Updated 6 Jul 2026 • 8 mins read
Cloud Pricing Models Explained and How to Choose
Cloud providers sell the same compute under very different pricing models: on-demand, reserved and committed-use, savings plans, spot, serverless consumption, tiered storage, and now token-based AI pricing. This guide explains how each cloud computing pricing model works, what it rewards, and how to combine them into a portfolio.
The most expensive misunderstanding in cloud computing is treating it as one product with one price. In reality, the same virtual machine can cost full rate, thirty percent of full rate, or ten percent of full rate depending on how you buy it, and the gap between those numbers is decided by pricing-model literacy, not engineering skill. Providers are not hiding this; they publish every model. Most teams simply default to the most expensive one.
This guide explains every major cloud computing pricing model, what each rewards and punishes, how the newest one, token-based AI pricing, breaks the old intuitions, and how mature teams combine models into a portfolio rather than picking a favorite.
Key takeaway Cloud pricing models trade flexibility for discount. On-demand is the ceiling: maximum flexibility, zero discount. Commitments (reserved instances, committed-use discounts) cut up to about 72 percent for one-to-three-year promises; savings plans cut up to the mid-sixties percent with more flexibility; spot capacity cuts up to about 90 percent in exchange for interruptibility; serverless bills pure consumption; and AI services bill per token, scaling with usage rather than infrastructure. No single model wins: the right answer is a portfolio, commitments under the stable baseline, on-demand and spot above it, reviewed quarterly.
The One Principle Behind Every Model
Every cloud pricing model is a position on one trade-off: how much flexibility you give up determines how much discount you receive. Promise nothing and pay list. Promise a year of usage and pay substantially less. Accept that your capacity can be reclaimed with minutes of notice and pay a fraction. Once you see the models as points on that single curve, choosing between them stops being a catalog exercise and becomes a matching problem: pair each workload's real flexibility needs with the model that prices them.
On-Demand: The Default and the Ceiling
On-demand (pay-as-you-go) bills resources by the second or hour with no commitment: start anything, stop anything, pay for exactly the interval between. It is the right model for spiky, experimental, and short-lived workloads, and the wrong model for anything that runs all year, because it is deliberately the most expensive rate on the sheet. Every other model is defined by its discount from this one, which makes unmanaged estates easy to diagnose: they are the ones paying ceiling prices for basement predictability.
Commitment Models: Reserved, Committed Use, and Savings Plans
Commitment models convert forecast confidence into discounts, and each provider implements the idea differently.
| Model | You commit to | Discount (up to) | Flexibility |
|---|---|---|---|
| AWS Reserved Instances / EC2 Instance Savings Plans | Specific instance configuration or family, 1 or 3 years | About 72% | Low: exact-match rewards, changes hurt |
| AWS Compute Savings Plans | Dollars per hour of compute, 1 or 3 years | About 66% | High: applies across EC2, Fargate, Lambda |
| Azure Reservations | Resource configuration, 1 or 3 years | About 72% | Low to medium, with exchange options |
| Azure Savings Plan for Compute | Dollars per hour, 1 or 3 years | About 65% | High across eligible compute |
| Google Committed Use Discounts | Resources or spend, 1 or 3 years | Roughly 55–70% by resource and term | Varies: resource-based deeper, spend-based flexible |
Two truths govern all of them. Commit to your stable baseline, never your peak: undercommitting leaves discount on the table, but overcommitting buys shelfware you pay for whether used or not. And treat commitments as a living portfolio, purchased incrementally, tracked for coverage and utilization, and rebalanced quarterly, which is exactly the practice our discount manager guide lays out.
Spot and Preemptible: The Deep Discount
Spot (AWS, Azure) and Spot or preemptible VMs (Google) sell spare capacity at discounts up to about 90 percent, with one condition: the provider can reclaim it on short notice. That makes spot spectacular for interruptible work, batch processing, CI builds, stateless and fault-tolerant services, rendering, much ML training, and unusable for anything that cannot die gracefully. The engineering price of admission is checkpointing, fast startup, and orchestration that replaces reclaimed capacity automatically; paid once, it converts a workload's fault tolerance directly into money.
