Research-backed calculator guide

GPU Cloud Egress Fees: Find the Data-Transfer Break-Even

Compare GPU cloud egress allowances and rates, calculate checkpoint-export cost, and find the data-transfer volume that cancels out a cheaper GPU hourly price.

What GPU cloud egress includes

Egress is data leaving a provider’s network. In a GPU workflow, that can mean a model checkpoint copied to your laptop, a dataset moved to another cloud, an inference response delivered to users, or a trained model exported to a serving platform. Ingress is data entering the cloud, but “free ingress” does not make the reverse path free.

Keep the network path explicit. Internet egress, movement between cloud regions, movement between availability zones, and a read from object storage can be different meters. A VM may have one network price while a bucket has a retrieval price. A managed NAT gateway can add processing cost even when the destination transfer rate looks low. A GPU marketplace may bundle bandwidth, cap it, or state terms per host rather than per account.

This matters for model checkpoint egress cost. A finished model is not necessarily the same size as a resumable training checkpoint. A checkpoint may contain optimizer state, gradients, metadata, and multiple precision copies. If you export every save instead of retaining checkpoints beside the GPUs, the export cadence can dominate the bill.

How to compare GPU cloud egress fairly

Use the following sequence for each provider. It is deliberately more conservative than copying one per-GB number into a spreadsheet.

1. Name the traffic. Is the data going to the public internet, another provider, another region, a CDN, or a service in the same region? Use the provider’s meter for that path.
2. Choose one unit. Keep GB and GiB visible. A decimal GB is 1,000,000,000 bytes; a GiB is 1,073,741,824 bytes. Do not silently compare a $/GB quote with a $/GiB quote as if they were identical.
3. Record the allowance scope. It may be account-wide, service-wide, region-wide, per instance, or included only in a plan. AWS says its 100 GB monthly internet allowance is aggregated across eligible AWS services and regions, with stated exceptions. ^[1]
4. Apply tiers in order. A provider can lower the rate after 10 TB or another threshold. Multiplying all traffic by the cheapest tier overstates the saving at smaller volumes and understates the bill at larger volumes.
5. Add non-egress meters. Include object-storage retrieval, request charges, NAT processing, inter-region transfer, and any private-link or CDN fees when they apply.
6. Repeat for the traffic pattern. One 2 TB export, 20 daily 100 GB exports, and 2,000 one-GB responses can have different operational implications even when the monthly total is similar.

Google Cloud’s published Compute Engine network pricing illustrates why the unit and destination matter: the page presents internet transfer rates per GiB and separates Premium and Standard Tier pricing, with destination-dependent bands and a monthly free allowance for the relevant service. ^[2] Azure likewise shows a first-100-GB allowance, separate inter-region prices, and different internet-egress tables for Microsoft’s Premium Global Network versus a transit-ISP routing preference. ^[3]

The core formulas

For a single rate tier:

estimated egress = max(0, outbound volume - free allowance) × rate

For multiple tiers:

estimated egress = Σ (billable units in tier × tier rate)

For a provider that is cheaper on GPU time but more expensive on egress:

egress break-even volume = compute saving ÷ net egress rate

Then add the allowance back if you want the total outbound volume at which the saving disappears:

total outbound break-even = free allowance + egress break-even volume

These are planning equations, not invoice reconciliation. Round only after calculating. If a provider bills in GiB, convert your workload volume to GiB before applying the provider’s rate.

Worked cost example: 2,000 units out in one month

Assume a North America workload exports 2,000 units during a 30-day month. “Unit” is intentionally labeled because providers do not all use the same binary or decimal convention. The traffic is internet egress from a compute workload. We assume the first paid tier and exclude storage retrieval, NAT, CDN, inter-region, taxes, support, negotiated discounts, and marketplace-specific bandwidth terms.

The table is useful because every output is inspectable. It is not a universal provider ranking. The AWS row uses a clearly labeled planning assumption for a common first paid tier; the AWS page confirms the aggregate allowance and tiered structure, but a real bill depends on the specific region, service, and rate tier. ^[1] The Google row keeps GiB because Google publishes this meter in GiB. The Azure row uses the North America/Europe Premium Global Network first paid tier and its published decimal-GB convention. ^{[2] [3]}

Now test a checkpoint pattern instead of a one-off export. Suppose a 140 GB finished model is exported once per day for 30 days. Monthly outbound volume is 140 × 30 = 4,200 GB. At the same AWS-style planning rate and 100 GB allowance, the estimate is (4,200 − 100) × $0.09 = $369. If a resumable checkpoint is 840 GB and exported once per week, using 4.3 weeks as a planning month gives 840 × 4.3 = 3,612 GB, or roughly $316.08 after the same allowance. If that 840 GB checkpoint is exported every day, the monthly volume becomes 25,200 GB, and a tiered schedule—not the first-tier rate multiplied across every byte—must be used.

This is where a GPU cloud data transfer out cost can overwhelm compute. Keeping the full checkpoint beside the GPU and exporting only deployable weights changes the volume, not just the provider. Before choosing a low-hourly-rate GPU, write down the number of exports, the size of each export, and whether a retained checkpoint crosses a provider boundary.

