Skip to content
Velloreth
Technology 11 min read

Understanding Cloud Cost Models

On-demand, spot, reserved instances, savings plans, and egress fees — how each pricing model works, what it costs you in flexibility, and when to use which.

By

Cloud pricing is designed to look simple and behave otherwise. The list prices are public, the discount mechanisms are documented, and yet most teams overspend — not because the models are opaque, but because the choices interact. A commitment discount on the wrong instance family is waste. A spot fleet without interruption handling is an outage. An architecture that egresses a terabyte a day can dwarf every compute saving you make.

This post is about the models themselves: what each one buys, what it costs you in flexibility, and how to sequence the decisions.

The spectrum from flexibility to discount

Every major cloud (AWS, GCP, Azure) offers some version of the same ladder:

ModelDiscountCommitmentInterruptibleBest for
On-demand0% (baseline)NoneNoSpiky, unknown, short-lived
Spot / preemptibleUp to ~90%None (capacity only)YesFault-tolerant batch work
Reserved InstancesUp to ~72%1–3 years, specific configNoSteady, understood baselines
Savings PlansUp to ~66–72%1–3 years, $/hourNoSteady baselines with flexibility needs

The trade is always the same: you give up flexibility (or accept interruption risk) and the provider gives you a lower rate. The skill is matching the model to the workload’s actual behavior, not its aspirational behavior.

On-demand: the default for a reason

On-demand is pay-per-second (or per-hour) with no commitment. Stop the instance, stop paying. It is the right model for:

  • Development and staging environments.
  • New workloads whose traffic pattern you do not yet understand.
  • Genuinely spiky traffic that averages too low to justify a commitment.
  • Anything with a lifetime shorter than a few months.

The mistake is leaving production on on-demand indefinitely. A 24/7 instance that has run unchanged for a year is the textbook candidate for a commitment discount. The other mistake is the opposite: buying a three-year reservation for a workload launched last month. Commitments should follow evidence, not enthusiasm.

Spot: cheap capacity with a warning label

Spot instances are spare capacity sold at a steep discount — AWS advertises up to 90% off on-demand. The provider can reclaim them with short notice (about two minutes on AWS) when the capacity is needed elsewhere.

That makes spot excellent for workloads that can checkpoint and resume:

  • Batch data processing.
  • CI runners and build farms.
  • Rendering and transcoding.
  • Stateless web tiers behind a load balancer, if the scheduler handles interruption.
  • Machine learning training jobs with periodic checkpointing.

It makes spot a poor fit for:

  • Databases and anything with in-memory state that cannot be checkpointed.
  • Single points of failure with no replacement path.
  • Workloads where a mid-job interruption is worse than a slightly higher bill.

The architectural requirement is interruption handling: your system must treat instance loss as a normal event, not an exception. On AWS, capacity-optimized allocation and diversified instance-type pools reduce (but do not eliminate) interruption risk.

Reserved Instances and Savings Plans

Reserved Instances

An RI is a 1- or 3-year commitment to a specific instance configuration (family, size, region, optionally AZ). In exchange you get up to roughly 72% off on-demand. Standard RIs give the deepest discount and the least flexibility; Convertible RIs let you exchange the configuration later for a smaller discount.

RIs still matter where you need capacity reservation — an AZ-scoped RI guarantees you can launch that instance type in that zone, which savings plans do not provide.

Savings Plans

Savings Plans are the more modern mechanism and what AWS recommends for most compute today. Instead of committing to an instance, you commit to a spend level (for example, $3/hour) for 1 or 3 years.

Two types:

  • Compute Savings Plans — up to roughly 66% off, applying across any EC2 instance family, region, OS, and also to Fargate and Lambda. Maximum flexibility.
  • EC2 Instance Savings Plans — up to roughly 72% off, locked to an instance family in a region. Deeper discount, less flexibility.

The practical advantage over RIs: if you change instance families, move regions, or shift a workload from EC2 to Lambda, the plan keeps applying. With RIs, each of those changes requires an exchange or a new purchase — and utilization quietly drops.

