Where in my statement do I show to not have understood such obvious assumption? Any consensus mechanism which is highly demanding on computational resources will obviously have fee's, they go hand in hand

I think the "problem" with "blockchain the technology", ie. outside the context of cryptocurrencies is two-fold:
1) Permissioned blockchains still seem rather pointless. For most use cases classical distributed databases (or distributed systems in general) are doing a much better job. Even in the iRobot takedown scenario the main properties that were missing are security and robustness. That you can do in a distributed manner without using "blockchain the technology", in a much more reliable and efficient way at a much lower cost.
2) Permissionless blockchains without PoW-based consensus algorithms so far are often either not really permissionless or not really secure or both. Most of the time you either have (a) committee-based validators or an other form of central coordination which mostly rids of permissionlessness -- which leads back to (1) above. Or (b) variants of Proof-of-Resource like PoS that -- as of now -- are still mostly work in progress in terms of security and robustness.
I guess there could be ways in which permissionless blockchains could support AI and vice versa but I think the overlap where they would complement one another are rather limited. I'm afraid in most cases you'll end up with either a shitty AI or a shitty cryptocurrency.