To say there is a moat or there is no moat is a bit silly because it's not a binary thing "moat or no moat". It's an oversimplification to a point where it obscures what is really happening, which is that some moats are getting smaller, sometikes significantly smaller.
The author correctly points out recent trends where things that were significant moats, like large text and image datasets are now more accessible to individuals.
But to conclude as a main point that therefore data is not a moat anymore is just absurd. Datasets serve different problem domains. We now have accessible datasets for internet text and code, and images from the internet. Talking about whether data is a moat only makes sense with context of what problem domains you are talking about.
Everyone can download the entire text corpus of wikipedia. That's not a moat.
Can everyone download an enormous high quality self driving car dataset, comparable to what Google has constructed from Google maps and many human labelers? Obviously that data is a huge moat.
How about a comprehensive protein dataset? That's something that's gotten much more accessible especially with Alpha Fold, which is great.
But there are still other problem domains where data is a huge moat because collecting the data is super duper expensive and/or requires an expert team of biologists, for example.
To make an analogy, data is like drugs. There are different classes of drugs. Different drugs are used to solve different problems. Some cure you, some harm you, some are well understood and some are not. Some are more accessible than others. But it would be folly to talk about drugs as being good or bad or cheap or expensive without context of problem domains.