Anthropic, OpenAI, GDM, and Meta spend more on training than other labs by an order of magnitude. If they felt safe reducing this spend they would. These labs fear getting outcompeted.