When an AWS bill climbs unexpectedly, grouping Cost Explorer by Usage Type usually puts a handful of cryptic strings near the top: USE1-DataProcessing-Bytes, USE1-NatGateway-Bytes, or USE1-DataTransfer-Regional-Bytes (or USE2- in Ohio and USW2- in Oregon). Looking at Usage Type on its own is an easy way to chase the wrong problem, though. AWS reuses the DataProcessing-Bytes label across completely unrelated services whose per-gigabyte rates differ by more than tenfold.
Maybe the question isn't why your VPC moved so many gigabytes last month, but rather which AWS service actually billed for them? Before touching a single route table, group Cost Explorer by both Service and Usage Type. Adding VPC Endpoints to trim a $0.045/GB NAT Gateway charge will do very little if most of your DataProcessing-Bytes line item is actually $0.50/GB CloudWatch Logs ingestion.
Disambiguating AWS Cost Explorer Usage Types
Standard on-demand rates in us-east-1 (checked November 2026; verify your region and any Enterprise Discount Program rates before estimating savings) map to very different traffic paths:
| Cost Explorer Service | Usage Type (us-east-1) |
Reference US East Rate | Traffic That Triggers the Charge |
|---|---|---|---|
AmazonCloudWatch |
USE1-DataProcessing-Bytes |
$0.50 / GB | Standard CloudWatch Logs ingestion, frequently inflated when VPC Flow Logs or verbose debug logs stream to CloudWatch Logs instead of S3 |
EC2 - Other / AmazonVPC |
USE1-NatGateway-Bytes (and NAT DataProcessing-Bytes) |
$0.045 / GB (+ $0.045/hr per NAT) | Every gigabyte traversing a Managed NAT Gateway in either direction, including S3 downloads, ECR image pulls, and external API calls from private subnets |
AWSELB |
USE1-DataProcessing-Bytes |
LCU component ($0.008 / LCU-hr) | Application or Network Load Balancer payload processing billed through Load Balancer Capacity Units |
AmazonVPC |
USE1-VpcEndpoint-Bytes |
$0.01 / GB (first PB tier) + $0.01/hr per AZ ENI | Traffic processed through AWS PrivateLink Interface Endpoints (ECR, Logs, STS, KMS, Secrets Manager, SQS) |
AmazonVPC / EC2 |
USE1-DataTransfer-Regional-Bytes |
$0.01 / GB each direction ($0.02/GB round-trip) | Traffic crossing Availability Zones within the same AWS region across EC2, RDS, caches, or misrouted endpoints |
You can run this read-only AWS CLI query to print your trailing 30-day spend and gigabyte volume broken down by service and usage type:
aws ce get-cost-and-usage \
--time-period Start=$(date -u -d '30 days ago' +%Y-%m-%d),End=$(date -u +%Y-%m-%d) \
--granularity MONTHLY \
--metrics "UnblendedCost" "UsageQuantity" \
--filter '{
"Dimensions": {
"Key": "USAGE_TYPE",
"Values": [
"USE1-DataProcessing-Bytes",
"USE1-NatGateway-Bytes",
"USE1-VpcEndpoint-Bytes",
"USE1-TransitGateway-Bytes",
"USE1-DataTransfer-Regional-Bytes"
]
}
}' \
--group-by Type=DIMENSION,Key=SERVICE Type=DIMENSION,Key=USAGE_TYPE \
--output table
One diagnostic trap is worth avoiding right here. Turning on VPC Flow Logs for a busy cluster and pointing them at CloudWatch Logs at $0.50/GB can easily cost more than the NAT Gateway charge you set out to investigate. Deliver VPC Flow Logs straight to an Amazon S3 bucket in Parquet format instead, and attach a 7- or 14-day lifecycle expiration rule.
