Does somebody know how they designed that?

It is not bigquery, because bigquery is fast but multiple seconds to respond to queries.

It could be bigtable, but it just seems expensive and keeping this many timeseries up to date all the time and few of them are looked at that often. Bigtable starts at $500 usd for a single node and you need multiple.

I understand that there is a lot of money in Cloud. I also tried to search and could not find any resources on how this is done on large scale.

How can I do something like that in my own applications?

I have looked at Postgres TimescaleDB and other timeseries databases, Prometheus, but you quickly end up with a lot of timeseries that takes a lot of memory to compute (Prometheus running out of memory).

They have made an article here talking about OpenTSDB https://cloud.google.com/architecture/monitoring-time-series-data-opentsdb