A company evaluating whether to move a workload off public cloud computes a lower monthly infrastructure bill for owned hardware. Which costs does that comparison typically omit?

What the naive comparison omits

1. People. Data-centre operations, hardware supply chain, capacity planning, physical security, network engineering — permanent functions that do not appear in an infrastructure spreadsheet and do appear in the budget. This is usually the largest omitted cost.

2. Utilisation. Owned capacity is sized for peak plus headroom and idle the rest of the time. Cloud is billed for what runs. A comparison of peak owned capacity against average cloud consumption flatters ownership substantially.

3. Lead time as a cost. Capacity acquisition measured in months means growth must be forecast, and a forecast miss is either an outage or expensive idle hardware. Optionality has value that does not appear as a line item.

4. Refresh and depreciation. Hardware has a life. The comparison must include replacement on a cycle, not just the initial purchase.

5. Redundancy and disaster recovery. A second site with its own capacity, or a cloud arrangement for it.

6. Managed service replacement. The cloud bill included managed databases, queues, load balancers, identity and monitoring. Each must now be operated, which is engineering time plus licence cost plus the operational risk of running it less well than a specialist.

7. Opportunity cost. Every engineer working on infrastructure is not working on the product.

What the comparison sometimes omits in the other direction

To be fair, the naive comparison also understates ownership benefits: egress pricing for bandwidth-dominated workloads does not improve with scale, and owning capacity changes the gradient of the cost curve rather than just its intercept. For a workload where storage or bandwidth is the dominant and growing cost, that gradient change can be decisive.

The conditions that make repatriation right

All of these, not a majority: the cost is dominant and growing; the workload is uniform enough that custom optimisation pays; scale is large enough that a few percent matters in absolute terms; growth is predictable enough to forecast; and the capability is core to the product rather than incidental.

Three yeses justify vertical integration. Two or fewer, and you are buying a distraction with a discount.

The framing

Total cost of ownership is a comparison of two operating models, not two bills. A comparison that produces a number without naming the headcount, the utilisation assumption and the refresh cycle is not a TCO analysis — it is an infrastructure quote.