Consumption Models: Serverless and Automatic Discounts
Serverless pricing removes the resource from the equation entirely: functions bill per request and per unit of execution time, serverless databases and warehouses bill per capacity-second or per data processed, and idle costs nothing. It suits variable and event-driven workloads and can invert at sustained high volume, where always-on capacity under commitment becomes cheaper, so the crossover point deserves an annual check. Google adds a related twist with sustained-use discounts, automatic price reductions on eligible resources as monthly usage accumulates, no commitment required, a reminder that the models differ enough across providers to move real money; our AWS vs Azure pricing comparison and AWS vs Azure vs GCP guide cover the cross-provider picture.
Storage Tiers and Data Transfer
Storage pricing rewards honesty about access patterns: hot tiers cost the most to store and least to access, and archive tiers invert that, with retrieval fees and minimum retention periods enforcing the bargain. Lifecycle policies that demote aging data automatically are the model working as intended. Data transfer is the pricing model nobody chooses and everyone pays: inbound is free, egress to the internet and traffic across regions and zones is not, and architecture, not negotiation, is the primary control. These two quiet models are where a large share of cloud waste accumulates.
The New Model: Token-Based AI Pricing
AI services introduced a pricing model with no infrastructure analog: per-token billing, where cost scales with the text a model reads and writes. Input and output tokens price differently, output typically several times input; model tier moves cost by an order of magnitude or more; and spend tracks product usage and prompt design rather than server count, which breaks every forecasting habit built on instances. The model comes with its own discount levers, prompt caching that cuts repeated input cost dramatically, batch processing at roughly half price, and provisioned throughput as the commitment analog, covered in depth in our token economics guide and AI cost optimization guide. Treat it as a first-class pricing model, because for many companies it is becoming a first-class line item.
How to Choose the Right Model per Workload
- Steady, predictable, always-on: commitments, sized to the baseline your usage history actually supports.
- Variable or unpredictable: on-demand, with autoscaling doing the flexing you are paying for.
- Interruptible and fault-tolerant: spot, with checkpointing and replacement automation.
- Event-driven or mostly idle: serverless, rechecked yearly against the always-on crossover.
- Data at rest: tier by access pattern, enforced by lifecycle policy rather than intentions.
- AI workloads: right-size the model tier, cache and batch aggressively, and commit only what usage history justifies.
Pricing Models and FinOps
Rate optimization, choosing and managing pricing models, is one half of the optimization pillar in FinOps, alongside usage optimization, and it is the half that finance can drive without touching a single workload. The prerequisite is the same as everywhere in FinOps: visibility and allocation first, so coverage, utilization, and effective savings rate can be measured per team and tracked as KPIs. Estimating under the right model also belongs at design time, not invoice time, which is what the provider calculators are for; start with our AWS pricing calculator guide if you are modeling a new workload today.
Conclusion
Every cloud computing pricing model is the same bargain at a different point: flexibility surrendered, discount received. On-demand buys freedom at the ceiling price, commitments and savings plans pay you for predictability, spot pays you for fault tolerance, serverless bills pure consumption, storage tiers bill honesty, and token pricing bills usage itself. The mistake is picking one and defaulting; the practice is running a portfolio matched to each workload's real flexibility and reviewing it quarterly. That portfolio needs measurement, coverage, utilization, and effective rates per team, across every provider and model you use, and giving you exactly that view is what Opslyft is built for.
FAQs
On-demand (pay-as-you-go), commitment-based models (reserved instances, committed-use discounts, and savings plans), spot or preemptible capacity, serverless consumption pricing, tiered storage pricing, data transfer pricing, and, for AI services, per-token pricing.
Per-second or per-hour billing with no commitment: start and stop freely and pay for exactly what runs. It offers maximum flexibility at the highest rate, making it right for variable and short-lived workloads and wasteful for anything steady.
Reserved instances commit to a specific resource configuration for the deepest exact-match discount, up to about 72 percent, but lose value when infrastructure changes. Savings plans commit to an hourly spend amount, discount slightly less, and apply automatically as workloads shift.
Up to about 90 percent versus on-demand across the major providers, in exchange for capacity that can be reclaimed on short notice. The savings are real for interruptible, fault-tolerant workloads and unavailable to anything that cannot restart gracefully.