This is not entirely true. Firstly because Lambda keeps getting cheaper with the move from per 100ms billing to pwe 1ms being a big contributor. but also that I see organizations handling millions of requests per hour at peak but because of the dynamic nature of Lambda, any cost increase during the peaks are entirely offset by the reduced cost during low traffic periods since Lambda and associated services cost $0 when its doing nothing. Add in other services that act the same way such as S3, DynamoDB, API Gateway, etc and you have a very powerful, redundant and scalable applications without even really trying. No to mention that because spinning things out is so quick you can get a feature in front of users earning revenue way faster than any traditional service that requires some effort to start up, configure and maintain. And the costs associated with having VM's of the right size up to handle peak loads that are much slower to ramp up and down, creating significant periods of over provisioned capacity you pay for.
Lambdas are not slow to boot because of the methods used to tune how hot they stay; in a production application doing millions of invocations a day we saw 0.03 % of all invocations suffering a cold start that added an additional 100ms latency only. And if you absolutly have to have warm Lambdas always for even that 0.03% you can use provisioned concurrency that always keeps a certain number of Lambda functions warm.
