Egress fees are charges applied when data leaves a cloud provider’s network. In practical terms, whenever you download files, move data outside the provider’s infrastructure, you may stack up egress costs.
You can think of egress frees as a “data exit tax.” While uploading data (ingress) is usually free, taking that data out is where providers charge.
Why Cloud Egress Fees Matter For GPU Workloads
GPU workloads are extremely data-intensive. Whether you're training large models, fine-tuning checkpoints, or serving inference at scale, data is constantly moving in and out of your environment.
This makes cloud egress fees especially important because:
- Model weights can be hundreds of GB
- Datasets are frequently transferred between environments
- Inference APIs send data externally on every request
- Teams often collaborate across regions or providers
Even small per-GB fees can compound into significant egress charges over time.
How Egress Charges Work Across Cloud Providers
Most major cloud providers follow a similar pattern: inbound data is free, outbound data is charged. However, pricing structures vary depending on destination, region, and volume.
| Provider | Egress fee | Naming* | Tiers |
|---|---|---|---|
| AWS (EC2) | ~ $0.09/GB | Data transfer | First 100 GB/month free. Discounts apply above 10 TB/month. |
| Google Cloud | ~ $0.12/GB | Egress pricing | First 200 GB/month free. Varies a by region. |
| Microsoft Azure | ~ $0.09/GB | Internet Egress | First 100 GB/month free. Discounts above 10 TB/month. |
| Vast.ai | $0-0.06/GB | Bandwith | Varies by host. Also has ingress fees. |
| Nebius | $0.015/GB | Storage - Egress traffic | Standard networking is Free. But download from object storage costs $0.015/GB |
AWS Egress Fees
AWS egress fees are tiered based on how much data you transfer out to the internet.
Typical structure:
- First 100 GB/month: Free
- Additional usage: Charged per GB
- Cross-region transfers: Additional fees
For GPU users exporting trained models or datasets, these costs can quickly scale into hundreds or thousands of dollars.
Azure Egress Fees
Azure egress fees also depend on where your data is going.
Key factors include:
- Data leaving Azure regions to the public internet
- Cross-region transfers within Azure
- Pricing differences between zones
Azure’s pricing can be harder to estimate, especially in multi-region deployments.
Google Cloud Network Egress Fees
Google Cloud network egress fees vary significantly based on destination:
- Same-region transfers: Often free or low cost
- Cross-region transfers: Charged per GB
- Internet egress: Highest cost tier
This makes distributed training or multi-region inference particularly expensive if not carefully managed.
Vast.ai Egress Fees
Vast.ai egress fees operate on a decentralized marketplace model rather than a fixed corporate rate.
- Host-determined pricing: Individual machine hosts set their own bandwidth prices, which typically range from free to $0.04 per GB.
- Inbound data costs: Unlike traditional clouds, some hosts on also charge ingress fees for uploading data to the instance.
- Risk of asymmetry: Renting a remarkably cheap GPU could be offset by a host charging higher-than-average rates to download your training datasets or export your model checkpoints.
Take a look at the real-world impact on this A100 80GB instance. At first glance, it's really cheap, but the data transfer fees are really high. For short workloads that need a lot of data moved into the instance, this can quickly get expensive.
| Uploaded data | Running time | Downloaded data | Actual cost per hour |
|---|---|---|---|
| 50GB | 1hr | 10GB | $3.35 |
| 50GB | 5hr | 10GB | $1.12 |
| 50GB | 10hr | 10GB | $0.84 |
Nebius Egress Fees
Nebius egress fees decouple standard network routing from backend data storage retrieval.
- Free network bandwidth: Standard internet ingress and egress traffic originating directly from your GPU virtual machines is completely free.
- Standard Object Storage ($0.015/GB): Reading data out of Standard Object Storage costs a flat fee per gigabyte.
- Enhanced Object Storage (Free): Upgrading to Enhanced storage eliminates retrieval fees entirely, offering zero-egress data processing at a higher monthly baseline storage cost.
Example: How Egress Fees Add Up
Let’s simulate a common scenario. You train a large model on a cloud GPU and need to:
- Download a 500 GB model checkpoint
- Export 1 TB of processed data
- Serve 500 GB of inference output
That’s 2 TB (2000 GB) of data leaving the cloud.
Here’s how that could look across providers:
| Provider | Total Data Egress | Cost Per GB | Estimated Cost |
|---|---|---|---|
| AWS | 2000 GB | $0.09 | $171.00 |
| Google Cloud | 2000 GB | $0.12 | $228.00 |
| Microsoft Azure | 2000 GB | $0.09 | $165.30 |
This is purely data transfer cost, your billing would also include GPU compute, storage, and other charges.
Now imagine doing this multiple times per month, or at larger scale. Costs can quickly spiral into thousands.
More importantly, if you want to migrate away from a provider, you must pay to extract your own data, reinforcing the lock-in effect.
How To Reduce Or Avoid Egress Charges
If you're using traditional cloud providers, you can reduce cloud egress fees by:
- Keeping workloads within a single region
- Minimizing unnecessary data transfers
- Using internal storage and networking where possible
- Caching frequently accessed data
However, these strategies add complexity and don’t eliminate the problem entirely.
The simplest solution is choosing a provider that doesn’t charge for egress at all.
Last Though on Egress Fees
Egress fees are one of the most overlooked costs in cloud GPU computing. While they may seem insignificant at first, they can quickly become a major expense for data-heavy workflows. By understanding how egress fees work you can make better decisions about your infrastructure.
If predictable pricing matters, choosing a platform without egress charges can simplify both budgeting and scaling. Thunder Compute’s transparent pricing model is designed to eliminate these hidden costs, so you can focus on building and deploying without surprises.
Learn how modern cloud providers structure their entire infrastructure model in what is GPU-as-a-Service.
FAQ
What are egress fees?
Egress fees are charges applied when data leaves a cloud provider’s network. In practical terms, whenever you download files or move data outside the provider’s infrastructure (such as downloading a 500 GB model checkpoint or exporting processed datasets), you may stack up these costs. While uploading data (ingress) is usually free, taking that data out acts as a "data exit tax."
Why are egress fees important for GPU workloads?
GPU workloads are extremely data-intensive. Model weights can be hundreds of GB, and datasets are frequently transferred between environments. Even small per-GB fees can compound into significant charges when training large models, fine-tuning checkpoints, or serving inference at scale.
What are the specific AWS egress fees for 2026?
AWS provides a Free Tier of 100 GB per month for data transfer out to the internet. Beyond that, the next 10 TB cost $0.09 per GB, while the next 40 TB drop to $0.085 per GB. Additional costs often include NAT Gateway processing fees at $0.045 per GB and Inter-AZ transfers at $0.01 per GB.
What are the Azure egress fees for 2026?
Azure egress fees start after a 100 GB monthly free grant. For the next 10 TB, the cost is $0.087 per GB when using the Microsoft Premium Global Network. If routing via a transit ISP, the cost is slightly lower at $0.08 per GB. Intra-continental transfers within North America or Europe are billed at $0.02 per GB, while transfers between Availability Zones cost $0.01 per GB.