Data transfer costs: the line item that ruins cloud budgets

Compute prices across providers differ by maybe a factor of three. Cloud egress costs differ by a factor of a hundred or more. For any workload that moves real volumes of data to users, that single line decides where you should be hosting, and no amount of instance right sizing will change it.

Why cloud egress costs vary so widely

Bandwidth is not expensive to provide at scale. The hyperscalers price egress well above cost because it is a reliable margin source and because it discourages moving data out of the platform. The European operators price it close to cost because bandwidth is what they sell.

The result is that a workload serving twenty terabytes a month can be a rounding error at one provider and a five figure annual line at another, on otherwise similar infrastructure.

What actually gets billed

Inbound is free everywhere. The billed categories are:

• Internet egress. Data leaving the provider to the public internet. The headline number, usually tiered so the rate falls with volume.
• Inter region transfer. Data moving between regions of the same provider. Billed, and frequently overlooked in multi region designs.
• Cross zone transfer. Data moving between availability zones within a region. Small per gigabyte, and significant for chatty microservices that were placed for resilience without thinking about the traffic.
• NAT gateway processing. On AWS, every gigabyte through a NAT gateway carries a processing charge on top of the transfer itself.
• CDN egress. Usually cheaper than direct egress, and the reason a CDN often pays for itself immediately.

Traffic between resources in the same zone on private addresses is generally free, which is a strong argument for keeping internal service communication on the private network.

The rough shape across providers

Without quoting figures that will be stale by the time you read them, the ordering has been stable for years:

• Most expensive. AWS, Azure and Google Cloud, at broadly similar rates with tiered reductions at volume.
• Substantially cheaper. Oracle Cloud, which includes a large monthly allowance and prices the overage well below the other three. This is a genuine structural difference.
• Cheaper again. The developer clouds. DigitalOcean and Linode pool a per instance allowance across the account and charge modestly beyond it.
• Cheapest. Hetzner and OVHcloud. Hetzner includes a very large allowance per server in European locations with inexpensive overage; OVHcloud includes generous or unmetered traffic on most European products, priced by port speed.

Always check current rates for your specific region, because the tiers and the regional variations are real.

Reducing the number

1. Put a CDN in front. Cached responses never reach your origin, and CDN egress is usually cheaper than direct egress. For static assets this is the single biggest lever.
2. Compress everything. Brotli or gzip on text, modern formats on images, sensible bitrates on video. A thirty percent reduction in bytes is a thirty percent reduction in the bill.
3. Keep internal traffic private. Service to service calls over public addresses are billed as egress and are also worse for security.
4. Watch cross zone chatter. Spreading a chatty service across three zones for resilience is correct, and it has a traffic cost worth knowing about.
5. Use VPC endpoints. On AWS, routing traffic to S3 and other services through gateway endpoints avoids NAT gateway processing charges entirely. See AWS VPC networking basics.
6. Host bulk data where bandwidth is cheap. There is no rule saying the whole system lives with one provider. Media and downloads can sit somewhere with generous included traffic while the application stays where it is.

When egress should decide the provider

If outbound volume is a meaningful fraction of your infrastructure profile, it should be the first factor rather than the last. Media hosting, file distribution, game asset delivery, backup targets and anything with large downloads all fall into this category.

For those workloads, the European operators are not merely cheaper, they are a different order of magnitude. The Hetzner guide and the OVHcloud guide cover what you give up in exchange, which is mostly managed services rather than reliability.

If the workload is an ordinary application with modest traffic, egress will not decide anything and you should choose on other grounds. Cloud pricing models compared covers what does move the number in that case.