Intel CEO Lip-Bu Tan signaled a potential return to memory manufacturing, calling the sector no longer a commodity business but a strategic asset in the AI era. In a TechSurge podcast, Tan described new memory architectures as one of his 'pet projects' and flagged the sector as ripe for innovation. He hinted at stacking DRAM directly onto CPUs, saying 'there are a lot of ways we can really do stacking together.' The company recently hired Seok-Hee Lee, former SK Hynix CEO, as executive vice president of its foundry division.
Intel has two parallel memory efforts in motion: XBM (Cross-Batch Memory), which eliminates the silicon interposer used in conventional HBM, leveraging chiplet interconnect instead for direct 3D stacking onto logic; and Z-Angle Memory (ZAM), a low-power stacked DRAM co-developed with SoftBank's Saimemory, targeting prototype by 2028 and commercialization by 2030. Both aim to address the memory-bandwidth bottleneck now strangling AI accelerators. Current market forecasts show memory supply sold out through 2027–2028 at elevated prices.
For architects: Intel's return to memory matters because memory is now the binding constraint in compute systems. Stacking on-die or in-package DRAM reduces latency and power, critical for both training and inference. XBM targets post-2030 deployment, but if commercialized, could differentiate custom-silicon strategies. However, Intel has failed at memory bets before (Optane, HMC); success requires ecosystem buy-in beyond Intel's own products.