AI data egress charges
There is no special AI rate. Training pulls, model transfer, and inference streaming all bill at the same per-GB cloud egress every other workload pays. What changes is the shape of the traffic, which is why AI teams get surprised.
How much are AI data egress charges?
AI workloads pay standard cloud egress: $0.09/GB on AWS for the first 10 TB (after 100 GB free), $0.12/GB on GCP Premium Tier, $0.087/GB on Azure Zone 1. There is no separate ML meter. The cost hides in three flows: pulling training data in (free ingress), moving datasets and checkpoints across regions or clouds ($0.02/GB cross-region on AWS, $0.09/GB out to another cloud), and streaming inference responses out at roughly 6 bytes per token.
The largest avoidable charge is re-pulling model checkpoints: a 1 TB checkpoint from your own S3 bucket costs about $90 in egress every fresh instance. Hosting weights on Cloudflare R2 (zero egress) removes it entirely.
Rates from published AWS / GCP / Azure / Cloudflare pricing, reviewed June 2026
[01]The three AI egress surfaces
Every AI egress bill is one of these three data flows. Two of them are avoidable with the right region and storage choices.
Training data pull
Pulling public datasets (Common Crawl, ImageNet, HuggingFace) into your cloud is ingress, which is free. Cost only appears on cross-region or cross-cloud re-egress.
Read more →Model & checkpoint transfer
Re-pulling a 1 TB checkpoint from your own S3 bucket to a fresh instance costs standard egress each time. The biggest avoidable AI egress bill. Zero on Cloudflare R2.
Read more →Inference response streaming
Every output token egresses ~6 bytes. Trivial for text at billions of tokens; grows fast for image and audio generation where each response is megabytes.
Read more →What AI egress actually costs
per month · AWS list ratesText inference egress is trivial; dataset and checkpoint movement is where the money goes. Train where the data lives, host re-pulled artefacts on zero-egress storage.
Model your own AI egress, or compare providers
AI egress is standard cloud egress, so the same calculator and comparison apply. Run your monthly transfer-out volume across AWS, Azure, GCP, and the zero-egress alternatives.
[02]Frequently asked
Q.01How much are AI data egress charges?
›
There is no separate AI or ML egress meter. AI workloads pay the same per-GB cloud egress as anything else: AWS charges $0.09/GB for the first 10 TB per month (after 100 GB free), GCP $0.12/GB at Premium Tier, and Azure $0.087/GB in Zone 1. AI-specific cost shows up in three data flows: pulling training datasets in (ingress, which is free on all clouds), moving datasets or model checkpoints across regions or clouds (standard egress, $0.02/GB cross-region on AWS or $0.09/GB out to another cloud), and streaming inference responses to users (egress at roughly 6 bytes per output token).
Q.02Does pulling a training dataset into my cloud cost egress?
›
No. Data transfer into a cloud (ingress) is free on AWS, Azure, and GCP. Pulling Common Crawl, ImageNet, The Pile, or a HuggingFace dataset into your training cluster costs you nothing in ingress. The cost appears only if you then re-egress that data: reading an S3 dataset from a compute instance in a different region ($0.02/GB on AWS) or copying it out to a different cloud (standard internet egress, $0.09/GB on AWS). The rule of thumb is to run your GPU cluster in the region where the dataset already lives.
Q.03How much does LLM inference egress cost?
›
Each output token leaves your cloud as roughly 6 bytes (4 bytes UTF-8 plus about 2 bytes of SSE/JSON framing), so a 1,000-token response is about 6 KB. Serving 1 billion output tokens per month egresses about 5.6 GB, which is under the 100 GB AWS free allowance and effectively $0. At 100 billion tokens per month (about 559 GB), you pay for 459 GB after the free tier, roughly $41 at AWS's $0.09/GB. Inference egress is small relative to compute for text; it grows fast for image and audio generation, where each response is megabytes not kilobytes.
Q.04Why is moving model checkpoints so expensive?
›
Model weights are large and re-pulled often. A 1 TB checkpoint pulled from your own S3 bucket to a fresh training instance costs about $90 in egress every time if the instance is out of region or out of cloud. Teams that spin up fresh instances weekly can pay hundreds of dollars a month just re-pulling weights. Hosting weights on HuggingFace (backed by Cloudflare R2, which has zero egress fees) or on Cloudflare R2 directly removes that charge entirely. This is the single biggest avoidable AI egress cost.
Q.05Is there a zero-egress option for AI workloads?
›
Yes. Cloudflare R2 charges $0 for egress at every volume, and Backblaze B2 charges $0.01/GB with 3x-storage free egress. HuggingFace serves its models and datasets from R2, so downloads are free at source. For any AI artefact your team re-pulls repeatedly (model weights, curated datasets, checkpoints), hosting it on a zero-egress store instead of an S3 or GCS bucket eliminates the per-pull charge. Storage is slightly more expensive on some of these, but for high-read AI assets the egress saving dominates.