Overview
- In May, AWS leadership told internal teams to identify and turn off underused EC2 instances to free CPU capacity after engineers began reporting multi-day waits for new CPU-backed servers.
- The shortage is driven by new agentic AI patterns that push more orchestration and tool calls onto CPUs, turning previously low-use virtual machines into a scarce resource.
- AWS has updated its Compute Optimizer to flag idle or 'ghost' instances by checking two-week CPU and network activity and is enforcing rightsizing and automated shutdowns across teams.
- Cloud customers are feeling the strain through tighter on-demand and spot supply and repeated price rises for machine-learning capacity in 2026, while contracted reservations have remained largely insulated.
- Amazon is adding physical capacity over the medium term, citing a recent 3.8 gigawatt power increase and plans to double capacity by 2027, but chip and data-center buildouts mean operational fixes and pricing are the immediate tools to rebalance supply.