A network account runs one NAT Gateway per Availability Zone and, behind it, twelve teams go out to the internet: package downloads, calls to external APIs, backups to S3 that nobody routed through an endpoint. The VPC bill shows one line for the NAT hours and another for the GB it processed, both against the network account. When the time comes to split it, the question matches the one for any orphan resource: whose traffic is this? Until now the answer was a script over the flow logs or a fixed percentage per team. Since November 16 AWS has an answer of its own, and the other three clouds remain at the script stage. This note covers all four; everything comes from the official documentation and the sources sit at the end.
What a NAT Gateway charges and why the line is a single one
AWS charges a NAT Gateway for each hour it stays available and for each GB it processes. The owner of the resource is the account it lives in; the GB come from the instances that send traffic, whether they sit in that account or in a spoke VPC that arrives through a Transit Gateway. Neither line knows who sent the bytes: lineItem/ResourceId is the NAT ARN and lineItem/UsageAccountId is the network account. That is why the split was always manual, and why the 11/16 announcement says most customers solved it with custom scripts and manual estimates.
AWS: Networking Flexible Cost Allocation, since 2026-11-16
Networking Flexible Cost Allocation (FCA) is an Amazon VPC feature that attributes the data transfer and data processing charges of a shared networking resource to the resources, subnets, VPCs, accounts or IP ranges that generated the usage. It supports NAT Gateway (both dimensions), data transfer between EC2 instances inside the same VPC, and VPC peering in the same Region. It changes neither the bill nor who pays: the account that owns the NAT keeps paying the total; what changes is how the Cost and Usage Report itemizes the usage. It costs nothing extra.
It works with a policy and an association. The policy says which dimensions to split, data-transfer and data-processing, and the reporting granularity: eni, ec2, subnet, vpc, account or cidr (an IP range with a label you define, for example 10.0.1.0/24 = team-b). The association applies the policy to specific NAT Gateways, or to every NAT Gateway in an account, a VPC or a subnet, including the ones you create later:
# policy-document.json: both dimensions, labelled by source range { "ReportingGranularity": ["cidr"], "Dimensions": ["data-transfer", "data-processing"], "CidrRangeSpecifications": { "PrefixesType": "source-prefixes", "Prefixes": [ { "CidrRange": "10.0.0.0/24", "CidrLabel": "team-a" }, { "CidrRange": "10.0.1.0/24", "CidrLabel": "team-b" } ] } } aws ec2 create-networking-flex-cost-allocation-policy \ --network-resource-type natgw \ --nat-fca-policy-document file://policy-document.json \ --show-back-enabled --description "Shared NAT by team" # returns FcaPolicyId (fcap-...) and ExpectedEffectiveAt aws ec2 associate-networking-flex-cost-allocation-policy \ --fca-policy-id fcap-0123456789abcdef0 \ --network-resource-ids nat-0123456789abcdef0 aws ec2 describe-networking-flex-cost-allocation-policy-associations \ --filters Name=resource-id,Values=nat-0123456789abcdef0 What you need to see it: the policy associated with showback on, the "Networking flexible cost allocation" preference turned on in Cost Management preferences by the management account (or a standalone account; a member account cannot), and a CUR, legacy or 2.0 with Data Exports, that includes split cost allocation data and resource IDs. A report without split cost allocation data keeps showing a single line even with the policy in effect. The association takes effect at its creation time plus 60 minutes, rounded up to the next hour (created at 2:01 PM, effective at 4:00 PM), and it does not apply retroactively.
The CUR gains one new line per resource that generated usage, with splitLineItem/SplitCost and splitLineItem/SplitUsage, splitLineItem/ParentResourceId pointing at the NAT ARN, and two AWS-generated tags: aws:networking-fca:allocated-account (the account that owns the resource that generated the traffic) and aws:networking-fca:allocation-policy (policy and version, for example fcap-0123456789abcdef0:2). Whatever FCA cannot attribute lands on an (unattributed) line. The original NAT line stays there with the total cost: if you add the split lines to the unblended one, you count it twice. Those lines go to the report of the account that owns the NAT and to the management account's report, not to the teams' accounts.
What FCA does not do, per its own limitations page:
- It does not split traffic that the NAT forwards to a virtual private gateway, a Transit Gateway, another VPC or a peering connection: it covers what leaves to the internet through an internet gateway, or reaches S3 and DynamoDB through a gateway endpoint, and nothing else.
- If spoke VPCs arrive through a Transit Gateway, the breakdown is by VPC and account even when the policy asks for subnet, ec2 or eni; and the Transit Gateway's own charges stay outside this policy.
