The main problem is that an AI training workload eventually ends, or (if inference is also included) at most generates more work proportional to the rate of use

The main problem is that an AI training workload eventually ends, or (if inference is also included) at most generates more work proportional to the rate of user queries. If you have 10 times as many workers, you are done in 1/10th of the time. Crypto on the other hand generates as more and more work as more miners join the network, so that the overall time taken remains constant. This is an essential part of the system, to prevent improvements in technology devaluing all previously mined crypto. The two have a fundamental incompatibility.