--- title: "NAT and Endpoint Charges Along an AI Data Path" description: "Leviathan Systems — Technical content Published 2026-10-20; updated 2026-10-20." canonical: "https://www.leviathansystems.co/articles/ai-data-path-nat-endpoint-processing-charges" last-updated: "2026-10-20" --- # NAT and Endpoint Charges Along an AI Data Path Leviathan Systems — Technical content Published 2026-10-20; updated 2026-10-20. A zero data-transfer line answers one pricing question. The path can still use a network service that charges for provisioned availability and processed data. Identify the actual NAT gateway or endpoint product, its location, its billing interval and its measured traffic before comparing proposals. Same-Region and private describe aspects of a connection; neither supplies a complete charge model. ## Locate the zero in the documented service path AWS's VPC pricing example sends a file from EC2 through NAT to S3. EC2 and NAT share an Availability Zone and use private IP addresses; S3 is in the same Region. Transfer costs zero in that example, while NAT hours and processing remain chargeable. That is a useful starting point for an AI dataset-staging proposal showing zero transfer. Preserve the source, destination, direction and location conditions when interpreting that line. A different transfer path needs its own applicable terms; the word S3 alone cannot establish them. There are separate pricing questions within the same movement. Transfer concerns the applicable movement category. Processing concerns data passing through the purchased network service. Availability concerns how long that service remains provisioned under its terms. Several applicable charge categories can accompany one movement without automatically being an accounting duplication. This comparison uses AWS terms checked October 20, 2026. It evaluates selected charge components, rather than an account bill or an approved network design. References: [NM1](https://aws.amazon.com/vpc/pricing/) ## NM-119: three NAT hours and 100 processed GB For an original fictional case, stipulate one ordinary NAT gateway provisioned and available for exactly three complete hours. The selected EC2-to-S3 traffic follows the same-AZ, same-Region conditions above. Separately stipulate that this traffic contributes 100 provider-metered GB of NAT processing during that interval. No real account or traffic measurement is represented. Let H mean an unevaluated charge per NAT-hour, P an unevaluated charge per processed GB, and D the applicable transfer charge for this selected traffic. The component is 3H + 100P + D. Under the source-qualified transfer case, D = 0, leaving 3H + 100P. That substitution removes the transfer term; it provides no reason to set H or P to zero. AWS's NAT pricing documentation identifies available hours and processed gigabytes as separate charging quantities. In this case, 3 multiplies the hourly coefficient and 100 multiplies the processing coefficient. Interchanging them would change the units and the meaning of the calculation. The coefficients are symbols, not reproduced AWS prices. No benefit tier, discount, tax or complete invoice is modeled. The example deliberately uses complete hours: the applicable NAT pricing terms bill a partial NAT Gateway-hour as a full hour, so an arbitrary elapsed interval cannot silently be substituted as a fractional hourly quantity. References: [NM1](https://aws.amazon.com/vpc/pricing/), [NM4](https://docs.aws.amazon.com/vpc/latest/userguide/nat-gateway-pricing.html) ## Establish processed data independently of dataset size The 100 GB premise is a service-meter quantity. It was not calculated from the size of a stored dataset, a completed object or the GPU job's output. GB here retains the provider meter's label; no conversion from GiB or from a payload-size display is assumed. For example, a dataset inventory can stay unchanged while a workload accesses some contents repeatedly. That inventory does not describe the number of traversals through a particular network service. Conversely, contents already available to the consumer need not cross the same service again. These observations explain why the buyer needs a path-specific quantity; they do not establish any precise processing total. A proposal should therefore connect its processing assumption to the selected traffic and interval. If the gateway also serves unrelated traffic, separate that population from the example's stipulated 100 GB before attributing a processing component to this dataset. An estimate for one dataset cannot be presented as the resource's complete meter population. The availability quantity has a different boundary. Three hours of a shared gateway's availability do not become three hours per dataset merely because several datasets use it. First identify the purchased resource and its charge; any later division among teams belongs to the existing shared-platform cost-allocation guide. References: [NM4](https://docs.aws.amazon.com/vpc/latest/userguide/nat-gateway-pricing.html) ## An S3 gateway endpoint changes the product being