NVIDIA chips are better for training due to the sheer prevalence of CUDA and NVIDIA being ahead of the curve in providing data centre hardware to scale up. They’re also not keeping stuff for just themselves.
When it comes to inference, most of that isn’t really relevant. TPUs or any other specialized hardware will have a higher ROI. But Google aren’t giving them to others either
I know you’re asking for a link, but it’s not really something other than Google can prove. However it is a reasonable extrapolation for anyone familiar with the product space to make.
I think there are different size Gemini models and some of them can fit on a phone with TPUs or similar and since each user will use their own phone, it offloads the computation.
Everything I have read/heard tells me that companies would much, much rather have access to Nvidia chips than Googles TPUs.