"The cloud is just somebody else's computer" is technically true and practically incomplete.
The useful part of cloud computing is precisely that somebody else owns a great deal of difficult infrastructure and exposes it through services you can consume. Sometimes that is an excellent trade, and we would rather say so than pretend otherwise.
You are buying capabilities, not virtual metal
A cloud virtual machine is easy to compare with a conventional server, because it is the same shape.
Managed databases, object storage, queues, global content delivery, serverless functions and machine-learning services are a different proposition. They remove operating work as well as hardware ownership.
If a managed service eliminates an entire subsystem your team would otherwise have to build, run and be woken up by, the premium can be extremely cheap.
Elasticity is real when you need it
Some workloads genuinely change by enormous amounts. Ticket launches, public events, batch processing and seasonal systems can have demand patterns where buying permanent capacity for the peak makes no sense at all.
Being able to take resources for a few hours and then give them back is worth paying for.
Geography can be difficult to reproduce yourself
Large platforms provide regions and edge networks around the world. If an application has to serve users close to many locations, replicate across continents, or satisfy a specific regional requirement, reproducing that footprint independently is usually impractical.
The trade is control for convenience
Managed services reduce the number of layers you operate. They also increase your dependence on somebody else's implementation, pricing and limits.
That can be a perfectly sensible trade. It should simply be recognised as one, because the more proprietary the service, the more expensive changing direction becomes later.
Use cloud because it solves a problem
We operate dedicated infrastructure because it fits the long-lived, predictable workloads we mostly host. We also use cloud services where they are the better tool, and we would rather explain that than defend a position.
There is no architectural virtue in avoiding a hyperscaler, just as there is none in putting every database on one. "Cloud first" and "cloud never" are both shortcuts around the same question, which is what you are buying with the abstraction.