Improving infrastructure efficiency for growing demand in the age of AI

Doing More With What’s Already Built

The industry’s default answer to AI’s growing appetite for compute and storage has been to build more: more data centers, more servers, more power. That answer is incomplete. Energy, cooling, hardware availability, and physical space are all finite, and the teams operating large fleets have to make the infrastructure they already own work harder.

Dropbox’s Infrastructure and Datacenter Engineering groups have spent more than a decade optimizing that full stack as one system, not as independent silos. Capacity planning, fleet optimization, hardware lifecycle management, power delivery, cooling, rack design, and facility planning all interact. A change in one layer creates ripple effects in another. Increasing storage density, for instance, can cut the number of servers required, while more powerful hardware can introduce new energy and cooling demands. Working through those tradeoffs is what allows engineering teams to create room for growth before expanding the physical footprint.

Planning Capacity Before It’s Needed

Efficiency decisions often begin months or years before hardware lands in a data center. When a user uploads a file or queries Dash, the response needs to be immediate. Delivering on that expectation means forecasting how customer demand and workloads will evolve, when new resources will be required, and whether the existing environment can absorb them.

Dropbox runs a hybrid model where its Magic Pocket blob store (built for the company’s scale) runs in colocated facilities where the company owns its own servers and networking gear. That arrangement gives engineers visibility across software, hardware, and the physical environment. As AI-driven products shift the scale and shape of demand, planning has to account for more than raw capacity numbers. New hardware must fit within a facility’s power and cooling limits, and those limits vary from site to site. The engineering work is in understanding those constraints early so capacity is added deliberately, with adequate headroom for growth, maintenance, and failures.

Keeping the Active Fleet in Balance

Once capacity is deployed, the assumption that workloads will behave as forecast goes out the window quickly. Customer behavior shifts, products evolve, and AI introduces new demands on compute, storage, memory, and networking. That makes efficiency an ongoing process, not a one-time optimization.

Several initiatives target different layers of this problem, all aiming to extract more useful capacity from what is already online.

Putting Idle Hardware to Sleep

Infrastructure is provisioned for peaks, which means most hardware is not needed at all times. Keeping every component fully powered during lulls wastes energy without delivering value. Dropbox’s Deep Sleep initiative addresses this by letting fleet management algorithms determine which servers can be placed in standby, spinning down hard drives into standby mode or powering off idle servers entirely. Servers can return to service within minutes, which lets Dropbox balance its energy savings against performance and reliability needs. For latency-sensitive workloads, selective drive spin-down is used rather than powering down entire machines. The core challenge is safely identifying hardware that can power down without compromising the capacity needed for operations.

Balancing Workloads Across the Fleet

Having enough aggregate capacity is only useful if it is located where the work is happening. One part of the fleet might be at its limit while another sits mostly idle. To avoid deploying new infrastructure to solve a localized constraint, Dropbox continually monitors spare capacity and workload distribution. When imbalances appear, teams rebalance workloads or bring dormant capacity online. Some adjustments run automatically; larger changes pass through engineering review. This constant adaptation helps the company avoid treating a local bottleneck as a fleet-wide shortage.

Packing More Data Per Drive

Storage density gains multiply across the fleet. Shingled magnetic recording (SMR) packs data more densely onto hard drives without changing their physical size. When each drive holds more data, fewer drives are needed for the same payload, which can mean fewer servers and racks, less cabling, and lower power and cooling overhead.

The compounding effects are visible in Dropbox’s efficiency metrics. Raw power consumption is not the full story, since overall energy use can rise as the business grows even when the fleet is becoming more efficient. Instead, Dropbox tracks watts per petabyte. Since 2020, that metric has improved by more than 50%, meaning it takes less than half the power to support a petabyte of storage today than it did at the start of the decade.

Extending Hardware Life and the Environment Around It

Replacing equipment too early leaves useful capacity on the table; keeping it too long introduces risk. Hardware does not automatically become unreliable at a certain age, and Dropbox relies on production data like annual failure rates to inform lifecycle decisions rather than following a fixed replacement schedule. When gear continues to meet operational standards, Dropbox extends its useful life, and hardware is repaired whenever that is practical. When a machine can no longer serve production needs, the company works with partners to resell or responsibly recycle it.

Lifecycle extension has limits, and newer hardware eventually needs to land. As each server generation increases compute and storage density, the physical environment has to keep pace. A machine that stores more data or delivers more compute typically draws more power and sheds more heat, changing rack design and cooling requirements.

A recent deployment of Dropbox’s seventh-generation servers illustrates the kind of constraint that emerges. The new hardware’s power requirements exceeded the capacity of the existing rack power design. Rather than rebuilding the facility infrastructure, the Hardware and Datacenter Engineering teams redesigned the rack power architecture, doubling the number of power distribution units (PDUs) per rack while keeping the existing busways. The new hardware went into production without major changes to the data center itself.

Optimization does not end with the silicon. As hardware evolves to meet AI-era demand, each generation must fit the power, cooling, and space that already exist, and provisioning decisions made years ago determine whether that fit is possible. The win is in making those decisions as part of a single system, where the full lifecycle from planning to retirement is accounted for.

Lessons From a Decade of Infrastructure Work

The efficiency work described here builds on more than ten years of engineering across Dropbox's infrastructure. Different teams have tackled capacity planning, fleet optimization, storage systems, hardware lifecycle management, and data center engineering. Each area addresses its own set of problems, but collectively they aim to extract more value from the infrastructure that runs the company's products.

This is not a one-time effort. Customer needs shift, hardware generations change, and new technologies bring both opportunities and constraints. The result is a continuous process of refinement, where each improvement lays the groundwork for the next.

These lessons have become more important as industry-wide demand for digital infrastructure grows. AI is accelerating the need for storage and compute, but the core engineering challenge remains the same as it always has been: infrastructure must scale without sacrificing reliability, efficiency, or resilience. As AI technologies evolve, Dropbox plans to keep building systems that support existing products while retaining the flexibility to accommodate whatever comes next.