How Does Amazon Split Memory Between AI and Core Customers?

Andy Jassy says Amazon has to split scarce memory between HBM for AI accelerators and DRAM for core, non-AI customers. He says 2026 capacity is sold out, 2027 is largely sold out, and a good chunk of 2028 is sold out. He does not think the shortage of data center space, power, and memory ends in 2027.

AWS leaves some capacity for on-demand use in its inference service, Bedrock, or EC2, but the big chunks are largely sold out. The AI business and the core business are both growing at rapid rates.

Andy Jassy says the starting point is being customer-focused. In the pandemic, he thinks competitors tried to optimize within quarters and drive as much revenue as possible while companies were cutting costs. AWS sent its field team to help customers cost-optimize, knowing that would cut revenue in the short and medium term. The same approach applies to core customers: help them succeed over a long period, and the business comes along with them. He says scarcity could instead be used to squeeze core customers out.