I got some data to share. Walking through the dump for the first 1000 entities (~19Mb), took 0.008 seconds per item, where in each step the following things where done:
* read line from file
* json_decode the line
* use the EntityDeserializer to turn the array into DataModel objects
Given that these entities are on average a lot bigger than the typical one found in Wikidata, it looks like the average deserialization time is a few milliseconds. So now I really wonder why people are blaming DataModel 1.0. Everything seems to indicate most time is spend in Wikibase.git and MediaWiki.