Micron’s top executive forecasts an ongoing memory shortage well into the late 2020s, as surging AI workloads push demand for advanced DRAM and HBM products beyond current supply capacity, with industry-wide implications for data centres and emerging AI applications.
Micron Technology expects the memory market to remain tight well beyond 2027 as artificial intelligence continues to soak up supply faster than manufacturers can add it, according to remarks from Executive Vice President and Chief Business Officer Sumit Sadana at KeyBanc Capital Markets’ Annual Technology Leadership Forum.
Sadana said customer demand has strengthened since Micron’s latest earnings report and that calendar 2027 now looks “even tighter than 2026”. He added that the company still has no clear view of when industry supply will catch up, reflecting a shortage that is being driven less by short-term pricing swings than by the long lead times needed to build and ramp advanced memory plants.
He said DRAM remains the main bottleneck for customers, ahead of power, land, data-centre space or logic wafers. The pressure is being amplified by the rise of generative AI, agentic AI and, further ahead, artificial general intelligence. Sadana said agentic AI workloads can require five to 30 times more tokens than comparable chat-based tasks, while deeper reasoning systems demand still more memory and bandwidth.
High-bandwidth memory, or HBM, is becoming increasingly important in that mix because processors can sit idle if they are waiting on DRAM data. Sadana said Micron has already had to balance HBM output against conventional DDR supply, with 100 bits of HBM3E production reducing DDR output by roughly 300 bits. He said the ratio could widen to about four to one with HBM4E, putting additional strain on standard memory availability.
The shortage is now affecting more than data centres, even if that segment remains the most acute. Sadana said some customers cannot obtain more than half the memory they want, despite elevated prices. He added that buyers are altering system configurations largely because they cannot secure enough supply, not simply because they are trying to lower cost. In his telling, that can leave processors underused and create deferred demand for higher-capacity systems once components become available.
Micron is also relying more heavily on strategic customer agreements, which Sadana described as different from older long-term memory contracts. He said the agreements extend across multiple years, with most related revenue expected to fall under terms running through the end of 2030. He added that the deals include binding purchase commitments, take-or-pay provisions and no contractual exits for customers. At the time of Micron’s earnings report, the company had announced 16 such agreements covering $22 billion in cash and cash-like commitments, including $18 billion in cash expected to sit on Micron’s balance sheet.
According to Sadana, some of those arrangements use market-based pricing, while most of the volume is covered by pricing bands with floors set to support gross margins above prior industry-cycle peaks. He said the contracts also deepen engineering collaboration with customers, including product and research road maps extending beyond 2030. He pointed to Micron’s HBM3E device, which he said uses 30% less power than the next-best product, and to work with Nvidia on lower-power DRAM for data centres.
Micron is increasing its planned U.S. investment to $250 billion from $200 billion over the coming years, Sadana said, while also expanding manufacturing in Japan, Taiwan and Singapore and back-end production in India. He said the company has committed $500 million to GlobalWafers for raw wafers and is taking part in a broader $3 billion supply-chain investment effort. Sadana also described Micron as the only company currently investing in front-end memory fabrication in the U.S., saying the Idaho 1 facility should come online in mid-2027, Idaho 2 at the end of 2028, with additional plans in New York and Virginia.
Looking further ahead, Sadana said customised HBM products are likely to become more important as HBM4E arrives and qualification processes remain costly and time-consuming. He also flagged “physical AI”, including robotics and humanoid robots, as a possible long-term demand driver. A humanoid robot, he said, could require hundreds of gigabytes of DRAM and terabytes of solid-state storage to support on-device computing, safety and responsiveness when cloud links are unavailable. Micron expects that market to develop later this decade and potentially accelerate in the early 2030s.
Disclaimer: This content is intended for informational purposes only. Readers are advised to exercise their own judgement, conduct due diligence, or consult a qualified expert before acting on any information provided.
