SoftBank and Intel form Saimemory, a company to develop a prototype within two years of stacked DRAM chips for AI that consume ~50% of the power of current HBM
TOKYO — SoftBank and Intel are developing a type of memory for artificial intelligence expected to consume much less electricity …
Context & Ripple Effects
Saimemory shifts the SoftBank-Intel connection from reported discussions over an AI chip—talks that did not meet SoftBank’s requirements—to a focused memory-development effort. It enters an HBM market already advancing through products such as Samsung’s higher-capacity HBM3E and investments in the packaging needed to stack DRAM dies.
First-order effects
- SoftBank and Intel now have a dedicated vehicle to pursue a stacked-DRAM prototype with a stated power target relative to current HBM, concentrating their near-term work on memory rather than a complete AI chip.
- The effort puts power efficiency alongside bandwidth and capacity as a design criterion for AI memory, while leaving commercial viability dependent on achieving the prototype target.
Second-order effects
- Incumbent HBM suppliers and advanced-packaging players face a clearer incentive to emphasize energy efficiency in addition to stacking density and throughput; SK Hynix’s planned HBM packaging capacity illustrates why packaging is central to that competition.
- AI-system builders could gain another memory-design path if the prototype succeeds, potentially broadening component choices beyond conventional HBM rather than changing current deployments immediately.
Third-order effects
- The project supports a longer-term shift in AI infrastructure from optimizing individual accelerators toward co-designing memory, packaging, and compute around power constraints.
- If lower-power stacked memory proves manufacturable at scale, differentiation in AI hardware may increasingly accrue to integrated memory-and-packaging architectures, not just processor performance.
The trend: AI infrastructure is moving toward memory-centric co-design as power consumption becomes as consequential as raw bandwidth for scaling AI systems.