Storing data in a cloud is inexpensive and putting it in is usually free. Taking it out is charged, and the asymmetry shapes architectural decisions long after the original choice.
The underlying cost is real but uneven
Providers operate their own long-distance networks and buy connectivity to other networks, and both carry genuine cost. Traffic leaving toward the public internet uses capacity that must be provisioned and paid for.
Traffic between machines inside one data centre uses equipment the provider already owns and has already sized. The marginal cost of that traffic is close to nothing.
Pricing broadly follows this shape, with internal transfer cheapest, cross-region transfer more expensive, and traffic to the outside world charged at the highest rate.
Charges are asymmetric by design
Free inbound transfer lowers the barrier to moving workloads in, which is a reasonable commercial choice for any provider competing for new customers.
Once data is resident, applications are built around it, and moving it later means paying the outbound rate on the whole volume at once. The cost of leaving rises with every month of accumulation.
This is why the charge functions as a switching cost regardless of intent. Even where the underlying network expense is genuine, the pricing structure has an effect on mobility.
Architecture responds to the pricing
Systems get designed to keep processing next to storage, because moving computation to data is cheaper than moving data to computation. That constraint influences how services are split.
Teams add caching layers and content delivery in front of storage specifically to reduce repeat egress, so a cost line becomes a permanent architectural feature.
Multi-cloud designs face the same arithmetic. Any workload that regularly reads data held in another provider pays egress continuously rather than once.
Regulatory pressure changed some terms
Rules in several jurisdictions have pushed providers to waive charges for customers permanently leaving, on the reasoning that exit fees impede competition.
Those waivers generally apply to full migration rather than to ongoing operational transfer, so day-to-day costs for hybrid architectures are largely unaffected.
Estimating it in advance is the hard part
Egress is difficult to forecast because it depends on user behaviour rather than on provisioned capacity. A feature that becomes popular changes the bill without any infrastructure change.
Backup and replication traffic is a common surprise, since copies sent outside the provider are billed at outbound rates even though nobody thinks of them as user traffic.
The reliable approach is to measure transfer volumes in a pilot before committing an architecture, because the rate card alone does not reveal how much a given design will actually move.