If you've spent any time reading hyperscaler bills, you know egress is the line that ruins forecasting. Compute is predictable. Storage is predictable. Bandwidth charges are not, because the price changes depending on where you send the bytes, who you peer with, and whether the moon is full. This is the practical guide.
The basic prices
As of 2026, the public-internet egress prices for the big three look like this:
- AWS: $0.09/GB for the first 10 TB/month, dropping in tiers. Effective rate at small-to-medium scale: $0.07–$0.09/GB.
- GCP: $0.12/GB out of premium tier, $0.085/GB out of standard tier.
- Azure: $0.087/GB after the first 100 GB/month, with tiering similar to AWS.
Inter-region (inside the same cloud) is roughly half price. Traffic to private-link, Direct Connect or ExpressRoute partners is materially cheaper. Public internet is the headline rate above.
In real numbers:
- 1 TB egress per month: roughly €80–€110.
- 10 TB egress per month: roughly €700–€900.
- 100 TB egress per month: roughly €5,000–€7,000.
These numbers compound. Compute might be €2,000/month, storage €500/month, egress €5,000/month. The cloud's "obvious" cost ends up being a third of your bill.
Where egress shows up
The naive picture is "user pulls a file, you pay for the file size". The real picture is much wider.
Inference responses. Every token streamed back to a client is billable egress.
Model checkpoint distribution. Pulling a 30 GB checkpoint to your laptop or to a different cloud is 30 GB of egress.
Backups to your own object storage at another vendor. If you back up to a different cloud than the one you're computing on, you pay egress on every backup.
Cross-region replication. Replicating to another region "for resilience" is paid bandwidth.
Data-pipeline source pulls. Some datasets you pull from elsewhere are charged at the source, not the destination. If you then process them and ship the result to another region, that is also egress.
Customer-facing artefacts. Generated images, audio, video. Anything you produce and serve back to a user is egress.
The teams that get caught are not the ones whose bandwidth is tiny. They are the ones whose product happens to involve moving GBs of generated content per user-month, which compounds fast.
Why hyperscalers price it that way
Egress is the one line item where hyperscalers have actual pricing power. The compute price floor is set by chip ASPs and competition. The storage price floor is set by storage hardware costs. Egress doesn't have a comparable floor. Bandwidth at scale costs the cloud single-digit cents per GB, and they charge an order of magnitude more.
It is also a moat. Once your data is in AWS S3, the cost of pulling it out to evaluate a competitor is itself a barrier to switching. The asymmetry (free in, expensive out) is intentional. It is known internally as "data gravity" pricing.
How EU GPU specialists handle it
There are three patterns.
Free egress on all traffic. Some specialists (Clodei included) do not charge per GB at all. The GPU price absorbs the bandwidth.
Free intra-EU, paid outside. Anything that stays in the EU is free, US-bound traffic is metered.
Per-GB pricing at lower rates. Half to a third of hyperscaler rates.
Specialists can do this because their margin model is the GPU itself. The bandwidth costs they actually pay (peering, transit) are a few cents per GB at the volumes they handle, so absorbing it into the GPU price does not move the per-hour rate by more than 10%.
The questions to actually ask a vendor
When comparing GPU vendors, the headline €/h is incomplete without:
What is the egress price per GB to the public internet? If the answer involves a calculator, that is the answer.
Is there a separate price for traffic between your regions? Common gotcha. "Same cloud" does not always mean "free".
What about pulling object storage to a different cloud? This kills multi-cloud setups.
Are there free tiers for egress? Many hyperscalers give you a free 100 GB/month. Useful for tiny workloads, irrelevant once you scale.
Is bandwidth metered per second or per month? Some specialists rate-limit instead of metering, which is useful or not depending on your workload.
Modelling egress before you commit
A short exercise that will save you a lot of money:
- Estimate your monthly bandwidth usage. Be generous. Most teams underestimate.
- Multiply by the headline egress rate of each vendor you're considering.
- Add that number to the headline GPU price.
- Compare that combined number, not the GPU price alone.
For an inference workload that streams 5 KB tokens per request and does 10 million requests/month, that is 50 GB. Trivial. For a SaaS that ships a 5 MB generated PDF per user-action and does 1 million actions/month, that is 5 TB. About €450 you did not have on your cost sheet.
When it doesn't matter
If your workload is genuinely closed-loop, meaning you train, store and serve all in the same cloud, the egress line is small. Hyperscalers are still expensive on a per-GPU-hour basis but you avoid the bandwidth tax.
The teams that get hit hardest are the ones with users on the public internet, multi-cloud setups, or backups to a different vendor. Those are the workloads where "free egress" is not a marketing claim. It is a 20–40% cut on the bill.
What we ship
Clodei does not bill egress, period. The €/h on the plan card is the €/h you pay. Bandwidth comes out of our margin, not yours. We can do that because GPU compute is the line item we are competing on. Bandwidth is a fixed cost we would rather absorb than turn into a calculator-induced friction at signup.
If you do the egress modelling exercise above and the number is large, that is the strongest argument for an EU specialist over a hyperscaler that no marketing page makes loudly enough.