Fungible Data Center

Overview

The Fungible Data Center is an architecture paradigm articulated by Google that enables operators to defer infrastructure commitment decisions until close to deployment time. The core idea is that standardized interfaces across power, cooling, mechanical/server-hall layout, and telemetry should make it possible to swap accelerator types — for example, choosing GPUs over TPUs — within approximately a 30-day window, rather than locking in decisions years in advance during facility construction.

Fungibility requires convergence across four domains, as outlined in the AI Data Infrastructure Standards Open Letter:

1. Power — standardized high-voltage DC distribution (e.g., Mount Diablo ±400 Vdc)
2. Cooling — standardized CDU interfaces (e.g., Project Deschutes)
3. Mechanical / Server Halls — common rack form factors and floor layout standards
4. Telemetry — standardized DC telemetry interfaces for monitoring and control

The concept applies across both Google-owned data centers and third-party colocation facilities. Because Google now operates both GPU and TPU generations simultaneously, infrastructure that can accommodate either without major re-engineering represents a significant operational and capital efficiency gain.