You open Cost Explorer, filter on EC2-Other, and there it is: NAT Gateway, sometimes larger than the instances behind it. It's one of the most common surprises on an AWS bill - and one of the easiest to fix once you know where the traffic comes from.
How NAT Gateway pricing works
A NAT gateway lets resources in private subnets reach the internet (and public AWS endpoints) without being reachable from outside. You pay for it in two ways:
| Charge | US East (N. Virginia) price | What drives it |
|---|---|---|
| Hourly charge | $0.045 per hour | Every hour the gateway exists - per gateway, or per Availability Zone for a regional NAT gateway. |
| Data processing | $0.045 per GB | Every GB that goes through the gateway, in either direction, whatever the destination. |
| Data transfer | standard rates | Charged on top: internet egress, and cross-AZ traffic when instances use a NAT gateway in another zone. |
Prices from the Amazon VPC pricing page for US East (N. Virginia) as of October 2026. Other regions are priced differently.
The hourly part is predictable. The data processing part is what grows silently: it applies to all traffic through the gateway - including traffic to AWS services such as Amazon S3, Amazon ECR or CloudWatch Logs, which often don't need to go through it at all.
A worked example
Take a typical production VPC with two Availability Zones and one NAT gateway in each, processing 5 TB per month:
| Line | Calculation | Monthly cost |
|---|---|---|
| Hourly charge | 2 gateways × 730 h × $0.045 | $65.70 |
| Data processing | 5,120 GB × $0.045 | $230.40 |
| Total NAT gateway | $296.10 |
If 70% of that traffic is actually going to Amazon S3 - backups, logs, data pipelines, container image layers - a free gateway endpoint (see fix #1) removes about $161 per month of processing charges from this single VPC. Multiply by your number of VPCs and environments.
First, find out what goes through your NAT gateway
- Cost Explorer. Group by Usage type and look for lines ending in
NatGateway-HoursandNatGateway-Bytes. The second one is your per-GB bill. - CloudWatch metrics. Each NAT gateway publishes
BytesOutToDestinationandBytesInFromDestination. Sort your gateways by volume to see which VPC and environment to look at first. - VPC Flow Logs. Enable flow logs on the NAT gateway's network interface, then query them (CloudWatch Logs Insights or Amazon Athena) to rank destination IP addresses by bytes. Matching those IPs against AWS's published IP ranges tells you which traffic goes to S3, ECR or other AWS services.
Once you know the top destinations, the fixes below become obvious.
7 ways to reduce your NAT gateway costs
1. Add gateway endpoints for S3 and DynamoDB - they're free
Gateway VPC endpoints for Amazon S3 and Amazon DynamoDB have no hourly and no data processing charges. Once added to your route tables, traffic to those services stops going through the NAT gateway. It's usually the single biggest win, it takes minutes, and it doesn't change your application code. Check bucket policies that restrict access by source IP: with an endpoint, requests come from your VPC instead.
2. Use interface endpoints for heavy AWS API traffic - when the math works
Other AWS services (Amazon ECR, CloudWatch Logs, AWS STS, Secrets Manager…) can be reached through interface endpoints (AWS PrivateLink). They are not free: in US East, $0.01 per hour per Availability Zone plus $0.01 per GB. Compared with $0.045 per GB through the NAT gateway, each GB saves $0.035, so one endpoint in one zone ($7.30 per month) pays for itself above roughly 210 GB per month of traffic to that service. Below that, keep the NAT gateway.
3. Keep traffic in its own Availability Zone
If instances in zone A use a NAT gateway in zone B, you pay cross-AZ data transfer on top of NAT processing. With zonal NAT gateways, give each private subnet a route to the NAT gateway in the same zone. AWS also offers a regional NAT gateway (launched in November 2025) that expands across zones automatically, which simplifies this setup; its hourly charge applies per zone in use.
4. Pull container images without crossing the NAT gateway
Container platforms pull images constantly - on every deployment and every scale-out. Amazon ECR stores image layers in S3, so the S3 gateway endpoint (fix #1) already covers most of that volume; add the ECR interface endpoints if the API traffic justifies it. For images from public registries, an ECR pull-through cache keeps a copy in your account instead of downloading them again and again.
5. Fix chatty workloads
Flow logs often reveal a handful of noisy talkers: an agent shipping verbose logs to a third-party SaaS, servers downloading the same OS packages on every boot, or a service polling an external API every second. Lower log verbosity, compress what you ship, bake packages into your images and cache external calls.
6. Remove NAT gateways from idle environments
Every NAT gateway costs at least $32.85 per month in US East (730 hours × $0.045), even with zero traffic. Dev, test and sandbox VPCs often keep one - or one per zone - for nothing. Delete them where workloads don't need outbound access, or recreate them only when needed with infrastructure as code.
7. Consider IPv6 with an egress-only internet gateway
For IPv6 traffic, an egress-only internet gateway provides the same "outbound only" protection without NAT processing charges - you still pay normal data transfer out. If your workloads and destinations support IPv6, it's worth evaluating for high-volume internet traffic.
Quick checklist
- Identify your
NatGateway-Bytescost per account and VPC in Cost Explorer - Add S3 and DynamoDB gateway endpoints to every VPC (free)
- Rank the remaining destinations with VPC Flow Logs
- Add interface endpoints only where traffic exceeds ~210 GB/month per service
- Route each zone to its own NAT gateway, or use a regional NAT gateway
- Delete NAT gateways in idle environments
NAT gateway is one of 38 items in our free AWS Cost Optimization Checklist. If you'd rather have someone go through your accounts and put a dollar figure on each saving, that's exactly what our FinOps Diagnostic does.
