Technology

The Real Cost of Cloud Egress, Snapshots, and Idle Servers

Cloud bills hide their biggest charges in egress, snapshot chains, and idle capacity. Here is the math behind each line.

Anthony Dami

Anthony Dami

Technology & Culture

Nov 12, 20269 min read
The Real Cost of Cloud Egress, Snapshots, and Idle Servers

Somewhere in the first quarter after a product launch, two graphs start to diverge. Requests served keeps climbing at the rate the business planned for, and the cloud invoice climbs faster. When someone finally filters the bill by service, the top line isn't the compute fleet or the database. It's data transfer, a charge that never appeared in the architecture diagram because the diagram ended at the load balancer.

Cloud pricing is built so that the expensive parts are structural rather than accidental. Providers publish hourly and per gigabyte rates, but the money accumulates in three places most teams never model: bytes leaving the network, backup copies nobody meant to keep forever, and capacity that is paid for while doing nothing. Each of those has a different mechanism and a different lever, which is why generic cost cutting usually misses all three.

Egress is a departure tax with almost no ceiling

Uploading into AWS, Azure, or Google Cloud is free on every major platform. Getting data out is billed per gigabyte, tiered by monthly volume, priced by region, and quoted on a page several clicks away from the instance pricing that everyone compared before signing. AWS lists $0.09 per gigabyte for internet egress from us-east-1, Microsoft lists $0.087, and Google Cloud lists $0.12. Those entry rates have moved remarkably little since 2020, which is exactly why so many cost reviews concluded egress was a rounding error and stopped looking there.

The conclusion held while applications were small. It stopped holding as media got richer, as logs and telemetry multiplied, and as models began streaming responses instead of returning a single payload. A workload that transferred five thousand gigabytes a month out of Google Cloud in 2021 now routinely moves thirty thousand, and the rate hasn't changed at all. Pricing stayed flat while the quantity it multiplied against grew three to seven times past the original migration forecast.

ProviderListed internet egress10 TB per month100 TB per month
Amazon Web Services (us-east-1)$0.09 per GB$922$9,216
Microsoft Azure$0.087 per GB$891$8,909
Google Cloud$0.12 per GB$1,229$12,288
Cloudflare R2$0$0$0

Modeled at flat entry rates, before volume tiers or private pricing agreements.

Run the numbers at scale and the line stops being a rounding error. One terabyte of outbound traffic at nine cents a gigabyte is about $92. Push a hundred terabytes a month through the same meter and the annual figure lands near $110,600 before any negotiated discount, which is a fully loaded mid-level engineer in most Western markets, spent entirely on moving bytes that already exist on disk. The bytes aren't new value. They're the same value crossing a boundary, and the boundary is what's billable.

The surcharges sit on adjacent lines

Searching a cloud invoice for the words data transfer out to the internet will find one charge and miss four others. A request that crosses an availability zone is metered in both directions, so a chatty service mesh or an application tier that talks to a database in a different zone generates billable traffic that looks free inside the console. A NAT gateway adds a per gigabyte processing charge on top of its hourly fee, which turns a well intentioned move toward private subnets into a permanent transfer bill. Pulling the same object from a CDN origin bills the origin transfer before the CDN bills the viewer, and cross region replication of a bucket bills both the write and the later read.

None of these show up under the heading teams are taught to watch, and that is the design. The published rate is the visible price. The architecture determines how many times each byte gets billed before it reaches a person.

Europe has begun to put a ceiling on the worst of it. The EU Data Act requires providers to stop charging business customers switching fees by 12 January 2027, and the major hyperscalers have since published waiver programs for customers migrating their data out. A regulatory limit on exit charges is the first structural constraint egress pricing has faced in a decade, and it applies inside Europe to any company able to document where its data is going. Everywhere else, the meter runs uncapped.

Snapshots bill like version control until the archive tier rewrites the chain

The first snapshot of a volume copies every block. Every snapshot after it stores only the blocks that changed, which makes a nightly backup policy feel nearly free for months. That feeling is accurate right up until someone decides the older copies should move to the cheap archive tier, where the rate is $0.0125 per gigabyte month against $0.05 for standard snapshot storage. The discount looks like a four to one win on the rate card.

It isn't, because archiving a snapshot converts it from an incremental record into a full copy of the volume. A retention chain that was cheap precisely because it deduplicated across nights becomes a stack of complete images, each billed on every block in the volume. There is also a ninety day minimum on the archive tier, with a pro-rated charge if you delete or fully restore early, and a one time restore fee of $0.03 per gigabyte when you need the data back.

