At KeyBanc's Annual Technology Leadership Forum, Sumit Sadana, Executive Vice President and Chief Business Officer of Micron Technology (MU.US), stated that the supply tightness in the memory chip market is expected to extend beyond 2027 due to sustained demand from the AI boom outpacing the industry's capacity expansion pace.

Sadana noted that customer demand signals have strengthened since Micron's latest earnings report, with the company now expecting the calendar 2027 supply-demand balance to be "tighter than 2026." He added that it remains unclear when supply will be able to match demand. For clients, the "primary constraint" is DRAM itself, not power, real estate, data center capacity, or logic wafers. He attributed the imbalance partly to the difficulty and long lead times involved in building and ramping up advanced memory chip fabrication facilities.

AI Demand Reshaping the Memory Market Landscape

Sadana characterized the current environment as fundamentally different from previous memory industry cycles, citing the expansion of generative AI, agentic AI, and future general-purpose AI applications. He noted that agentic AI workloads may require 5 to 30 times the number of tokens needed for traditional chat interface tasks, while deep reasoning workloads demand even more. He also highlighted the growing importance of high-bandwidth memory (HBM) in AI systems, explaining that processors can idle while waiting for DRAM data, making higher memory bandwidth and capacity critical for improving system utilization.

Micron has previously discussed the trade-off between HBM production and regular DDR memory supply. Sadana stated that producing 100 bits of HBM3E reduces DDR output by approximately 300 bits, a roughly 3:1 capacity displacement ratio. With the arrival of HBM4E, this ratio could approach 4:1, further straining conventional memory supply. While data centers are the most concentrated area of demand, Sadana noted that demand is high across all segments. Some data center clients, even willing to pay higher prices, can secure less than half of their target procurement volumes. He emphasized that customers are adjusting system memory configurations primarily due to supply constraints, not pricing. Although reducing DRAM capacity per system could allow clients to ship more units, it may also lower processor utilization and create a latent demand for higher-capacity configurations once supply improves.

Strategic Customer Agreements

Sadana detailed Micron's strategic customer agreements (SCAs), which he said differ fundamentally from historical long-term contracts in the memory industry. These agreements span multiple years, with most SCA-related revenue falling under terms extending through the end of calendar 2030. Unlike previous arrangements, Sadana noted that SCAs include binding purchase commitments, "take-or-pay" clauses, and no contractual option for clients to exit. As of Micron's latest earnings, the company had announced 16 agreements involving $22 billion in cash and cash-equivalent commitments, with $18 billion in cash expected to remain on Micron's balance sheet.

Some agreements use market-based pricing, while most cover volumes with price ranges where the floor is set above gross margin levels seen in previous industry cycle peaks, Sadana said. He added that these agreements also foster deeper engineering collaboration with clients, covering product and R&D roadmaps extending beyond 2030. He cited Micron's HBM3E product as an example, claiming it consumes 30% less power than competitors' offerings, and mentioned collaboration with Nvidia to bring low-power DRAM into data centers.

Investment and U.S. Manufacturing

Sadana stated that Micron is increasing its planned U.S. investment from $200 billion to $250 billion, while also investing in its global network of back-end manufacturing facilities in Japan, Taiwan, Singapore, and India. He revealed that Micron has committed $500 million to GlobalWafers for raw material wafers and is participating in a broader $3 billion supply chain investment program. Sadana noted that Micron is the only company investing in front-end memory chip manufacturing in the U.S., with its Idaho Fab 1 expected to begin production by mid-next year and Idaho Fab 2 expected by the end of 2028. The company also plans to build a wafer fab cluster in New York and invest in Virginia, including the introduction of 1-alpha DRAM technology. Sadana expects Micron's U.S. manufacturing footprint to command a pricing premium and provide clients with greater supply chain resilience.

HBM and the Physical AI Opportunity

Looking ahead, Sadana pointed to opportunities for customized HBM products with HBM4E. Due to the costly and time-consuming certification and co-engineering processes, he expects most HBM projects to use one or two suppliers. He specifically highlighted "physical AI," encompassing robotics and humanoid robots, as an emerging long-term demand driver. Sadana noted that a single humanoid robot may require hundreds of gigabytes of DRAM and several terabytes of solid-state storage to support onboard computing, safety features, and rapid response when cloud connectivity is unavailable. While the robotics field is still in its early stages, he expects this segment to grow in the latter part of this decade and potentially enter a faster growth phase in the early 2030s.

Micron Technology is a global semiconductor company specializing in designing and manufacturing memory and storage solutions. Its portfolio includes DRAM, NAND flash, solid-state drives, memory modules, and embedded memory for a wide range of computing and electronic devices. Micron serves data centers, enterprise and cloud infrastructure, client computing, mobile devices, automotive systems, and industrial applications, while also offering products under the Crucial brand for consumers. Founded in 1978 and headquartered in Boise, Idaho, the company has grown into a multinational manufacturer with R&D and production facilities worldwide.