Dropbox has outlined how a decade of infrastructure optimization is helping it absorb growing demand from AI without treating new data-center capacity as the only answer. Its work spans forecasting, fleet utilization, storage density, hardware lifecycles, and rack-level power delivery, much of it predating the current AI boom.
The International Energy Agency projects that global data-center electricity consumption will roughly double by 2030 as AI workloads expand, making the ability to extract more useful capacity from existing infrastructure increasingly valuable.
IEA (2025), Global data centre electricity consumption, by equipment, Base Case, 2020-2030, IEA, Paris
Dropbox combines its Magic Pocket storage system with colocated data centers where it manages its own servers and networking equipment. That gives its engineers visibility from software workloads down to racks, power, and cooling, allowing capacity constraints to be addressed at several layers of the stack.
The company begins before hardware is deployed, forecasting demand months or years ahead so additional capacity can be added deliberately while preserving headroom for failures, maintenance, and changing workloads.
Once hardware is running, Dropbox’s Deep Sleep system powers down idle servers or places unused disks into standby. Servers can return to service within minutes, allowing the company to retain spare capacity without paying the full energy cost of keeping every provisioned component continuously active.
When capacity exists but is unevenly distributed, Dropbox instead moves workloads across the fleet. Rebalancing can prevent a local hotspot from triggering additional hardware purchases while unused capacity remains elsewhere in the infrastructure.
Storage growth creates a different problem: fitting more data into the same physical footprint. Dropbox has adopted technologies including shingled magnetic recording to increase storage density, allowing each rack to hold more data without expanding at the same rate as customer demand.
The company measures the result using watts per petabyte and reports that power efficiency across its storage infrastructure has improved by more than 50% since 2020. Dropbox uses the metric because total electricity consumption can still increase as the service grows, even while the amount of power required to store each petabyte continues to fall.
Hardware replacement has also become part of the efficiency effort. Rather than retiring equipment solely according to a fixed age, Dropbox says it uses observed reliability and failure rates to determine how long systems can remain in service. Extending useful hardware lifetimes reduces the amount of new equipment required to maintain a given level of capacity.
The physical limits of the facility became visible when Dropbox introduced its seventh-generation servers. Their higher power requirements exceeded the existing rack design, so engineers doubled the number of power distribution units per rack while retaining the existing busways rather than rebuilding the underlying data-center infrastructure.
That kind of constraint is likely to become more common as AI systems push rack density higher. Gartner forecasts that AI-optimized servers will consume more electricity than conventional data-center servers by 2027, increasing pressure not only on compute capacity but also on power delivery and cooling.
Dropbox’s experience shows that responding to rising infrastructure demand is not only a question of building more data centers. Capacity can also be created by powering down idle hardware, moving workloads, packing more data into existing racks, extending reliable hardware lifetimes, and adapting the physical infrastructure around denser systems.