The mechanics compound. For a one terabyte volume with two percent of its blocks changing daily, a thirty night retention policy stores the base copy plus twenty nine increments, about 1,618 gibibytes in total, which bills near $81 a month at the standard rate. Archive all thirty nights instead and you are billed for thirty full copies of 1,024 gibibytes at the archive rate, roughly $384 a month, and a single restore still costs about $31 on top. The cheaper tier costs 4.7 times more, and the same invoice would show a lower per gigabyte rate the whole time.

What you doWhat gets billedMonthly cost
30 night chain, standard tier1,024 GiB base plus 29 increments of 20.5 GiB at $0.05$81
Same 30 nights, each archived30 full copies of 1,024 GiB at $0.0125$384
Restore one archived snapshot1,024 GiB at $0.03 per GB, one time$31
Delete a middle snapshotNothing, if newer snapshots still reference those blocks$0

That last row is the one that catches people during cleanup projects. Deleting a snapshot in the middle of a chain can free nothing at all, because later snapshots still point at the blocks it introduced. Storage only falls when the last reference disappears, which means a retention policy built on age alone will quietly hold data far past the date on the spreadsheet.

Idle capacity is the largest bucket and AI refilled it

Flexera's fifteenth State of the Cloud Report, published in March 2026, puts self reported waste at 29 percent of infrastructure and platform spend, up two points from the year before and the first increase since the firm started tracking the number downward. The report attributes the reversal to AI workloads that bill in unpredictable bursts, and it found that 76 percent of large enterprises now spend more than $5 million a month on public cloud while 17 percent blew through their budget entirely. Idle resources and overprovisioned instances account for the bulk of the waste.

The mechanism is that idle compute is invisible because it is healthy. An instance running at eight percent utilization still passes its health check, still answers a ping, and still receives its full hourly charge. Autoscaling policies keyed to CPU thresholds never fire on workloads whose load arrives as short spikes, Kubernetes requests get set once during a migration and never revisited, and a GPU reservation bought to serve a ninety minute daily peak bills for all 1,440 minutes in the day. Add AI inference that arrives in bursts with no historical baseline and the forecast stops being useful within a quarter.

What actually changes the number

  • Measure egress per active user or per thousand requests, not as one aggregate line, so the growth rate becomes visible before the invoice does.
  • Move static assets and model weights to an object store that charges nothing for egress, and keep the compute that serves them close to that store.
  • Build snapshot lifecycles around the chain rather than the calendar, and test a restore before trusting the archive tier math.
  • Right size on utilization percentiles rather than averages, then schedule shutdown for anything with a predictable quiet window.
  • Put a budget alarm on data transfer specifically, since general spend alarms get normalized within two months.

For inference workloads the placement decision matters more than the price negotiation, and the same economics that pushed teams toward open weight models also pushed them to keep the model and its training data on the same side of a network boundary. Moving a checkpoint across regions on every restart is a recurring egress charge dressed up as an operational inconvenience, and it is one of the few costs that scales with how often things go wrong.

The bill is an architecture review that arrives monthly in dollars. Egress measures how much state has to travel, snapshots measure how long old decisions linger, and idle capacity measures how badly a team can predict its own demand. None of those are accounting problems, so none of them get fixed by renegotiating a rate card. The commercial question worth watching is whether buyers of AI capacity start demanding contracts that meter finished outputs instead of bytes in flight, because as long as transfer is billed per gigabyte, a provider's cheapest route to revenue growth is the fastest growing line on a customer's invoice.

Strategic Takeaways

  • ✓At AWS's $0.09 entry rate for internet egress, a hundred terabytes a month costs about $9,200 before volume tiers, and application data volumes have grown three to seven times past what most migration plans assumed.
  • ✓Archived EBS snapshots look four times cheaper per gigabyte, but archiving converts an incremental chain into full copies, so a thirty night retention policy costs about five times what the same chain costs in the standard tier.
  • ✓Flexera's 2026 State of the Cloud Report puts wasted cloud spend at 29 percent, the first increase in five years, and idle plus overprovisioned compute makes up most of it.
Anthony Dami
About the Author

Anthony Dami

Technology & Culture writer at Largern, focusing on AI, consumer hardware, film, internet culture, and the routines and tools behind everyday life.

Comments (0)

Join the Discussion

No comments yet

Be the first reader to share your perspective on this story.

Related Coverage