priced Consider a separately proposed S3 gateway-endpoint path whose eligibility the network owner has established. AWS documents no additional charge for gateway endpoints, and its NAT example specifically distinguishes their absence of endpoint hourly and processing charges. This is a statement about that product, not a rule that every endpoint or private connection is free. Eligibility matters before economics. The S3 gateway-endpoint documentation requires the endpoint and S3 buckets to be in the same Region. It also excludes access from on-premises networks and identifies other remote connectivity restrictions. A proposal requiring such access cannot assume that this endpoint replaces its required path. An endpoint charge of zero leaves other purchased services to be evaluated. S3 storage and requests have their own pricing categories. If a NAT gateway remains separately provisioned for other traffic, removing this dataset's traffic from its processing population does not erase that retained resource's availability charge. Keep two comparisons visible: the processing attributable to the selected traffic and the availability of resources still purchased for the environment. Claiming the whole NAT component as avoided would require a different premise from the one just stated. No resource removal or traffic change is proposed here. References: [NM1](https://aws.amazon.com/vpc/pricing/), [NM2](https://docs.aws.amazon.com/vpc/latest/privatelink/vpc-endpoints-s3.html), [NM4](https://docs.aws.amazon.com/vpc/latest/userguide/nat-gateway-pricing.html), [NM6](https://aws.amazon.com/s3/pricing/) ## A two-AZ interface endpoint contributes six endpoint-AZ-hours An interface endpoint is a different offer. AWS's interface-endpoint documentation identifies hourly usage and data processing charges. The PrivateLink pricing terms apply hours in each Availability Zone where the endpoint remains provisioned and bill partial endpoint-hours as full hours. Independently of the NAT case, stipulate one interface endpoint provisioned in two AZs, each for exactly three complete hours. Separately stipulate 100 provider-metered GB processed across that endpoint in total. There are 2 x 3 = 6 endpoint-AZ-hours. With E as an unevaluated charge per endpoint-AZ-hour and Q as an unevaluated charge per processed GB for this selected component, the expression is 6E + 100Q. Using 3E would omit one AZ's provisioned interval. Using 200Q would instead duplicate the stipulated aggregate traffic: the 100 GB was supplied as the endpoint total, not as 100 GB per AZ. The AZ count multiplies the availability quantity; it does not automatically multiply a separately supplied aggregate data quantity. AWS also notes that some services can include their interface-endpoint cost within the service price. The applicable service's pricing information identifies those cases. An absent separate line therefore calls for an inclusion check. If a component is already included in a quoted service amount, adding it again as a separate amount would duplicate that cost. These are interface-endpoint terms. Gateway Load Balancer, resource and tunnel endpoints are different products. The two-AZ premise establishes neither a required architecture nor an availability result. The formulas supply no monetary winner, throughput result, security assessment or proof that the proposed paths meet the same connectivity requirement. References: [NM3](https://aws.amazon.com/privatelink/pricing/), [NM5](https://docs.aws.amazon.com/vpc/latest/privatelink/create-interface-endpoint.html) ## Turn the zero-transfer quote into a reproducible proposal Return the fictional NAT proposal to the account and network owners with a concrete question: does the quoted zero refer to the exact movement and location conditions documented above? A response should identify the actual source, destination and intervening service products. Same-Region access is too broad to answer that question on its own. Then connect the quantities to that path. The case needs one identified NAT resource, three complete available hours and 100 GB attributable to the selected traffic. If the real proposal uses another interval, partial-hour rules must be applied under its actual terms. If its 100 GB comes only from a dataset-size estimate, the processing premise remains unresolved. Neither missing interval evidence nor an unexplained data quantity should be treated as zero. For an endpoint alternative, ask the owners to establish which product can serve the required source and destination. The fictional interface offer needs two identified AZs and the provisioned interval in each; its data quantity was aggregate. The gateway-endpoint offer needs its S3 access eligibility established before its additional-charge statement becomes relevant. These are different service proposals until their functional requirements have been reconciled. Finally, ask which amounts are separate, included elsewhere or outside the selected comparison. Keep any retained NAT availability visible, identify the transfer category actually claimed, and