AI memory boom reshapes chip spending and China power grid

This briefing captures the less glamorous but absolutely critical side of the race to AGI: supply chains and power grids. If memory really grows from single digits to around 30% of hyperscale data center capex in just three years, that’s evidence that the bottleneck is shifting from raw GPUs to the bandwidth and capacity needed to keep them fed. For leading AI companies, this changes which partnerships and contracts matter most – HBM vendors and DRAM fabs become as strategic as GPU suppliers, and firms with long-term memory allocations, like Nvidia, gain even more leverage.([finance.sina.com.cn](https://finance.sina.com.cn/headline/2026-04-05/doc-inhtkvrf5532463.shtml))

On the energy side, China explicitly linking “compute–electricity coordination” into its national work report signals that AI is now a first-class driver of grid planning, not an afterthought. State Grid and China Southern Power Grid building “new-type” power systems for AI-era loads suggests an emerging model where governments treat compute as critical infrastructure, much like highways or ports. That could accelerate large-scale training and inference deployments inside China while other regions struggle with aging grids and transformer shortages.

If memory and power become the binding constraints on scaling models, whoever solves those constraints fastest effectively pulls the AGI timeline forward. This story hints that Chinese policymakers and suppliers see that opening and are moving aggressively to occupy it.