- Private NAT Gateways have no support yet. With a Network Firewall or Gateway Load Balancer endpoint between the source and the NAT, that traffic bills as usual to the NAT owner.
- Traffic inside a single Region. Available at no charge in every commercial Region except Bahrain and UAE, in the China Regions, and not in GovCloud.
- A resource holds one policy at a time; up to 100 policies per Region per account, up to 100 associations per policy, up to 100 CIDR ranges per policy. You can associate resources your account owns, nothing shared with you.
- If a network ACL on the NAT subnet drops traffic that the source security group allowed, FCA still counts it and the charges can go up: block it at the source with the security group.
- Volume: for each allocated usage type and operation, the report adds one line per resource that generated usage, per hour. 200 instances against one NAT with a one-usage-type policy means 4,800 lines a day; reporting by VPC or by account reduces them.
AWS before FCA, and what stays manual
For traffic before the association, or in the Regions and paths FCA does not cover, the path remains VPC Flow Logs. In the NAT ENI's record, srcaddr is the private IP of the NAT's own interface and pkt-srcaddr is the original IP of the instance that sent the packet; for a regional NAT Gateway, interface-id comes as - and srcaddr already carries the original IP. Summing bytes by pkt-srcaddr over a period gives you the split, and then you map each IP to its ENI, its instance and its account. That is what FCA does on its own; with flow logs you do it yourself and repeat it every month.
Azure: the NAT Gateway is a resource, and Cost Management splits by tags
In Azure the NAT Gateway is a resource of its own that you attach to subnets of a VNet, so the cost already reaches Cost Analysis with a resource ID. The split per team depends on tags, and Cost Management has two rules: you must apply tags directly to the resource, and they do not inherit from the resource group unless you enable the tag inheritance setting, which applies subscription and resource group tags to child resource usage records within 24 hours, and covers new data in the current month, not past months.
az tag update \ --resource-id "/subscriptions/<sub>/resourceGroups/<rg>/providers/Microsoft.Network/natGateways/<nat>" \ --operation merge --tags owner=team-x cost-center=1234 # then: Cost Management + Billing > Cost analysis > Group by: Tag > owner That answers "who owns the NAT", not "who owns the traffic" when subnets from different teams leave through the same gateway. For the second one, VNet flow logs: you enable them at the virtual network level and they record traffic for every supported workload, gateways included, with bytes sent and received per flow tuple. Microsoft recommends disabling NSG flow logs on the same workloads first, to avoid recording the traffic twice. The aggregation by source IP remains yours, or Traffic Analytics does it at a per-GB processing rate.
az network watcher flow-log create --location eastus --resource-group <rg> \ --name vnet-flow-log --vnet <vnet> --storage-account <storage> GCP: Cloud NAT charges per VM and per GiB, and the per-VM metric exists
Cloud NAT charges a rate per hour per gateway based on the number of VMs that use it, capped at 32 (a gateway with more VMs pays the capped rate), plus a price per GiB processed in both directions, plus data transfer out. The gateway lives on a Cloud Router, so in the BigQuery billing export the cost lands on the router's project, with the resource labels and, in the detailed export, resource.name.
For the split, GCP publishes a per-VM metric: compute.googleapis.com/nat/sent_bytes_count (and received_bytes_count) on the gce_instance resource, with the labels nat_gateway_name, nat_project_number, router_id and ip_protocol, sampled every 60 seconds. In Metrics Explorer: resource VM Instance, metric "Sent bytes count" in the NAT group, grouped by instance_name or by the VM's team label. From there to a percentage per team is a spreadsheet, not a script over logs. The gateway logs, if you enabled them, read like this:
gcloud logging read 'resource.type=nat_gateway' --project=<project> --freshness=1d --limit=20 OCI: the NAT Gateway has no SKU; the split goes by tag and by VNIC
OCI's networking pricing page has no NAT Gateway SKU: what you pay is outbound data transfer, with the first 10 TB per month free by origin (North America, Europe and UK; APAC, Japan and South America; Middle East and Africa). The cost reports carry one row per resource per hour with product/resourceId and the tags/ columns, come out every six hours and stay for one year; any tag shows up in Cost Analysis and in the reports, with no need to mark it as cost-tracking.
oci network nat-gateway update --nat-gateway-id <ocid> \ --freeform-tags '{"owner":"team-x"}' oci usage-api usage-summary request-summarized-usages --tenant-id <tenancy-ocid> \ --granularity DAILY --query-type COST \ --time-usage-started 2026-11-01T00:00:00Z --time-usage-ended 2026-11-21T00:00:00Z \ --group-by '["resourceId"]' To know which VNIC generated the traffic, VCN flow logs: you enable them for every VNIC in a VCN or a subnet, or for specific instances, and each record carries data.sourceAddress, data.bytesOut, oracle.vnicocid and oracle.vnicsubnetocid. As in Azure and GCP, the sum per team is yours to do.