Tracing Top NAT Destinations With v5 VPC Flow Logs and Athena
Once NAT Gateway bytes confirm as a real driver, how do you see what is actually crossing the gateway? Default v2 VPC Flow Logs only record srcaddr and dstaddr, which collapse NAT traffic onto the NAT Gateway's own private network interface IP. Turn on v5 custom fields including pkt-srcaddr and pkt-dstaddr delivered to S3, then query Athena against your NAT Gateway's private IP (using 10.0.1.25 below as an example) to rank the heaviest remote destinations:
SELECT
pkt_srcaddr AS internal_workload_ip,
pkt_dstaddr AS remote_endpoint_ip,
dstport,
ROUND(SUM(bytes) / 1073741824.0, 2) AS gigabytes_transferred
FROM vpc_flow_logs
WHERE year = '2026' AND month = '10'
AND (dstaddr = '10.0.1.25' OR srcaddr = '10.0.1.25')
GROUP BY pkt_srcaddr, pkt_dstaddr, dstport
ORDER BY gigabytes_transferred DESC
LIMIT 25;
From there, compare the top remote_endpoint_ip addresses against AWS's published ip-ranges.json prefixes. That tells you right away whether private-subnet workloads are paying NAT rates to reach AWS services like S3, ECR, CloudWatch, and DynamoDB, or talking out to external SaaS endpoints such as observability agents, webhook receivers, and package mirrors:
curl -s https://ip-ranges.amazonaws.com/ip-ranges.json | \
jq -r '.prefixes[] | select(.region=="us-east-1") | "\(.ip_prefix)\t\(.service)"' | head -n 30
Add S3 and DynamoDB Gateway Endpoints Carefully
AWS offers two kinds of VPC Endpoints, and their billing models could hardly be more different. Gateway VPC Endpoints only exist for Amazon S3 and DynamoDB. They work by inserting private AWS prefix lists (com.amazonaws.us-east-1.s3 and com.amazonaws.us-east-1.dynamodb) into your subnet route tables, with no hourly interface fee and no per-gigabyte processing charge for same-region traffic. Because Amazon ECR stores container image layers in S3, adding an S3 Gateway Endpoint also pulls image-layer downloads off your NAT Gateways during container autoscaling and CI builds. Interface VPC Endpoints, powered by AWS PrivateLink, cover services like ECR API/DKR, CloudWatch Logs, KMS, STS, Secrets Manager, and SQS—and those bill both an hourly fee per Availability Zone ENI and a per-gigabyte rate.
Free does not mean zero operational risk, however. Adding a Gateway Endpoint injects a more specific prefix-list route into a live VPC route table. As AWS Gateway Endpoint documentation points out, existing long-lived TCP connections to S3 or DynamoDB that opened through a NAT Gateway may keep following the old path until closed or experience connection resets when routes shift. Roll out route-table changes during a planned maintenance window, check that your endpoint policy permits every bucket your applications touch, and confirm long-lived worker pools reconnect cleanly.
Check Break-Even Volume Before Deploying PrivateLink Endpoints
Does every AWS service need an Interface Endpoint? Usually not. It is not unheard of for a team to deploy fifteen PrivateLink Interface Endpoints across three Availability Zones in the name of cost savings, only to watch the monthly VPC bill go up. So long as public internet egress keeps your NAT Gateways running anyway, the NAT hourly charge stays put, which means each Interface Endpoint has to save enough per-gigabyte processing fees to pay for its own network interfaces.
The math is straightforward. In an illustrative three-AZ US East VPC, one Interface Endpoint costs $21.90 a month in fixed ENI charges (3 AZs x 730 hours x $0.01/hr). Shifting traffic from a $0.045/GB NAT Gateway onto a $0.01/GB PrivateLink endpoint saves $0.035 per gigabyte. Divide $21.90 by $0.035, and a service needs roughly 626 GB per month across three AZs—or about 417 GB per month across two AZs—before an Interface Endpoint lowers your bill.
Busy clusters pushing CloudWatch Logs or frequent ECR control-plane pulls can pass 626 GB a month easily. Low-volume control-plane calls like KMS or Secrets Manager almost never justify dedicated Interface Endpoints on cost alone, unless strict private-subnet compliance rules forbid NAT egress altogether. You can plug in your own AZ count and monthly gigabytes with our AWS NAT Gateway vs. VPC Endpoint Cost Calculator before provisioning anything.
Trim Cross-AZ Chatter Without Giving Up Redundancy
Sometimes DataTransfer-Regional-Bytes rivals NAT spend on the very same invoice. When it does, check whether workloads are crossing Availability Zones out of habit rather than resilience.
If you deploy an Interface Endpoint, place an ENI in every active compute AZ and enable Private DNS so pods and instances talk to a local endpoint interface instead of paying $0.02/GB round-trip to reach another zone. Look at cache, queue, and database read patterns too: read-heavy worker pools or service meshes can often prefer healthy same-AZ read replicas or cache nodes when latency rules allow, while keeping multi-AZ failover intact on primary databases and load balancers. And of course, turning on gzip or zstd compression on log forwarders and OpenTelemetry exporters shrinks outbound payloads and NAT processing bytes at the same time (see our APM and Datadog cost optimization guide).
Pair transfer cleanup with a recurring monthly cloud cost checklist and our broader guide on reviewing an AWS bill for unnecessary costs.