A cloud estimate can look convincing while missing the costs that change the decision: a second database copy, traffic between zones, or a network gateway that runs all month. The way to avoid that is to estimate one workload from measured usage, then run a small version of it and compare the bill with the forecast.

This is a moderate job. Allow half a day if you already have usage reports and a proposed cloud design; allow longer if you need to measure traffic or untangle shared infrastructure. You’ll finish with a monthly run-rate estimate, a separate migration budget, and a record of what the test deployment did—and didn’t—prove.

Illustrative image: cloud computing costs

You’ll need:

  • Access to the workload’s server, storage, backup, and network metrics
  • A proposed cloud region and basic architecture
  • A spreadsheet and your provider’s pricing calculator
  • Permission to view billing data, set alerts, and run a limited test
  • Someone who can confirm performance and recovery requirements

If terms such as virtual machine, object storage, and network gateway are unfamiliar, start with how cloud servers, storage, and networks work together. You need to know which components your design uses before you can price them.

1. Set the boundary of the workload

Write down what the estimate includes before collecting prices. For a small web application, that might mean two application servers, one database, a load balancer, backups, monitoring, and traffic to users. Record the proposed region, hours of operation, availability requirements, and expected growth period.

Separate the monthly cost of running the workload from the one-time cost of moving it. The migration budget may include data transfer, tooling, testing, staff or contractor time, and a period when old and new systems run together. Mixing those expenses into one monthly number makes both harder to judge.

Decide what is shared, too. If several applications use the same network connection or support plan, note how you’ll allocate its cost. Don’t quietly assign the whole charge to one workload—or leave it out of every workload estimate.

2. Measure what the existing system actually uses

Collect at least a few weeks of representative data, including a busy period. Record CPU and memory use, disk space, disk activity, database size, backup retention, and traffic volume and direction. For a seasonal business, a quiet-month average isn’t enough; capture a peak or model it separately.

Look for a shape, not just a total. A server that sits idle overnight may be scheduled to shut down if the application allows it. A production database that needs to be available around the clock cannot. Count test and staging systems separately, including the hours people actually leave them running.

If your application serves a website, the differences between shared hosting and a VPS can also help clarify whether you need a cloud design at all. A full cloud deployment adds flexibility, but it may add components and operating work that a simpler hosting option avoids.

3. Translate the measurements into a proposed design

Choose the type and number of cloud resources that could meet the workload’s needs. For each, record its size, region, operating system, expected running hours, and whether you need redundancy. Include separate rows for production, staging, and disaster recovery rather than treating them as one server.

Be careful with “like for like.” Eight virtual CPUs in an old server do not establish that a particular eight-vCPU cloud machine will perform the same way. Database licensing, disk speed, memory pressure, and short traffic spikes may drive the choice. Start with a plausible size, then mark performance-sensitive choices for the test.

Draw the traffic paths between components. This makes it much easier to spot services—such as load balancers, gateways, and private connections—that a list of servers alone would miss.

4. Price compute by running hours

For each machine or compute service, multiply the expected quantity by its billable hours and the applicable rate. A continuously running server is commonly modeled at about 730 hours for an average month; a weekday-only test server needs its own schedule. Account for the chosen region, operating system, and any software license charges.

Use the provider’s calculator rather than a rate copied from an old proposal. The AWS Pricing Calculator’s EC2 estimate lets you set region, instance details, usage, storage, and data transfer; the Azure pricing calculator uses expected consumption and can show agreement pricing for eligible signed-in users; and the Google Cloud pricing calculator lets signed-in users view billing-account pricing. Replace defaults with your measured assumptions.

Build the first estimate using the payment terms you could actually use at launch. Model a reservation or longer-term commitment as a separate comparison, not as an automatic discount: a cheaper rate is little help if the test shows you bought the wrong size.

5. Price storage and copies separately

List the primary data, attached disks, database storage, snapshots, backups, logs, and retained test data. Note both the initial amount and likely monthly growth. Storage charges can depend on provisioned capacity, stored data, requests, performance settings, or retrieval, depending on the service; use the pricing fields for the exact products you chose.

A 500 GB database with a 500 GB backup is not one 500 GB storage line. Nor should you assume that a stopped test server costs nothing: on AWS, EBS volumes continue to incur storage charges while an EC2 instance is stopped.

Set a backup-retention assumption now. If you are deciding how much of your data belongs in cloud storage rather than on local media, the trade-offs between cloud and external-drive backups deserve their own review. For this estimate, the essential question is how many copies you will keep, where, and for how long.

6. Map and price every meaningful data path

Estimate monthly GB for traffic to users, traffic between application components, cross-zone or cross-region replication, backup transfers, and connections to systems that will remain outside the cloud. Label each path by direction. “800 GB of transfer” is not enough information to select a rate.

