Dropbox says a decade of infrastructure optimization is helping it absorb AI workloads without treating new data centers as the only source of capacity. Its approach combines long-range forecasting, tighter fleet utilization, denser storage, extended hardware life and targeted power upgrades.
The company’s Deep Sleep system shuts down idle servers or places unused disks on standby, while keeping spare equipment available to return within minutes. Engineers rebalance workloads when free capacity exists in the wrong part of the fleet, preventing a local hotspot from triggering unnecessary purchases. Shingled magnetic recording packs more data into each rack, and observed reliability—not a fixed age—determines when hardware is retired. Dropbox reports that electricity used per petabyte across its storage infrastructure has fallen by more than 50% since 2020.
Physical limits still appear. Seventh-generation servers exceeded the power design of existing racks, so engineers doubled power-distribution units while retaining the facility’s busways. Total energy use can also rise even when each stored petabyte becomes more efficient. The lesson is therefore not that optimization removes the need for construction, but that software placement, sleep states and hardware lifecycle decisions can delay it and preserve failure headroom. That matters as AI servers increase both rack density and the pressure on power and cooling systems.