Market expectations are growing that Nvidia could switch the memory in its next-generation AI inference accelerator, Rubin CPX, from GDDR7 back to high-bandwidth memory. The view is driven by assessments that large-scale production inference workloads require more bandwidth and capacity than GDDR7 can efficiently provide. Rising inference demand cited by CEO Jensen Huang and comments on supply constraints have added momentum. A switch would require redesign and could affect HBM supply-demand outlooks.

TeraView, a terahertz technology solutions company, said on Feb. 26 it secured additional orders for its EOTPR terahertz-wave inspection system from a global semiconductor foundry and a smartphone company. Both customers had already adopted EOTPR. The foundry, which supplies key parts to AI chipmakers, extended a maintenance contract ahead of adding automation functions in the second quarter. The system detects defects in HBM-related micro bumps, TSVs and flexible PCBs.