He’s referring specifically to local RAM used to run local AI models, whichl he argues is unused most of time. Say if you use your local AI for 4 hours a day, it’s idle for 20. Where the same model in a cloud might serve users across time zones 24/7.
It’s a similar argument about private car ownership, where a car might be used 5% of the day for a lot of people while 95% of the time it requires storage space, aka a parking spot somewhere.