technically in order for something to be turing complete it needs infinite memory

It’s possible (likely, even) to have a chip fast enough for inference, but not fast enough or with enough memory to do meaningful training runs. Like the current DGX spark.

There is a reason why Google has tpu8i and tpu8t