You assumed the issue relates only to compute costs. Even with a small payload, bandwidth fees can add up quickly when millions of clients are polling.
Building the paid features announced 3 years ago would pay for compute as well as bandwidth. If the service runs at a loss after that, shut it down or start reaching out to these companies that have integrated it into their products and ask them to sponsor the service, which will certainly be cheaper for the company compared to the cost of dev time to change to some other service that can shutdown just as quick, QA the updates, iterate, approve, ship, wait for adoption, triage bugs, manage all the people doing those things, etc.
Or better yet update the TOS to be free and unlimited for non-commercial use, and to contact sales for commercial use with an estimate of request volume, then sign a phat enterprise contract and rest easy
