An online box office sells 800 tickets on a normal day, then 40,000 in one hour on opening night. A newsroom publishes an investigation on Sunday evening and traffic multiplies by ten. A B2B SaaS doubles logins after a viral LinkedIn integration. None of these projects "needs cloud" all the time. They need occasional elasticity — and that is when usage billing stops being hype and becomes a tool.
Stable load vs unpredictable load: two cost logics
| Load profile | Often optimal product | Why |
|---|---|---|
| Near-constant traffic | VPS, dedicated, managed shared | Fixed price, no bill surprise |
| Short violent spikes | Cloud + targeted autoscaling | Pay surplus for a few hours |
| Predictable seasonality | Fixed server + manual scale-up | Known Black Friday → plan ahead |
| Unpredictable growth | Cloud with budgets and alerts | Avoid rebuying hardware every quarter |
Cloud is not inherently "faster". It is billed by the minute and provisioned on demand. Its price is justified when the annual average under-uses an oversized server.
Running cloud 24/7 for VPS-level load is a mistake. Running a fixed VPS to absorb a ×50 spike is another.
When cloud actually earns its price
Four scenarios where elasticity outweighs the cloud premium:
1. Peak amplitude >> average. You use 2 vCPU steady and 32 vCPU two days a year. A 32-core dedicated machine wastes money; temporary cloud scale-up can be rational.
2. Uncertainty window. Product launch, funding round, TV campaign: you do not know the ceiling. Scaling beats saturating.
3. Disposable environments. Load preprod, resilience tests, CI sandboxes: create and destroy resources instead of buying idle hardware.
4. Occasional multi-region. A local event needs temporary presence without a permanent datacenter.
Invoice lines that surprise teams
Poorly bounded unpredictable load produces unpredictable bills:
| Line item | Risk | Mitigation |
|---|---|---|
| Egress | Misconfigured CDN, large files | Edge cache, compression, included bundles |
| Snapshots / backups | Long retention, high frequency | 7/30-day policy, lifecycle |
| Load balancer | Always-on instance | Remove outside prod if possible |
| Public IP / NAT | Service multiplication | Consolidate, IPv6, architecture |
| Object storage | Unpurged logs | Lifecycle rules, quota alerts |
Ask for an estimator with your worst-case scenario before production cutover. Many teams discover egress after the first big spike — too late to renegotiate architecture.
Cloud without autoscaling: half the promise
Autoscaling only helps if you have:
- reliable metrics (CPU, latency, queue length);
- thresholds tested in preprod;
- a scale-down plan so 20 instances do not linger after the spike;
- budget caps and alerts at 50% / 80% / 100%.
Otherwise cloud becomes an expensive VPS with a complex UI. For known seasonality (Black Friday), documented manual scale-up sometimes beats miscalibrated autoscaling.
The peak: cloud buys optionality, not magic
Sales teams sell elasticity as universal insurance. FinOps reminds you permanent insurance has a price. A mature decision crosses load curve, cost of downtime, and ops skill to steer scale.
Decide and move forward without blind spots
Plot average and peak load over twelve months — or estimate at least three credible scenarios if history is missing. Then simulate the cloud bill including egress, storage, and snapshots, and compare that total to a correctly sized fixed server with room to grow. Set budget alerts before the first production deployment, and document a manual scale-up procedure even if autoscaling is planned.
Explore cloud offers in our directory and the guide Preparing WooCommerce for Black Friday for a concrete seasonal spike case.
Frequently asked questions
Is cloud always more expensive than a VPS?
No. At flat load, fixed wins. Cloud wins when short spikes make year-round oversizing more expensive.
Which hidden cloud costs should you watch?
Egress, storage, snapshots, IPs, load balancers. Set caps and alerts.
Do you need autoscaling from day one?
Only with tested metrics and thresholds. Otherwise start with documented manual scaling.
Public cloud or sovereign European cloud?
Compare total cost and latency first; add legal constraints as a secondary filter.
Cloud is not a religion. It is a billing option — profitable when unpredictability costs more than elasticity.