Training costs for foundation models like GPT-5 and Gemini 2/3 are expected to exceed $1 billion just to develop the base model, and largest current data centers use about 200,000 chips; at this trajectory of spending, $500 billion would be consumed relatively quickly.