Work out scale out vs scale up instantly with clear inputs, formula shown and shareable results.
When pricing is linear in size, scale-out and scale-up cost the same per unit of capacity, so the decision comes down to granularity and failure domain. Large instances round up more coarsely, wasting capacity, and each one takes a bigger share of the fleet with it when it fails. Scale-up wins where per-instance licensing, memory locality or single-threaded performance dominate.
Cost comparison
scale out = ceil(required / small units) x small cost; scale up = ceil(required / (small units x multiplier)) x small cost x multiplier
Scale out, generally. Smaller failure domains, finer autoscaling granularity and easier rolling updates all favour more, smaller instances.
Per-socket or per-instance licensing, large in-memory working sets that must not be partitioned, and workloads that cannot be parallelised across processes.