Cloud Egress Costs: The Line Item That Decides Your Provider

Cloud storage is cheap. Cloud compute is competitive. Moving data out of a cloud is expensive, and for any workload serving significant volume, egress is the number that decides which provider you should be on.

The pricing shapes

There are three broad models:

Per-gigabyte with a small free allowance. The hyperscaler model. A modest monthly free tier, then per-gigabyte charges that decrease in tiers as volume rises. Rates vary by region and destination.

Large included allowance, then metered. The VPS model. Each instance includes a substantial monthly allowance, often pooled across the account, with modest overage rates.

Unmetered on many configurations. OVHcloud's position on much of its range, subject to fair use.

The practical gap is large. A workload shifting several terabytes a month can cost a meaningful amount in hyperscaler egress and effectively nothing on a provider with a generous allowance. For media delivery, backup targets, file distribution or large API responses, that difference frequently exceeds the entire compute bill.

The internal charges nobody plans for

Egress to the internet is the obvious cost. These are the ones that appear unannounced:

• Cross-availability-zone transfer. Billed per gigabyte in each direction on several platforms. A microservice architecture spread across three zones pays for every internal hop.
• NAT gateway processing. Charged per gigabyte in addition to an hourly rate. Private subnets pulling container images or OS packages through NAT generate real volume.
• Cross-region replication. Priced as egress. Database replicas in a second region bill continuously.
• Load balancer processed bytes. Part of the capacity-unit model on the hyperscalers.
• Egress to your own on-premises network over a direct connection, which is cheaper than internet egress but not free.

VPC endpoints or private service connections can remove a large share of NAT charges by keeping traffic to managed services off the gateway. On storage-heavy workloads this usually pays for itself immediately.

Modelling it before you commit

The calculation is not difficult and is skipped almost universally:

1. Estimate monthly outbound volume. Average response size multiplied by monthly requests, plus media and downloads, plus backup transfers leaving the provider.
2. Separate CDN-cacheable from origin traffic. Static assets should be served by a CDN; only cache misses reach your origin.
3. Add internal transfer — cross-zone, NAT-processed and replication volume.
4. Apply each candidate provider's rates to the resulting figures.
5. Compare total cost, not compute cost.

Doing this before choosing a provider takes an hour and occasionally reverses the decision entirely.

Reducing egress

In rough order of effect:

1. Put a CDN in front of everything cacheable. CDN egress is cheaper than origin egress almost everywhere, and cache hits never touch your infrastructure.
2. Compress responses. Brotli or gzip on text, JSON and HTML typically removes 60–80% of the bytes.
3. Serve modern image formats. WebP and AVIF at appropriate dimensions, rather than full-size originals scaled in the browser.
4. Cache aggressively with correct headers. Long max-age on immutable assets with content-hashed filenames.
5. Keep chatty services in the same zone. Cross-zone hops are avoidable architecture, not a law of nature.
6. Use VPC endpoints for managed-service traffic.
7. Paginate and trim API responses. Returning fields nobody reads is pure egress.

Where provider choice genuinely changes the answer

Some workloads are egress-dominated by nature:

• Video and audio delivery
• Software, game and asset distribution
• Public file hosting and backup targets
• Image-heavy catalogues without a CDN
• Data export and reporting pipelines

For these, a provider with a large included allowance — a Hetzner account, a DigitalOcean account — or with unmetered configurations such as an OVHcloud account is not a marginal saving. It changes the business model of the service.

For workloads that are compute-heavy and egress-light — batch processing, internal tooling, APIs with small responses — egress barely registers and you should choose on other criteria.

Frequently asked questions

Is inbound data charged? Generally no. Inbound is free on essentially every major provider, which is why upload-heavy workloads are cheap to host.

Does traffic between instances cost anything? Within the same private network and zone, usually not. Across availability zones, often yes. Across regions, always.

Do CDN costs replace egress costs? Partly. You pay CDN egress instead of origin egress, usually at a lower rate, and cache hits never reach your origin at all. The net is almost always favourable.

Can I negotiate egress rates? At large volumes, yes — with the hyperscalers this is a standard part of enterprise agreements.

The comparison page lists the providers in the catalogue so you can weigh bandwidth policy alongside compute price.