Operations & cost
If you are searching for cloud costs or staring at an AWS bill that feels too high, you are not alone. Most teams we meet do not have a "wrong cloud" problem — they have unclear ownership of what is running, who may start it, and what should be shut down again.
We see it on AWS, on Vercel and on Cloudflare. The platform differs; the pattern does not: small, invisible decisions that each feel cheap until they land on the same invoice. This guide is the frame we use when the question is not "should we switch clouds", but "why is the bill rising, and what stops the rise".
TL;DR
Key takeaways
- Growth rarely explains the whole increase — egress, idle compute and forgotten environments do.
- Unclear ownership is more expensive than the wrong instance size: what has no owner does not get turned off.
- Observability without cost tags is theatre — you need to point to a team or product per line.
- Reserved capacity and commits only help once you know the baseline — otherwise you lock in noise.
- Moving to a "cheaper" platform does not fix a governance problem; it relocates it.
- Start with an honest week of inventory before you talk about architecture changes.
Why the cloud bill feels unpredictable
On-prem was expensive in a predictable way: you bought capacity and it sat there. Cloud is cheap in an unpredictable way: you pay for what you forget to turn off. That difference is why many CFOs feel cloud "ran away", even when the product has not grown much.
Three mechanisms drive it. First, creation is frictionless — a preview database, a staging cluster, an extra region takes minutes. Second, deletion is rarely anyone's job. Third, the invoice lags the decision: you see the rise weeks after someone "temporarily" started something.
That is why switching vendor is the wrong first move. AWS, Vercel and Cloudflare price differently, but all three reward discipline and punish forgetfulness. If you do not know which resources belong to which product, a new invoice will just look confusing in a different console.
The goal is not the lowest possible bill. The goal is a bill you can explain: this line is product X, that line is staging you chose to keep, and that line is something you are shutting down this sprint.
The five drivers we see again and again
Egress is the quiet killer. Data leaving a region, a CDN or a database costs — and it is easy to design into chatty APIs, large assets without caching, or backups copied across cloud boundaries without anyone noticing. When traffic grows, egress grows linearly while you may have assumed "flat" hosting.
Idle compute is next. Dev environments running over the weekend, GPU instances waiting for jobs, Kubernetes nodes overprovisioned "just in case". Cloud makes it tempting to leave things on; the economics punish it. Autostop, schedules and aggressive TTLs on temporary environments are boring work — and often the highest-impact first move.
Storage without lifecycle is third. Snapshots, logs, old object buckets, orphaned volumes. It feels free until it is not. A lifecycle policy is not optimisation — it is hygiene.
Fourth is managed services without usage thinking: managed databases sized too large, always-on queues, more Redis than you need. Managed is right when it saves ops time. It is wrong when it is a shortcut to avoid designing load.
Fifth — and most important — is missing tags and ownership. Without cost allocation tags, project names or team labels you cannot act. You can only wonder. Any cost control that starts without inventory ends in guesswork.
AWS, Vercel and Cloudflare — different bills, same pattern
On AWS the classics are underused EC2/ECS tasks, NAT gateways, cross-AZ data transfer, and RDS instances that never scaled down after launch. The billing console can show you this — but only if tags are consistent. Without tags, Cost Explorer is a museum of unexplained lines.
On Vercel the rise often looks different: build minutes, serverless invocation spikes, bandwidth, and image optimisation. That is not "AWS in disguise". It is a platform where frontend traffic and preview deploys are visible. If marketing runs heavy campaigns, or you regenerate too many pages without cache discipline, the bill rises — even when the backend is cheap.
On Cloudflare, Workers, R2 egress policies and image/video pipelines can surprise you if you assume edge is free. Edge is often cheaper than origin for the right workloads — but "move everything to the edge" without measuring request patterns is still a guess.
The point is not to rank the platforms. The point is that each platform has a small set of knobs that explain 80% of the rise. Find those knobs before you redesign the architecture.
Symptoms vs causes
What you typically see
The symptoms that make someone open a ticket.
- The invoice is higher than last quarter with no clear explanation
- A team says "cloud got expensive" without pointing at a service
- Staging looks like production in size
- Extra regions "because we might need them"
- Nobody can say what the most expensive line belongs to
What is typically the cause
The underlying drivers behind the symptoms.
