Improving Infrastructure Efficiency to Meet Surging Demand in the AI Era

As AI demand surges, the need to expand infrastructure such as data centers, servers, and power is growing. However, simply building new infrastructure is not enough. The key lies in maximizing the efficiency of existing infrastructure amidst constraints such as energy, cooling, hardware availability, and physical space.

Dropbox views and optimizes infrastructure not as individual components but as a single integrated system. It manages capacity planning, hardware lifecycle management, power supply, cooling, and rack design in an interconnected manner. For example, increasing Storage density can reduce the amount of hardware, but more powerful servers generate new energy and cooling requirements. By understanding these interrelationships, Dropbox makes decisions from a holistic system perspective.

To operate infrastructure efficiently, the following strategies are implemented:

• Demand forecasting and planning: Predicting changes in customer workloads to determine when additional resources are needed and reviewing whether new hardware fits within facility constraints.
• Leveraging a hybrid model: Combining Magic Pocket, its own storage system, with Colocated data centers that utilize specialized provider facilities to ensure visibility from software to the physical environment.
• Continuous Fleet optimization: AI-powered features impose new requirements on computing, storage, memory, and networking. Dropbox continuously adjusts its active fleet by reducing hardware usage during low demand or moving workloads to locations with available resources.

Additionally, to reduce the energy consumed by idle hardware, methods such as Deep Sleep are used to prevent unnecessary power consumption, and the focus is on maximizing the available capacity of existing resources before expanding infrastructure.