Memory supply gap stretches beyond 2028 as cloud capex tops US$725 billion

The surge in cloud capex to $725B reflects AI workloads’ insatiable demand for HBM and DDR5 bandwidth, forcing memory makers to lock in 3D stacking, TSV, and CoWoS capacity years ahead of schedule. This accelerates heterogeneous integration, straining EDA and advanced packaging test equipment supply chains. Regulatory risks loom as the U.S. and EU consider export controls on AI-optimized memory akin to those on GPUs—non-U.S. fabs without domestic production face client attrition. Strategically, Samsung leverages its HBM4 lead to anchor NVIDIA, SK hynix bets on CPO to bypass bandwidth limits, while Micron retreats to niche segments. Over the next 18 months, smaller cloud providers lacking long-term memory deals will be priced out of AI infrastructure, catalyzing third-party memory pooling as a new service layer.