Stop arguing ideology — score workloads
There is no universal answer. Score each workload across four axes: cost predictability, data sovereignty, latency tolerance, and team operational maturity.
Workloads that score poorly on more than two axes for one model belong in the other.
When cloud wins
Spiky or unpredictable demand, global distribution, fast time-to-market, managed services (queues, ML, analytics), and small ops teams.
If you re-architect for cloud-native (auto-scaling, managed databases, serverless), the TCO advantage is real.
When on-premise wins
Steady-state workloads with predictable utilisation, very large data volumes with high egress costs, low-latency control systems, and strict data residency.
A well-run virtualised cluster often beats cloud TCO for workloads above 70% steady utilisation.
Hybrid is the realistic default
Most mid-market companies end up hybrid: core ERP/file/DB on-prem, collaboration in Microsoft 365, dev/test in the cloud, DR replicated to a sovereign region.
Hybrid only works with a unified identity, network and observability layer.
The real cost of cloud
Always model 3-year TCO including egress, observability, support tiers, and the engineers needed to operate it. Lift-and-shift without refactoring is almost always more expensive than on-prem.
Conclusion
Make the choice workload by workload, not company-wide. Revisit it every 18 months as pricing, regulation and your own scale evolve.
