The market thinking we're going to be oversupplied with compute is utterly stupid (Save this).

Claims

• Only 20.6% of U.S. firms currently use AI in regular business functions, per Goldman Sachs and Census Bureau data, with even leading sectors barely past one-third adoption.
• Nvidia GPU orders have grown to $1 trillion through 2027, double a year ago, with lead times stretching to nearly a year and all three HBM suppliers sold out for 2026.
• Data center vacancy sits at 1% for the second consecutive year, with 92% of capacity under construction already pre-leased.
• The binding constraint on AI infrastructure is now power, not chips, with grid interconnection queues exceeding five years in most U.S. regions.
• Agentic AI workloads demand 10 to 100 times more compute than simple chatbot queries, meaning the shift from training to inference multiplies rather than reduces demand.