The two numbers that make commitments work

A commitment has two independent health metrics:

  • Coverage — what percentage of your eligible usage is covered by the commitment. Low coverage means you are still paying on-demand for work that should be on the plan.
  • Utilization — what percentage of your commitment is actually used. Low utilization means you bought more than you need and are paying for nothing.

Aim for high coverage and high utilization simultaneously. If you cannot have both — because usage is dropping or shifting — you bought the wrong size of commitment, or you bought too early.

Egress: the bill people forget

Every major cloud charges for data transfer out to the internet. AWS S3 internet egress, for example, is roughly $0.09/GB for the first 10 TB (after a 100 GB monthly allowance), with the rate declining at higher volumes. A service transferring 10 TB a month to users pays roughly $900 a month in egress alone, before storage and request charges.

Two architectural facts follow:

  1. Architecture decisions dominate cost more than instance selection. Moving from a chatty API that returns full objects to one that returns trimmed payloads can save more than any reservation discount.
  2. Zero-egress storage changes the math. Cloudflare R2 charges $0/GB for egress (storage is $0.015/GB-month, with request charges beyond a free tier). For media, downloads, and public assets served to the internet, the storage-vs-egress split alone can make R2 an order of magnitude cheaper than S3 Standard — even though S3’s per-GB storage rate is only moderately higher.

A worked example, illustrative of the shape of the numbers (rates as commonly published through 2026; verify current pricing before designing around them):

Workload: 1 TB stored, 10 TB served/monthS3 StandardR2 Standard
Storage~$23~$15
Internet egress~$890$0
Request chargesSmallSmall
Total order of magnitude~$910~$20

The gap disappears when data never leaves the provider network — transfers from S3 to EC2 in the same region, or to CloudFront, are free. Egress optimization is an architecture exercise, not a billing exercise.

Storage and request costs are not one line

Object storage pricing has three meters, and teams usually watch the wrong one:

  • Storage — dollars per GB-month. Usually the smallest of the three for active data.
  • Requests — per-thousand or per-million operations. Class A (writes, lists) cost more than Class B (reads). A chatty application with millions of small GETs can accumulate real request spend.
  • Retrieval — in colder tiers (S3 Intelligent-Tiering, Glacier, R2 Infrequent Access). A “cheap” archive becomes expensive when you need to read it back in bulk.

Snapshot and backup storage has the same trap in a different costume: snapshots accumulate silently, and a year of daily database snapshots can exceed the cost of the database itself.

A sequence that works

The order of operations matters more than any single tactic.

1. Right-size before you commit. Look at actual CPU, memory, and network utilization over at least 30 days. Most overprovisioned instances are two sizes too large. A t3.medium doing 8% CPU work does not need a reservation — it needs a t3.small.

2. Turn off what should not run. Non-production environments on nights and weekends. Idle load balancers, unattached volumes, old snapshots, orphaned IPs. These are free savings with no architectural risk.

3. Push interruptible work to spot. CI, batch jobs, rendering — anything that can checkpoint. This is often the single largest percentage saving available, and it does not require a multi-year bet.

4. Commit the stable baseline. After rightsizing, measure your steady 24/7 usage for a month. That number is your savings plan or RI target. Start with a 1-year term unless you have high confidence in the 3-year outlook; the discount difference between 1 and 3 years is real, but so is the lock-in.

5. Attack egress and data movement. Cache more aggressively, trim API payloads, put a CDN in front, and evaluate zero-egress storage for public assets. This is architecture work, so it runs on a longer timeline than billing changes — start it early anyway.

6. Monitor coverage and utilization monthly. Commitments drift as traffic changes. A plan that was right-sized in January can be over- or under-committed by July.

Provider-specific notes worth knowing

  • AWS currently steers customers toward Savings Plans over RIs for most compute. Spot is mature and well-documented; capacity-optimized allocation is the default recommendation.
  • GCP offers committed-use discounts (CUDs) analogous to RIs, and preemptible VMs analogous to spot. Automatic discounts also apply for sustained use without any commitment, which changes the baseline math.
  • Azure has reserved VM instances and Spot VMs, plus savings plan–style offerings. Azure’s egress tiers differ slightly from AWS; model them separately rather than assuming parity.

