Source: Bloomberg (paywall)
Microsoft's plan to scale from 12GW to 38GW represents a $150+ billion bet that AI model training and inference will become the dominant workload in cloud infrastructure—roughly 13GW of that new capacity dedicated to specialized silicon. Hyperscalers are shifting capital allocation away from balanced splits across general compute, storage, and networking toward front-loaded investment in custom AI chips and the thermal and electrical infrastructure required to cool and power them. The constraint is no longer compute capacity; it's electrical grid availability and the geopolitical race to secure rare earth materials for chip manufacturing.