Find the data-transfer break-even

Suppose Provider A charges $2.00 per GPU hour and Provider B charges $1.60 per GPU hour. Provider B saves $0.40 per hour. At 500 GPU hours per month:

monthly compute saving = ($2.00 − $1.60) × 500 = $200

Assume Provider B charges $0.09 per paid GB for the relevant outbound path and includes 100 GB. The paid egress volume that consumes the saving is:

$200 ÷ $0.09 = 2,222.22 GB

The total outbound break-even is approximately:

2,222.22 + 100 = 2,322.22 GB

Below roughly 2.32 TB outbound, the hourly saving is larger than the modeled egress charge. Above it, egress alone consumes the saving. This does not mean Provider B is worse. It means the decision depends on where data lives, how often it leaves, and whether the workload can keep storage and serving on the same platform.

For a response-serving workload, estimate output volume from requests: requests × average response bytes. For training, estimate dataset reads and checkpoint exports separately. For migration, model the one-time transfer and the recurring traffic that remains after cutover. If the provider’s free allowance is pooled, apply it once at the correct scope rather than once per GPU.

Use the Self-hosted GPU inference cost calculator when you need to combine GPU time, utilization, storage, and serving assumptions. The egress worksheet here is a network layer, not a replacement for a full inference TCO model. For an application that also retrieves documents or writes embeddings, compare that network estimate with the RAG, embedding and vector database cost calculator. If the GPU is one component of a larger product build, the AI agent development cost estimator helps keep implementation and operating costs separate.

Provider free-egress comparisons: what to verify

“Free egress” is not a complete comparison field. Ask what is free, where it applies, and what happens after the cap.

• Allowance scope: Is the allowance per account, per project, per VM, per GPU, or per month across a service family?
• Destination: Does “free” cover the public internet, only traffic inside the provider, or only a storage product’s download path?
• Capacity limit: A provider may advertise free transfer while applying a bandwidth cap, fair-use policy, or throttling after a threshold.
• Service boundary: GPU VM traffic and object-storage retrieval can be billed by different products.
• Region: Rates vary by source region, destination, route, and network tier. Azure’s table is a visible example of this regional and routing split. ^[3]
• Contract treatment: Enterprise agreements, credits, and migration programs can change a list-price result. Treat a pricing page as a list-price reference, not a quote.

A good request for written confirmation is: “For this GPU instance or endpoint, in this region, how much internet data transfer out is included each month; is the allowance pooled; what is the rate by tier; and are storage reads, cross-region traffic, NAT, or bandwidth caps separate?” Save the answer with the date and the product SKU.

Assumptions, exclusions, and limitations

The comparison above is intentionally narrow. It estimates internet egress only. It does not include GPU runtime, persistent disks, object-storage capacity, storage retrieval, request operations, NAT gateway processing, load balancers, CDN transfer, private connectivity, inter-region transfer, taxes, credits, support, or negotiated pricing. It also does not assume that a marketplace’s “no egress fee” statement means unlimited throughput or identical treatment for every storage product.

Rates and product terms change. The primary pricing pages were checked on November 2, 2026, but a planning workbook should retain the source URL, checked date, region, service, unit, allowance, and tier used. AWS notes that its aggregate internet-transfer tiers cover a defined set of services and regions; Google Cloud publishes separate network and CDN paths; Azure distinguishes internet and inter-region transfer. ^{[1] [2] [3]}

The result is an estimate, not a vendor quote. If the decision is material, run the exact workload through the provider’s calculator or billing export and ask sales or support to confirm the path. Never compare an hourly GPU number from one region with an egress number from another and call the result a total cost.

FAQ

What is GPU cloud egress?

GPU cloud egress is data leaving a GPU provider’s network. It includes a checkpoint downloaded to a workstation, a dataset copied to another cloud, a model exported for serving, or inference responses sent to users. The applicable charge depends on the destination, region, service, unit, allowance, and tier. Ingress into the provider is a different direction and does not automatically make outbound transfer free.

How do I calculate cloud GPU egress fees per TB?

Start with the provider’s unit. For a single tier, subtract the included allowance from the outbound volume and multiply the remaining units by the rate. If the provider quotes $0.09 per decimal GB and you send 2,000 GB with 100 GB free, the estimate is (2,000 − 100) × $0.09 = $171. For a tiered schedule, split the billable volume across tiers. Do not use “per TB” as a substitute for checking whether the vendor means 1,000 GB, 1,024 GB, or a binary GiB.

Is a GPU provider with free egress always cheaper?

No. Free egress can be valuable, but it is only one line in the total cost. Compare GPU hourly rate, expected utilization, storage, bandwidth limits, storage retrieval, region, reliability, and the exact destination covered by the free policy. A free internet-egress allowance may not cover object-storage reads or cross-region traffic. Ask the provider to confirm the complete path in writing.

What is the model checkpoint break-even?

The checkpoint break-even is the export volume at which a cheaper GPU rate no longer offsets its network cost. Calculate monthly compute saving first: (alternative hourly rate − chosen hourly rate) × GPU hours. Divide that saving by the applicable paid egress rate, then add the included allowance if you want total outbound volume. Repeat for the actual checkpoint size and export cadence. A resumable checkpoint can be much larger than the finished weights, so exporting every save can move the break-even sharply.

Does keeping the checkpoint beside the GPU avoid egress?

It can avoid an external network-egress event while the checkpoint stays within the provider’s storage and compute boundary, but it does not prove that storage is free. Persistent storage, object-storage operations, retrieval, and later export can still be billed. Confirm the provider’s product-specific terms and whether the storage is in the same region and service boundary as the GPU.

Should I include inference responses in the comparison?

Yes, if responses leave the provider. Estimate request count × average response bytes, then add streaming overhead and any retries or duplicated delivery. If a CDN, gateway, or load balancer sits in the path, model its charges separately. A small response can make egress immaterial; large generated files, audio, video, or retrieved documents can make it significant.