Cross-cloud comparisons of “which is cheapest” are almost always wrong at the margins, because the workloads differ. Compare your specific workload on each provider’s calculator, not list-price abstractions.

Right-sizing in practice

Rightsizing sounds abstract until you look at a week of utilization graphs. The typical patterns:

  • A t3.large (2 vCPU, 8 GB) running at 5% CPU and 1.5 GB memory. A t3.medium or even t3.small handles the same load. On-demand, that is roughly a 50–75% reduction in the instance line; on a commitment, the absolute savings are larger because the commitment is smaller too.
  • A database instance sized for a peak that happened once, eight months ago. Vertical rightsizing of managed databases (RDS, Cloud SQL) has more friction than EC2/VM changes — storage and compute are coupled, and some changes require downtime — but the savings compound monthly.
  • GPU instances for workloads that do not touch the GPU. An surprising number of inference endpoints run on GPU hardware because someone copied a template. Check utilization before you reserve or commit anything.

The measurement bar is simple: at least 14 days of metrics, look at p95 utilization (not average, which hides the peak), and only downsize if p95 stays comfortably under the next size down. Overshooting into an instance that cannot handle the daily peak costs more in incident time than the savings are worth.

Multi-cloud cost notes, short version

Cross-provider comparisons are usually noise, but a few structural differences matter when you model a workload:

  • Egress structures differ. AWS tiers internet egress starting around $0.09/GB after a 100 GB monthly allowance. Azure and GCP have their own tiers and allowances. Zero-egress options like Cloudflare R2 exist specifically because egress is the line item where architectures diverge most in cost.
  • Sustained-use discounts. GCP applies automatic sustained-use discounts to eligible VM usage without any commitment, which changes the baseline against which a committed-use discount is measured.
  • Committed-use shape. AWS leads with Savings Plans ($/hour commitments); GCP uses committed-use discounts (resource-level commitments); Azure uses reserved VM instances with a marketplace of savings-plan-style offerings. Model each on its native terms rather than forcing one provider’s mental model onto another.
  • Support plans scale with spend. A support plan that was negligible at $5k/month can be a line item worth revisiting at $50k/month.

None of this replaces running your actual workload numbers through each provider’s calculator. It does prevent the classic error of assuming AWS’s egress tiers apply on Azure, or that a GCP sustained-use discount makes a CUD redundant.

The bill of materials mindset

Treat your cloud bill like a materials list, not a mystery:

  • Compute hours by purchase model.
  • Storage by tier, with age.
  • Data transfer out, by destination.
  • Request/operation counts.
  • Support plan and marketplace subscriptions (the quiet line items).

Review it monthly for thirty minutes with the breakdown open. Most teams that do this find one or two structural surprises in the first month — an egress pattern, a forgotten environment, a commitment that stopped being utilized — that pay for the habit many times over.

Cloud cost is not a procurement problem you solve once. It is an ongoing property of your architecture, your traffic, and your commitment choices. The models are simple. Using them well is the discipline.

Frequently asked questions

What is the difference between reserved instances and savings plans?
Reserved Instances commit to a specific instance family, region, and optionally availability zone for 1–3 years. Savings Plans commit to a dollar-per-hour spend level for 1–3 years; Compute Savings Plans apply across instance families, regions, Fargate, and Lambda, while EC2 Instance Savings Plans are narrower but cheaper. AWS generally recommends Savings Plans over RIs for most compute because they are simpler to keep utilized.
How much can spot instances really save?
AWS advertises discounts of up to 90% versus on-demand for Spot. The real number depends on the instance type and region, but 70–80% is common. The cost is interruption: AWS can reclaim the instance with roughly two minutes of notice when it needs the capacity back.
Why is my cloud bill high when my compute usage looks low?
Often it is egress, storage operations, or idle resources rather than instance hours. Data transfer out to the internet is a common surprise, as are unattached volumes, old snapshots, and load balancers with no traffic. Review the cost breakdown by service before optimizing compute.
When should a small team buy a savings plan?
After you have rightsized and have at least a few months of stable, predictable baseline usage. If your baseline is small or still changing quickly, on-demand plus scheduled shutdowns of non-production environments usually saves more than an ill-fitting commitment.