The Cloud Bill, Line by Line: A Primer on What You Actually Pay For
Archi3 min read
Contents
The invoice is a map of your architecture#
A cloud bill looks like a phone book. Lines for compute, storage, network, databases, and a support charge. Each line maps to a decision someone made. The money question is simple: what does one customer, one order, or one request cost to serve? The invoice answers only after you split it.
Compute: hours, size, and idle time#
Virtual machines bill by the second or hour. A server that runs all month uses about 730 hours. At $0.10 per hour, that is $73 before storage and backups. Double the size, the rate can more than double. Teams often pick a size for peak load and leave it there overnight. The waste shows up as low CPU use, not as a line item called waste.
Containers hide the same math inside a cluster. A managed Kubernetes control plane may cost about $0.10 per hour, roughly $73 per month, before worker nodes. Each pod requests CPU and memory, and the scheduler reserves those requests even when the app sleeps. Money leaks through requests set too high.
Storage: cheap per gigabyte, expensive per habit#
Object storage is priced per gigabyte per month, plus requests and retrieval. A bucket with 10 TB at $0.023 per GB per month costs about $235. That sounds small until snapshots, old versions, and logs pile up. Block storage for databases costs more, and provisioned IOPS can cost more than the capacity.
Backups have a second bill: restore time. A cheap archive tier can add hours to recovery. The finance choice is not the lowest storage rate. It is the cost of downtime compared with the cost of fast disks.
Network: the line that surprises people#
Data going out to the internet is usually charged per gigabyte. One terabyte is 1,024 GB. At $0.09 per GB, that is about $92. Cross-zone traffic inside the same cloud can also carry a charge. So can a NAT gateway, which bills hourly and per gigabyte processed.
This is why media sites, API providers, and backup tools watch egress. A product with heavy downloads can lose money on a flat monthly price. Before pricing a plan, estimate outbound bytes per user.
The services around the server#
Load balancers bill hourly and by capacity units. Managed databases bill by instance size, storage, backup retention, and sometimes I/O. Observability tools bill by ingested gigabytes, hosts, or seats. A logging pipeline that keeps everything for a year can cost more than the compute that produced the logs.
Support is a percentage of spend on many plans. A 10 percent support charge on a $20,000 monthly bill is $2,000. That number changes when the bill changes, including when it rises because of a mistake.
Discounts are contracts#
Cloud providers sell savings plans and reserved capacity. The discount is real: 30 to 60 percent is common in sales conversations. The risk is a commitment to a shape of usage you may not need. A three-year reservation on a database family can become a bill for capacity you cannot use after a migration.
Spot instances cut prices further, but they can disappear with short notice. They fit batch jobs and test runners, not checkout. The finance team should see commitments as debt-like: a fixed obligation with a breakage cost.
Allocate before you optimize#
Tags and accounts decide whether the bill is manageable. A shared untagged cluster becomes an argument at the end of the quarter. Set a few rules: every resource has an owner, an environment, and a product tag. Put shared network and observability costs into a pool and split them by usage.
Then compute unit cost. Take the monthly bill, subtract one-time items, divide by orders shipped or active users. If the bill rises 20 percent and usage rises 25 percent, the unit cost fell. That is a better signal than the total.
A worked example#
A small API runs 6 virtual machines at $0.12 per hour, about $526 per month. Storage and backups add $180. A managed database adds $420. A load balancer and NAT gateway add $90. Egress for 2 TB adds $184. Logging adds $250. Support at 10 percent adds about $165. Total is roughly $1,815 per month.
Now change one habit: turn off three idle machines at night and weekends, about 60 percent of the month. Save $158. Move logs to 30-day retention and save $90. The total drops near $1,570 without touching the product.
Cloud money is not one bill. It is the sum of defaults. The defaults are set by engineers, approved by finance, and paid every month.
More articles