FAQ
Does FCA move the charge to the team's account? No. The documentation repeats it on every page: the allocated lines carry the lineItem/UsageAccountId of the account that owns the NAT and the total does not change. That is showback, not chargeback; the internal charge remains your job.
Do I see it in Cost Explorer? The documentation we read talks about the Cost and Usage Report, legacy and 2.0 with Data Exports, with split cost allocation data; we saw no FCA view in Cost Explorer and we do not claim one.
Why does my policy split nothing? Three causes: the association has not reached its effective time yet, the preference is off in the management account, or the CUR does not include split cost allocation data. And if you created two policy versions back to back, the intermediate one never takes effect: the association always uses the latest.
Does it work for Transit Gateway? No: FCA supports NAT Gateway, EC2 instances and VPC peering; Transit Gateway charges remain unallocated.
What Control Cloud does with this
I build Control Cloud. The map lists AWS NAT Gateways per VPC, with state, subnet and public IP, OCI ones per VCN and IONOS ones per datacenter, and it knows that a NAT's ENI and Elastic IP are not orphans even without an instance, so it does not offer to delete them. Real costs come from Cost Management in Azure, from the Usage API in OCI and from Cost Explorer in AWS on demand; GCP stays estimated because it has no spend API. What it does not do: it does not read the CUR or FCA's split lines, it does not read flow logs from any cloud, and it does not list Azure NAT Gateways yet. The split per team happens with what is in this note; the map tells you the number of NATs you have and in which VPC. If the traffic comes from a resource with no owner, the note on who created this resource is the previous step, and the ones on public IPs that bill and what survives when you delete a VM are the rest of the inventory.
Sources
- AWS What's New, "AWS announces Networking Flexible Cost Allocation for Amazon VPC" (2026-11-16): until now the CUR showed networking usage as a single item and customers split it with custom scripts and manual estimates.
- Amazon VPC User Guide, "Networking Flexible Cost Allocation" (what it attributes, supported resources, effective time +60 min rounded up, Regions, no charge); "How Networking Flexible Cost Allocation works" (policy, association, dimensions, granularities, showback); "Working with Networking Flexible Cost Allocation" (console, CLI create/associate/describe-networking-flex-cost-allocation-policy, Cost Management preference, versions); the FCA page for NAT gateway costs (networking-flex-cost-allocation-natgw: covered paths, Transit Gateway by VPC and account, network ACL); "Limitations" (single Region, private NATs, Network Firewall and GWLB, quotas).
- AWS Data Exports User Guide, "Understanding split cost allocation data" § networking (tags aws:networking-fca:allocated-account and aws:networking-fca:allocation-policy, line item formula) and "Example of split cost allocation data for networking" (splitLineItem/SplitCost, SplitUsage, ParentResourceId, (unattributed) line, which reports carry it).
- Amazon VPC User Guide, "Pricing for NAT gateways" (per hour and per GB) and "Flow log records" (srcaddr, pkt-srcaddr, interface-id - on regional NAT).
- Microsoft Learn, "Understand Cost Management data" (tags applied directly to the resource, no inheritance from the resource group); the tag inheritance page (enable-tag-inheritance: subscription and resource group tags on child records, 24 hours, current month); az tag update reference; "Virtual network flow logs overview" (VNet level, gateways included, bytes per tuple, disable NSG flow logs first, Traffic Analytics per GB); "Manage virtual network flow logs using Azure CLI" (az network watcher flow-log create).
- Google Cloud, "Cloud NAT pricing" (per hour by VMs capped at 32, per GiB processed, data transfer out); the Cloud NAT monitoring page (nat/docs/monitoring: nat/sent_bytes_count metric on gce_instance, labels, 60 s; gcloud logging read 'resource.type=nat_gateway'); "Structure of standard data export" and "Structure of detailed data export" (labels, resource.name).
- Oracle Cloud, "Networking pricing" (no NAT Gateway SKU; outbound data transfer, first 10 TB/month free by origin); "Cost and Usage Reports Overview" (one row per resource per hour, product/resourceId, tags/, every 6 hours, one year); "Tagging Overview" (any tag works in Cost Analysis and reports); oci network nat-gateway update and oci usage-api usage-summary request-summarized-usages references (--group-by resourceId); "VCN Flow Logs" and "Details for VCN Flow Logs" (enablement points, data.sourceAddress, data.bytesOut, oracle.vnicocid).



