--- title: "Short AI Jobs Can Cross More Than One EC2 Billing Minimum" description: "Leviathan Systems — Technical content Published 2026-10-21; updated 2026-10-21." canonical: "https://www.leviathansystems.co/articles/ai-ec2-running-period-minimum-idle-cost" last-updated: "2026-10-21" --- # Short AI Jobs Can Cross More Than One EC2 Billing Minimum Leviathan Systems — Technical content Published 2026-10-21; updated 2026-10-21. Compare short-job compute offers by each instance's running periods before applying the documented 60-second minimum. Eight separate 30-second periods produce 480 billed seconds; one gap-free 240-second period produces 240. But one period containing the same eight intervals plus seven 100-second running-idle gaps produces 940. These fictional quantities isolate eligible ordinary EC2 On-Demand instance usage. Task count, actual running seconds and minimum-adjusted billed seconds answer different questions. ## Apply the minimum to the instance-running period AWS's EC2 documentation, checked October 21, 2026, describes On-Demand usage billed by running seconds with a 60-second minimum. Its lifecycle guide applies a fresh minimum whenever an instance returns from stopped to running. Running idle time is billable. Rebooting continues the existing billing period without another minimum. This comparison assumes an offer subject to that ordinary per-second On-Demand instance-usage basis. It establishes no particular instance type, operating system, software-license eligibility or price. Spot, commitments and hibernation need their own applicable terms. For one running period lasting t seconds, the billed quantity is max(60, t): use the larger of 60 and t. For several periods, calculate that quantity for each period and then add the results. A 75-second period contributes 75 billed seconds under this basis; the minimum does not round every period into whole minutes. Keep each instance's periods separate even if different instances run simultaneously. Their summed quantities are billed instance-seconds, which can exceed the elapsed wall-clock interval. A task finishing, an application restarting or a period of low utilization is insufficient evidence of a new provider billing period. References: [EC2-OD](https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/ec2-on-demand-instances.html), [EC2-STATE](https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/ec2-instance-lifecycle.html) ## EC243: eight short periods or one compact period Consider fictional comparison EC243: eight short AI jobs, each assigned 30 seconds of instance-running time. Any startup or preparation time occurring in the running state is already included in those supplied durations. These are stipulated inputs, not measured workload timings. The offers use the same eligible instance-usage basis and, initially, the same unspecified rate. Offer A gives each job a separate 30-second instance-running period. Its actual running time totals 8 x 30 = 240 seconds. Each period is shorter than the minimum, so the billed quantity is 8 x max(60, 30) = 480 instance-seconds. Intervening time is stipulated to be outside the running state and adds nothing to this particular usage component. Offer B contains all eight jobs in one uninterrupted 240-second running period, with no gaps. Its billed quantity is max(60, 240) = 240 instance-seconds. The initial minimum is already satisfied within that period. Eight job identities have not created eight minimums. A and B therefore have equal stipulated running time but different billed quantities. Adding A's 240 running seconds first and applying the minimum only once would incorrectly return 240 billed seconds, omitting 240 seconds of minimum uplift. Applying the floor before summing preserves the difference. References: [EC2-OD](https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/ec2-on-demand-instances.html), [EC2-STATE](https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/ec2-instance-lifecycle.html) ## Running idle gaps can reverse the result Offer C keeps the eight 30-second allocated intervals but inserts seven 100-second idle gaps between them, all inside one uninterrupted running period. The intervals contribute 240 seconds and the gaps contribute 7 x 100 = 700 seconds. Its running period is therefore 940 seconds, and max(60, 940) = 940 billed instance-seconds. C has only one minimum, yet its billed quantity exceeds A's 480 by 460 seconds. B and C both contain eight jobs in one running period, but their billed quantities differ by the 700 running-idle seconds. A and B both contain 240 actual running seconds, but their billed quantities differ because of the repeated minimums. Neither task count nor running-period count alone reproduces all three results. The example also supplies a useful comparison boundary. A has 240 seconds of minimum uplift above its 240 actual running seconds. For a single period containing the same 240 allocated seconds plus a total of G running-idle seconds, the billed quantity is 240 + G. It equals A's 480 when G is 240 seconds; below that gap total it is lower, and above it higher. Use this boundary to compare submitted timelines. The application owner still determines which timelines are feasible. References: [EC2-STATE](https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/ec2-instance-lifecycle.html) ## Connect the running timeline to the offered rate A useful offer identifies the proposed running periods for each instance, the tasks assigned to them and the idle time included. Identify the state boundaries behind each duration and label it as a forecast or an observation. An application-runtime subtotal cannot establish the provider's running duration if it excludes preparation or other running time. Retain actual running seconds and minimum-adjusted billed seconds as separate quantities. For EC243, the sufficient quantity return is A: eight 30-second periods, 240 actual and 480 billed instance-seconds; B: one 240-second period, 240 actual and billed; C: one 940-second period, including 700 idle seconds, with 940 actual and billed. A response containing only eight jobs or 240 seconds leaves the comparison unresolved. Let r denote the applicable charge per billed instance-second. At the same rate, these components are 480r, 240r and 940r. If the offers have different rates, retain them separately: 480rA, 240rB and 940rC. A lower billed-seconds quantity alone cannot establish a lower monetary component across different rates. The actual offer must supply the exact product and charging basis, applicable AWS Region, currency, pricing date and rate. Preserve the rate's stated unit when applying it; an hourly rate would multiply the billed seconds divided by 3,600. EC243 provides no numerical price or account-level forecast. References: [EC2-OD](https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/ec2-on-demand-instances.html), [EC2-STATE](https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/ec2-instance-lifecycle.html) ## Use the component return with its remaining decisions The returned running periods, minimum-adjusted quantities and applicable rates are enough to reproduce this instance-usage comparison. Keep the three results together: 480 versus 240 shows the effect of repeated minimums at equal running time; 940 shows how running-idle gaps can outweigh that uplift. The appointed application and platform owners must establish whether the proposed grouping, preparation, available capacity and deadlines permit each timeline. Equal stipulated durations establish no real performance or scheduling equivalence. The related GPU Cloud Availability guide addresses the separate capacity and usable-window question. The appointed account owner must also complete the environment's charge scope. Use Stopping a GPU Instance Leaves Other Billing Clocks Running for the retained-resource question instead of extending this one-component calculation to the entire bill. A reproducible comparison can close the minimum calculation while those other decisions remain open. If the wider program also includes a separately approved customer-equipment installation, the related Colocation Infrastructure Installation service and project inquiry address agreed physical IT work at the facility. Share a nonconfidential equipment and site summary for that separate scope. Cloud scheduling and billing decisions remain with their appointed owners. References: [EC2-STATE](https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/ec2-instance-lifecycle.html), [L1](https://www.leviathansystems.co/services/colocation-installation) ## Primary sources - [EC2-OD — AWS EC2 User Guide: Purchasing On-Demand Instances](https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/ec2-on-demand-instances.html). Checked 2026-10-21. Supplies the running-seconds basis and 60-second minimum for the ordinary On-Demand usage comparison. No actual rate or product eligibility is inferred. - [EC2-STATE — AWS EC2 User Guide: Amazon EC2 instance state changes](https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/ec2-instance-lifecycle.html). Checked 2026-10-21. Explains the minimum for each stopped-to-running transition, billable running idle time and continuing reboot billing. The example excludes hibernation. - [L1 — Leviathan Systems: Colocation Infrastructure Installation](https://www.leviathansystems.co/services/colocation-installation). Checked 2026-10-21. Defines customer-side physical IT installation and facility interfaces for a separately approved equipment project. ## Related Leviathan resources - [Stopped compute and retained resources: the separate billing intervals](/articles/gpu-cloud-stopped-compute-retained-resource-billing) - [GPU cloud availability: quotas, capacity and usable windows](/articles/gpu-cloud-capacity-availability) - [Colocation installation for a separate approved equipment project](/services/colocation-installation) - [Discuss that separate physical installation scope](/contact?service=colocation-installation&from=%2Farticles%2Fai-ec2-running-period-minimum-idle-cost) - [Four Completed Reads Can Return Fewer Bytes Than Four Requested Buffers](/articles/ai-io-uring-read-completion-returned-byte-count) ## Is there also an approved customer-equipment installation project? For that separate physical scope, share a nonconfidential facility and equipment summary, intended IT installation work, access dependencies and required handoff to discuss scope and availability. [Discuss the scope](/contact?service=colocation-installation&from=/articles/ai-ec2-running-period-minimum-idle-cost)