It depends on what you are mining, and your cost modeling. For instance, NVIDIA cards have better memory bandwidth and both dagger-hashimoto and equihash are memory hard algorithms, and NVIDIA cards have better expected resale value (though this will vary based on your expected sales channel.)

AMD is generally more popular among hobbyist miners, I think, because early on they had a performance advantage in Ethereum and hobbyists tend to follow each other like a herd because they have limited time to do experimentation and exploration (and this is not a criticism of them, they have day jobs and it takes a lot of effort to even get miners going, let alone optimize them across multiple parameters.)

I thought they use AMD cards because they consume less power

Those two statements are basically equivalent when it comes to mining. cost/hash/second is the metric you optimize for. AMD cards can do the integer operations required with less power (and faster!)

NVidia actually consumes less power in most games.

But AMD cards use less power with integer performance.

This is the reason, Nvidia is good for floating point. Which is why people use them for deep learning.

The main reason is library support. Still no widely supported (or ported) equivalent to Cuda/CuDNN for AMD.

Power does not matter that much if you spend $3 of it to earn $100 every month. It mostly depends on payback time.