My belief is that Google was betting more on DeepMind's vision, which is closer to "thinking" than what LLMs models do.

LLMs have demonstrated some usefulness for sure, but it's no AGI. It makes many mistakes and doubles down on them at times. If Google can nail AGI on easier taks / games / simulations / etc, they'll win that race, cause it'll be easier to add a language layer to a thinking model than to tweak a language model into thinking.