- Egress and chatty integrations without cache
- Idle environments without autostop or an owner
- Storage and snapshots without lifecycle
- Overprovisioned managed services
- Missing tags, budgets and alert thresholds
A week that usually creates control
Not a transformation programme — a concrete working week before you talk bigger architecture.
- 01
Inventory without polish
List compute, databases, storage, CDN and third-party SaaS billed through cloud. Mark owner, environment and product. Anything without an owner is a shutdown candidate.
- 02
Find the top 10 lines
Sort by cost. For each line: is it production, staging, or forgotten? If you cannot answer within an hour, it is a governance problem — not a capacity problem.
- 03
Turn off the safe things first
Forgotten snapshots, old preview environments, idle GPUs, unused load balancers. Capture the decision in a short note so people dare to act.
- 04
Add tags and budgets
Introduce minimum tags (team, environment, product). Set budget alerts at the level where someone actually responds — not only at the whole-org level.
- 05
Only then: architecture
Now you can talk cache, region strategy, reserved capacity or platform change with data. Before that it is guesswork dressed as strategy.
“The most expensive cloud resource is the one nobody thinks is theirs. Cost control is ownership — not another prettier dashboard.”
When an architecture change is actually the answer
Sometimes the bill is high because the architecture is wrong for the workload. Chatty microservices across AZs, synchronous file processing in the request path, or a database used as a queue. Then inventory alone will not help — you need to change the design.
The signs: the most expensive lines clearly belong to the core product, they grow with traffic, and you can explain them. That is not forgetfulness. That is a trade-off you chose (or inherited), and it needs renegotiation.
Here hybrid is often the honest answer: keep frontend and cache close to users, move heavy jobs onto queues, and avoid paying egress for data you could have kept closer together. That is not a religion about AWS vs edge — it is placing work where it is cheapest to run.
If you are considering switching the whole platform to save money, ask first: will the new platform make ownership clearer, or just hide the same habits behind a different price list? Habits travel with you.
Cost control now — or architecture later?
Left is the situation — right is what we typically recommend as the next step.
Situation
You cannot name an owner for the most expensive lines
Next step
Inventory, tags and shutdown before any platform talk
Situation
Staging and previews match production in size
Next step
Autostop, smaller sizing and TTLs — not reserved capacity
Situation
Egress grows linearly with a known integration
Next step
Cache, batching or placing workloads closer to the data
Situation
Baseline is known and usage is stable
Next step
Consider commits/reserved only on the stable portion
Situation
Core product traffic drives the bill in an explainable way
Next step
Architecture review: queues, cache, region — not just "right-size instances"
Checklist before you declare cloud "too expensive"
If you are missing several of these, the problem is governance — not necessarily the platform.
- Every production resource has a named owner and a product tag
- Idle and preview environments have autostop or an expiry date
- Storage has lifecycle rules for logs, snapshots and cold data
- Budget alerts exist at team or product level
- You can explain the top 10 cost lines in under an hour
- You have separated growth-driven cost from forgetfulness-driven cost
Questions we hear again and again
Is AWS just more expensive than Vercel or Cloudflare?
Not automatically. AWS is broader and therefore easier to misuse at scale. Vercel and Cloudflare can also get expensive if traffic, builds or edge workloads grow without cache discipline. Compare the concrete workload — not the vendor brand.
Should we go multi-cloud to avoid lock-in and save money?
Rarely as a first step. Multi-cloud increases complexity and often egress between systems. Portability in the app layer is smarter than running two clouds "just in case" while you still lack tags on one of them.
How do we know if the increase is healthy growth?
If cost follows a business metric you understand — orders, active tenants, rendered pages — and you can point to the service driving it, it is typically healthy. If the bill rises while those metrics stay flat, it is forgetfulness or architecture.
Do FinOps tools alone help?
They help with visibility. They do not replace decisions. Without a person who is allowed to turn something off, dashboards become decoration. Start with ownership and a shutdown routine; add tooling once the routine exists.
When does switching platform make sense to save money?
When you have a known baseline, can explain usage, and the new platform clearly cheapens that specific workload — not because someone promised "cheaper cloud". Switching without governance just moves the same bill into a new console.
If the bill rose without a clear explanation
Let us find what actually costs.
A short walkthrough of your top lines is often enough to separate growth from forgetfulness — before you rebuild or switch platforms.