What is elasticity in cloud computing, and what does it save?

What is elasticity in cloud computing, and what does it save?

· Updated

Aron Wagner

Aron Wagner

CEO & Co-Founder

Elasticity means capacity follows load: machines get added as traffic climbs and removed as it drops, usually automatically. So you pay for your average load and can still serve your peak. What that's worth depends on how far your peak sits above your average: a flat load saves nothing, and a spiky one can cut compute by more than half.

What does elasticity mean in cloud computing?

NIST lists "rapid elasticity" as one of the five essential characteristics of cloud computing: capabilities "can be elastically provisioned and released, in some cases automatically, to scale rapidly outward and inward commensurate with demand." If you've seen "elastic computing" or "elastic cloud", they mean the same thing.

Speed is why anyone cared. The 2010 paper A View of Cloud Computing describes the change: adding or removing resources "at a fine grain (one server at a time with EC2) and with a lead time of minutes rather than weeks."

There are two ways to add capacity. Scaling out and in adds and removes machines. Scaling up and down swaps a machine for a bigger or smaller one. Elasticity usually means the first: a new machine joins without restarting the ones already serving traffic.

What does elasticity save?

Take a managed Kubernetes cluster at our list prices. It needs 2 workers (a worker is a VM that runs your containers) for most of the day and 8 for 4 hours in the evening. Size it for the peak and you run 8 workers around the clock: $766.80 a month. Let it scale and the same month costs $295.05.

[ Workers in one day ]
Workers running across one example day, elastic against sized for the peakElastic: 2 workers for 20 hours and 8 workers from hour 18 to hour 22, costing $295.05 a month. Sized for the peak: 8 workers all 24 hours, costing $766.80 a month. The average of the elastic day is 3 workers.06121824Hour of the day0238Sized for the peak: $766.80Average: 3 workersElastic: $295.05
Workers are 4 vCPU, 8 GB, 200 GB on Standard (Intel), billed by the hour for a 720-hour month, with one $12 load balancer and the free first control node. List prices on November 2, 2026.

The 6 extra workers run 4 hours a day for 30 days. That's 720 worker-hours, exactly one worker for a 720-hour billing month: $94.35. So the elastic bill is 3 workers plus the load balancer, the same as a fixed 3-worker cluster. The fixed one just can't serve the evening. That's the deal: the bill of your average, the capacity of your peak.

Here's the rule for your own traffic: on hourly billing, the compute saving is 1 minus your average divided by your peak. Here that's 1 minus 3/8, or 62.5%: $471.75 off $754.80 of workers. That 2010 paper says that for many services "the peak workload exceeds the average by factors of 2 to 10", which by this rule is a saving of 50% to 90%. To get your numbers, write down how many machines you'd need for each hour of a normal week, then take 1 minus the average divided by the highest.

Treat the result as a ceiling. You'll keep a minimum and some headroom, real traffic curves are rounder than this example, and workers take time to arrive and leave.

The saving also shrinks as the peak gets longer. Each extra 4 peak hours a day adds $94.35, until the two bills meet at 24 hours.

[ Bill by peak hours ]
The elastic bill rises in a straight line with peak hours until it meets the sized-for-peak bill at 24 hoursMonthly bill for a cluster of 2 workers scaling to 8. 0 peak hours: $200.70. 4 hours: $295.05. 8 hours: $389.40. 12 hours: $483.75. 16 hours: $578.10. 20 hours: $672.45. 24 hours: $766.80, the same as running 8 workers all the time.$0$400$80004812162024Peak hours a daySized for the peak: $766.80$200.70$295.05$389.40$483.75$578.10$672.45
The same cluster as above, 2 workers scaling to 8.

So look for a low baseline and a short, tall peak. Even David Heinemeier Hansson, who took 37signals off the cloud, makes an exception for load where "the baseline is a sliver of your largest needs": "There's nothing like the cloud when that happens."

How does cloud elasticity work?

An autoscaler watches load. A rule you set, within a minimum and maximum, adds machines when load climbs. A load balancer spreads traffic across them. When load drops, the rule removes machines, and you stop paying for each one once it's gone.

Kubernetes has two layers of this. The Horizontal Pod Autoscaler adds copies of your app, called Pods. By default it checks every 15 seconds and waits 300 seconds before scaling Pods down, so the count doesn't flap. When the new Pods don't fit on the workers you have, node autoscaling adds workers, and workers are what's on your bill.

What is the difference between elasticity and scalability?

Scalability is whether a system can grow. Elasticity is whether it grows and shrinks with demand on its own, and it needs scalability first. A cluster that grew to 8 workers and stayed there is scalable. It isn't elastic: it's still billing for 8 at 3 a.m.

Agility gets mixed in with these. It's how quickly you can build, test and launch things, not how much capacity you're running.

TermWhen it happensWhat it changes
ElasticityMinutes, all dayYour capacity and your bill, hour by hour
ScalabilityAs the business growsThe most load you can handle
AgilityEvery releaseHow fast you ship

When doesn't elasticity work?

Three things get in its way.

Spikes faster than boot time. A new VM can take minutes to be ready, and a freshly started process runs slower than one that's warmed up. A spike that arrives in 30 seconds outruns both. If you know when it's coming, raise the minimum first: a sale that goes out at 9:00 wants its workers by 8:50.

State. A database can't lose a machine the way a web tier can, and Kubernetes won't delete a StatefulSet's volumes when you scale it down. Sessions kept on one machine have the same problem. Put state in a managed database or object storage, and make the rest elastic.

Overload. When something behind your app is already saturated, more copies send it more traffic and client retries add to it, so autoscaling can make that worse. The maximum is what stops it: a limit of 8 workers holds this cluster's compute to $754.80 for a 720-hour month, even when a bug is driving the load.

Does a traffic spike raise your bandwidth bill?

Not on American Cloud. We don't charge egress, so compute is the only line that moves with a spike. On AWS, 5,000 GB out of US East in a month is $441, at $0.09 a GB past the free 100 GB, and autoscaling doesn't lower it: that bill follows your traffic however you scale.

How do you set up elasticity on American Cloud?

Turn on autoscaling for a managed Kubernetes cluster and set a minimum and maximum worker count; our MCP server's scaling tool takes the same range. Nodes bill by the hour for the hours they run, there's no prepaid term to give up, and the first control node is free. For VMs, auto-scaling pairs with a load balancer at $12 a month.

VMs do give you a choice of billing. A 4 vCPU, 8 GB, 200 GB VM is $92.50 on a monthly term and $94.35 by the hour for a full month. Keep the VMs that never get deleted on the monthly term. Hourly wins for any VM that's deleted for more than about 14 hours of a 720-hour month, which is nearly every machine that scales in.

To size your own cluster, try the pricing calculator.

Contact us