Innodisk is expanding its DDR5 memory portfolio for AI and edge computing systems, with a focus on balancing raw performance against the reliability demands of industrial deployments. Tom Lee, director of DRAM product management at Innodisk USA, spoke about how AI is reshaping the memory market and what that means for product development teams.
Lee said the rapid growth of AI has structurally changed the DRAM market, ending its traditional cyclical pattern. "Compute units such as GPUs and TPUs have massive demand for HBM, LPDDR, and DDR5, leading to the crowding out of traditional DRAM production capacity and causing supply tightness," he said.
The shift isn't limited to data centers. "AI servers and data centers were mainly focused on training. Now, they are gradually starting to shift toward inference led by edge AI, meaning an expansion from servers to edge AI and AI PCs," Lee said. "This will be another wave that causes memory capacity and specifications to surge."
What AI systems need from memory
Lee identified four requirements driving next-generation memory: higher bandwidth, larger capacity, better power efficiency, and strong reliability. For industrial applications, he said reliability can outweigh raw speed. "Many of these systems operate 24/7 in harsh environments, so features like wide-temperature support, long product longevity, and consistent long-term performance are just as important as speed."
For product development teams, the balance between performance and reliability is shifting. AI workloads moving to factories, transportation systems, robotics, and other edge environments require a different combination of capabilities than server-class hardware. "Customers are looking for the right balance of performance, power efficiency, reliability, thermal capability, and long-term support," Lee said.
Those requirements affect how product teams plan for memory, especially as the shift from DDR4 to DDR5 continues to play out in per-generation design cycles.
Building memory for edge and industrial AI
Innodisk developed its AI Memory Series to address bandwidth, capacity, and power efficiency. For industrial uses, the company adds wide temperature support, conformal coating, and anti-sulfuration treatment to protect against shutdowns in demanding environments.
"When we talk about mission-critical applications, reliability is no longer just a product feature; it's a system requirement," Lee said. "Even a single memory error can affect overall system availability, so we place as much emphasis on long-term reliability and validation as we do on performance."
Lee said the future direction is not just faster memory, but the right memory for the application. "As AI expands from cloud computing into edge AI, industrial automation, robotics, and autonomous systems, memory requirements will become much more diverse," he said. "Customers will place greater emphasis on long-term reliability, power efficiency, product longevity, and supply continuity."
Why this matters for product development
Product development teams planning AI systems should plan for two budget realities. First, technology demand for HBM and DDR5 is expected to tighten general DRAM supply, so lead times and allocation matter as much as performance specs. Second, reliability requirements differ significantly by deployment context; a railroad control system and an AI PC do not share the same thermal or longevity constraints.
The broader shift toward edge placement means product teams may want to re-evaluate existing memory specs rather than porting over server-grade parts. For teams looking for a frame for integration under disruption, an AI Learning Path for Product Managers covers how to evaluate AI hardware trade-offs. A broader resource on AI for Product Development offers working guidance on aligning hardware choices with ship expectations.
Your membership also unlocks:
