Really depends on what you’re trying to achieve. Take cost for example.

Let’s say that to serve a decently available service you need at the very least a Elastic Load Balancer (~18$/month) and two decent size instances (let’s say c4.large, so ~150$/mo). So roughly, 170$/mo.

With API Gateway+Lambda, assuming you need 128MB RAM for the function, and about 1 second of compute for each call, you can get a million calls for roughly 6$. So for 170 bucks a month you can get at least 28 million calls per month, which is roughly 11 requests/sec. Not bad. However, c4.large instances can handle way more than that.

Our data platform handles almost 200 billion calls each month, if we needed to pay 6$ per million requests (only for the data collection layer) we would go bankrupt very quickly. We charge 1$/million requests, and that includes our transformation and insertion layers, our salaries and more :-)

To sum up, up to a certain scale, there’s a huge benefit (in terms of headache and cash) to using a serverless architecture, but there comes a point where it just gets too expensive.