Both Google and Amazon charge based on compute heft per second (Amazon in GB-seconds, Google in a mix of GB-seconds and GHz-seconds). When you multiply that out, it overtakes EC2 pricing at a point; I've done the math out for different types of sites before, and it obviously varies by stack, but you hit those breakpoints a lot sooner than you might otherwise expect. One project's workload was something like 10 req/sec and the Lambda costs were more than three load-balanced m4.larges (edit: load-balanced for fault tolerance more than load) due to RAM usage that was per-request in Lambda but amortized across all requests on instances.

Serverless stuff can make a lot of sense for low-utilization tasks under this model, but for applications with nontrivial load you may be paying for a lot of GB-seconds on an ongoing basis. Couple that with the additional latency inherent to spin-up times when putting Lambda behind an HTTP endpoint, and it also runs the risk of being a kinda-janky experience for the user, too (though this can be mitigated), and I'm very not-a-fan past toy/proof-of-concept problems.

Perfect explanation as usual.