This simulation uses a Time-to-Train proxy derived from regional hardware
constraints.
Global North (Baseline)
Assumes unrestricted access to H100/H200 clusters and massive energy grids.
China
Penalty Factor: 1.5x - 2.0x
Despite high capacity, export controls force the use of less efficient chips (e.g., H20 or
domestic alternatives), increasing training time.
India
Penalty Factor: 3.5x - 4.0x
Rapid growth, but total FLOPs capacity is significantly lower than China/USA. Heavily
reliant on cloud imports with queue latency.
Global South
Penalty Factor: >15x
Severe lack of local data centers. Training relies on high-latency, expensive cross-border
cloud access or older hardware.