Then check whether a path passes through a separately billed service. For example, an AWS NAT gateway has both an hourly charge and a per-GB processing charge; other applicable data-transfer charges may also matter. Internet egress, cross-zone traffic, and a gateway’s processing fee should not be collapsed into a single guessed rate.

Give uncertain traffic a range. If last month’s outbound traffic was 600 GB but a product launch could double it, price both cases. This is usually more honest—and more useful—than adding an unexplained percentage to every line.

7. Add support and always-on extras

Go back through the architecture for costs outside the main compute, storage, and transfer rows: load balancers, public IP addresses, gateways, monitoring, log ingestion and retention, security services, and licensed software. Add the support level the team would actually need, allocating an account-wide plan where appropriate. The AWS Pricing Calculator has a separate support estimate, for example.

Pay special attention to resources that remain allocated when nobody is using the application. On AWS, public IPv4 addresses can incur charges whether they are in use or idle. Record who owns test resources and when they should be removed; an “idle” line in the spreadsheet is a cost to manage, not an assumed discount.

8. Build a baseline and a high-usage estimate

Put the calculator results into a spreadsheet with columns for resource, quantity, unit, rate, monthly cost, and the source or date of the rate. Keep one-time migration expenses on another sheet. Save the calculator configuration so another person can inspect the exact sizes and options you selected.

Here is the arithmetic for a deliberately simplified, hypothetical workload—not a provider quote:

Monthly item Assumption Estimated cost
Compute 2 servers × 730 hours × $0.12/hour $175.20
Primary storage 500 GB × $0.10/GB $50.00
Backup storage 100 GB × $0.05/GB $5.00
Outbound transfer 800 GB × $0.08/GB $64.00
Gateway 730 hours × $0.04, plus 800 GB × $0.04 $61.20
Support allocation Estimated share $40.00
Total Before other applicable charges $395.40

The rates are invented solely to show the method. The example charges the same 800 GB for outbound transfer and gateway processing because it assumes that traffic follows both billable paths; your design might not. Add any missing services, applicable taxes, and contract terms before treating a result as a budget.

Duplicate the sheet for a high-usage case. Change the variables that could genuinely move—traffic, server count, storage growth, or running hours—and show the difference. That range gives the pilot something specific to test.

9. Set billing alerts before deploying

Create a budget for the pilot and send alerts to someone who can act on them. Set thresholds below the maximum you are willing to spend, then check what the budget includes: support or shared-account charges may not appear in a workload-scoped view.

Treat an ordinary alert as a warning, not a brake. Azure budget notifications do not stop resources or consumption, and Google Cloud alerts-only budgets do not cap spending. Some products offer separate spend-control options, but their coverage and effect must be checked for your services. Billing data can also arrive late, so alerts cannot replace a deployment limit and an owner watching the test.

10. Run a controlled test deployment

Deploy the smallest version that still exercises the proposed design: the intended machine type, storage class, network path, logging, and backup setup. Tag or otherwise isolate its resources so you can identify its charges. Give the test a fixed end time and avoid production data unless its use has been approved and secured.

Generate representative work, not merely an idle server for an afternoon. Test ordinary demand and a short peak; move enough data to expose network paths, while keeping the test within its approved cost limit. Record what ran, for how many hours, and how much data moved. Without those measurements, a small pilot bill cannot reliably predict a full-month bill.

11. Reconcile the estimate with the charges

Once the provider has processed the test usage, compare billed units with your recorded units line by line. If compute cost differs, check instance size and running hours. If network cost differs, inspect traffic direction, zones, and gateway processing. For storage, check provisioned disks, snapshots, and logs. Separate a wrong usage assumption from a wrong rate.

Don’t multiply a short pilot’s total by 30 and call it the monthly forecast. Hourly resources, backup schedules, growth, peak traffic, and shared support charges may not scale that way. Update the relevant quantities in the spreadsheet instead, and mark charges that the pilot was too small or too short to validate.

12. Remove the test resources

Delete the pilot’s resources after you have the measurements you need, following your organization’s data-retention rules. Check disks, snapshots, public IP addresses, gateways, and logs as well as compute. A stopped server is not the same as a cleaned-up deployment.

The estimate has worked when another person can reproduce its assumptions, every major billable component has a row, and the pilot’s differences have explanations. If charges still appear that nobody can map to the design, keep investigating before approving the migration.

For a regulated workload, a complex network, major software licensing, or a large business-critical move, bring in a cloud architect or specialist cost engineer. They can test performance and contractual assumptions that a calculator—and a small pilot—cannot settle on their own.