carry surrounding service charges alongside the component formulas. This makes it possible to review omissions and duplication without treating a missing price or unverified path as a favorable result. With those inputs returned, account owners can apply the relevant quote to comparable service requirements. An unresolved path or meter remains an open input to that decision. References: [NM1](https://aws.amazon.com/vpc/pricing/), [NM2](https://docs.aws.amazon.com/vpc/latest/privatelink/vpc-endpoints-s3.html), [NM3](https://aws.amazon.com/privatelink/pricing/), [NM4](https://docs.aws.amazon.com/vpc/latest/userguide/nat-gateway-pricing.html), [NM5](https://docs.aws.amazon.com/vpc/latest/privatelink/create-interface-endpoint.html), [NM6](https://aws.amazon.com/s3/pricing/) ## Keep the billing review and physical project with their owners The immediate review belongs to the appointed account and network owners. The related stopped-instance and storage-duration articles address separate resource and retention intervals. GPU Cloud or Colocation explains operating responsibility; the multi-site network guide addresses transport and physical handoffs. Those decisions can matter to the wider project without determining this path's processing meter. If the wider program also needs physical connections between compute, storage or network equipment, Leviathan's Structured Cabling & Fiber Installation service and the linked project inquiry provide the relevant route to discuss that work. Share a nonconfidential equipment outline, approved endpoints and routes, requested cabling work, material responsibilities and required connection records so scope and availability can be discussed. That inquiry concerns agreed physical installation and handoff. Cloud network configuration, endpoint selection and account billing interpretation remain with the appointed cloud and customer teams. References: [L1](https://www.leviathansystems.co/services/structured-cabling) ## Primary sources - [NM1 — AWS: Amazon VPC pricing — ordinary NAT Gateway terms and same-AZ S3 example](https://aws.amazon.com/vpc/pricing/). Checked 2026-10-20. Ordinary NAT charge categories, partial-hour treatment and the qualified EC2-to-S3 example support this component comparison. - [NM2 — AWS: Gateway endpoints for Amazon S3](https://docs.aws.amazon.com/vpc/latest/privatelink/vpc-endpoints-s3.html). Checked 2026-10-20. Additional-charge statement, same-Region applicability and restrictions on remote access establish the boundary of the gateway-endpoint comparison. - [NM3 — AWS: PrivateLink pricing — interface endpoints](https://aws.amazon.com/privatelink/pricing/). Checked 2026-10-20. Per-AZ provisioned hours, processed gigabytes, partial-hour rounding and service-specific inclusion of endpoint costs. Rates for other endpoint products are outside this comparison. - [NM4 — AWS: Pricing for NAT gateways](https://docs.aws.amazon.com/vpc/latest/userguide/nat-gateway-pricing.html). Checked 2026-10-20. Confirms distinct availability and processing quantities used in the symbolic NAT case. - [NM5 — AWS: Access an AWS service using an interface VPC endpoint](https://docs.aws.amazon.com/vpc/latest/privatelink/create-interface-endpoint.html). Checked 2026-10-20. The interface-endpoint overview identifies its hourly usage and data-processing meters. - [NM6 — AWS: Amazon S3 pricing — separate service cost components](https://aws.amazon.com/s3/pricing/). Checked 2026-10-20. Storage, requests and transfer are separate S3 pricing categories; an endpoint's additional-charge statement does not settle them. - [L1 — Leviathan Systems: Structured Cabling & Fiber Installation](https://www.leviathansystems.co/services/structured-cabling). Checked 2026-10-20. Agreed physical cabling to approved endpoints and routes, with material responsibilities and connection records established for the project. ## Related Leviathan resources - [Stopped GPU compute and separately retained resource charges](/articles/gpu-cloud-stopped-compute-retained-resource-billing) - [AI dataset revision lifetime and minimum storage duration](/articles/ai-dataset-storage-minimum-duration-billing) - [Shared platform cost allocation after the resource bill is established](/articles/shared-ai-platform-cost-allocation-versus-resource-bill) - [GPU cloud or colocation: ownership and operating responsibilities](/articles/gpu-cloud-vs-colocation-ownership) - [Multi-site AI networking: transport and physical handoffs](/articles/multi-site-ai-cluster-network-handoff) - [Structured Cabling & Fiber Installation](/services/structured-cabling) - [Discuss a separate physical compute-to-storage cabling project](/contact?service=structured-cabling&from=%2Farticles%2Fai-data-path-nat-endpoint-processing-charges) ## Does the wider project need physical cabling? Share a nonconfidential equipment outline, approved endpoints and routes, requested cabling work and required records to discuss physical scope and availability. [Discuss the scope](/contact?service=structured-cabling&from=/articles/ai-data-path-nat-endpoint-processing-charges)