive customers on three-year terms. At the time of Micron's last earnings report, the company had signed 16 SCAs backed by $22 billion in cash and cash-like commitments, of which $18 billion is straight cash landing on the balance sheet; Sadana confirmed additional SCAs have been signed since then.
The pricing mechanics are notable for investors modeling margin durability. Most SCA volume carries a price band: a ceiling tied to the pricing prevailing when each agreement was signed (the initial 16 were priced off Micron's fiscal Q2 pricing, with subsequent agreements priced at whatever the market rate is at signing), and a floor set at a level Sadana said guarantees gross margin "well above any prior peak in the cycles of the industry." The agreements also carry an evergreen structure allowing years to be added to the back end, which Sadana described as customers signaling they view memory as a decade-plus strategic relationship rather than a cyclical hedge. Asked directly whether these agreements could be broken the way COVID-era LTAs were once supply normalizes, Sadana pointed to the binding terms and the absence of contractual exits, while also noting pointedly that customers who have treated memory "tactically" in the past are the ones struggling most for allocation today.
Data Center Demand: Customers Getting Under Half of What They Want
Sadana disclosed that in the data center specifically, Micron is frequently unable to meet more than half of customer demand, and that this constraint shows up as a multiyear signal rather than a one-quarter blip. He was explicit that despite historically high pricing, customers are not backing off: "They're eager to get more supply because they are not able to meet their own business case requirements." On the debate over server memory de-specification, Sadana pushed back on the idea that this reflects price sensitivity, framing it instead as a rationing response to supply scarcity — customers trading average capacity per unit for higher unit volumes — while noting that aggregate DRAM demand across those units is still rising even after accounting for lower average density per server.
Physical AI and Robotics: The Next Demand Wave Could "Dwarf" Today's AI Cycle
Perhaps the most striking forward-looking comment centered on physical AI and humanoid robotics, which Sadana described as still in its infancy but capable of eclipsing the current AI-driven memory cycle. He pointed to expectations that each humanoid robot will require hundreds of gigabytes of DRAM and terabytes of SSD storage to support fast response times and functionality without constant cloud connectivity. Multiplied across an eventual global deployment base, Sadana argued the aggregate memory requirement "is going to just be astronomical," projecting that this vector begins accelerating later this decade and hits an exponential growth phase in the early 2030s — a timeline that, if accurate, extends the memory supercycle narrative well beyond the current data center buildout.
HBM Economics and the Shift Toward an ASIC-Like Supplier Model
On HBM, Sadana reiterated the increasingly damaging trade ratio between high-bandwidth memory and conventional DRAM supply — producing 100 bits of HBM currently displaces 300 bits of DDR capacity at the HBM3E generation, a ratio that worsens to roughly 4:1 at HBM4 and HBM4E, further squeezing the wafer supply available for standard DRAM. More importantly for competitive positioning, Sadana argued that HBM design and qualification costs are prohibitive enough that most customers will not qualify all three HBM suppliers for a given platform, pushing the market toward a structure where most projects are single- or dual-sourced. "We think that the market is going to become very much like an ASIC-like market more and more," he said, extending this thesis beyond HBM to other differentiated memory products such as data center low-power DRAM, where Micron says it was sole-sourced with Nvidia until supply constraints forced a second source.
Competitive Positioning Against Samsung and SK Hynix in the U.S.
Responding to a question about Samsung's Texas expansion and SK Hynix's Indiana investment, Sadana drew a sharp distinction: Samsung's Texas investment is in logic foundry capacity, not memory, and SK Hynix's Indiana site is back-end assembly and packaging, not front-end fabrication. Micron, he said, remains the only company building front-end leading-edge memory fabs in the U.S., citing its Idaho fabs coming online in mid-2027 and late 2028, a New York fab cluster beyond that, and a Virginia site dedicated to long-life-cycle and automotive DRAM. Micron is baking a pricing premium for U.S.-made supply into its planning, betting that customers will pay up for supply chain resilience.
DRAM Capacity Constraints Are Spilling Into NAND
Asked about the risk of customers dynamically offloading cold data from DRAM to flash to reduce DRAM requirements, Sadana said the opposite dynamic is actually playing out: AI systems, particularly those running agentic and reasoning workloads with lengthening context windows, are so DRAM-constrained that overflow data is being pushed into NAND rather than the reverse. He noted that agentic AI interactions require five to thirty times the tokens of a simple chat interface for comparable tasks, intensifying both the bandwidth and raw capacity requirements simultaneously — a dynamic he said is stressing both of Micron's product lines rather than cannibalizing one in favor of the other.