While that's a great slogan: The infra you need to create and update models dwarfs inference infra. Without that liability, you won't get a working model.
No, the infra itself isn't the moat. Being in an economic position that allows you to expend the resources necessary to run training is one, though. And the cost & speed advantages you realize from utilizing scale (and economies thereof) locally instead of renting means you'll have infra as a signifier of that economic advantage.
And at that point, the additional inference infra is cost-efficient enough that you might as well just have it too, because it's both profit and a better barrier than obfuscation or licensing.
In theory you need the infra V1 to help you design an improve infra v2, and so on and so on. This is a race, not fortification building.
Looking for moats makes sense in a commercial world. I thought the point of the article was that this time its different. You don't need a moat to prevent competitors moving into your market. Nation states are building their own castles, regardless of cost or business case.
This time is, as every single time before, not actually different. The singularity has no calendar appointment for end of 2027, and businesses still care about making money, and will for the foreseeable future.
Nation states with endless pockets are committing to winning the AI arms race.
Business tasked with winning the race will make heaps of money, and everyday people and companies won't be choosing whether or not AI is worth it, we'll be paying with our taxes regardless.
I don;t think the Manhattan Project was business as usual, and I suspect this will be similar.
