KPI Card breaking only after being published

Been a while since you asked your question but I ran across this post while trying to solve my own problem, thought I'd mention it.  I had a similar issue with a Card visualization, which turned out to be an issue for the entire page.  This page revolved around average time to closure of helpdesk tickets.  When designing and viewing the data in Power BI Desktop, no issues.  Each ticket has a column for the time it was closed, so one of my main filters was to ignore any tickets where that value was blank (i.e. the ticket wasn't closed yet).

This dashboard was already in Azure and linked up to a promoted semantic model that was refreshing on a regular schedule.  What I observed was when I would publish this report and then immediately refresh the dashboard in Azure, the numbers would look correct for a few seconds, but slowly started reverted to extremely large (small? 🙂 ) negative numbers.  What I eventually found was that while I was connected into my data source on prem, those values being blank was fine, but when the semantic model was storing that data in Azure, it was converting the blank values to 12/30/1899 12:00:00 AM.  So suddenly, I was pulling in all these tickets that had a -124 year closure time, causing my numbers to go wack.  No idea how to fix those dates from filling in, but I do know how to filter out that date, which essentially fixed my issue.

This obviously isn't the same as what OP posted but wanted to throw it in here really quick in case someone else comes across this in the future.  Check your semantic model in Azure and make sure it isn't transforming your data somehow.