Unit 2.4: Cost Architecture and FinOps
Introduction
Most cloud cost is decided at design time, long before a FinOps report shows it. Choosing a compute class, a storage class or a way to scale the database fixes a cost curve that later tuning can only bend. This unit treats those choices as architecture, then adds the commercial layer: commitments, allocation and alerting.
All prices in this unit come from the IONOS CLOUD (EUR) price list, which applies to IONOS CLOUD GmbH and is the list a German customer such as FinCorp uses. Prices are net, per hour for compute and per 30 days for storage. The unit ends by creating FinCorp's first cost alert in the DCD.
1. Compute: Contention Is the First Lever
1.1 Dedicated Core Versus vCPU
The two Compute Engine classes differ in how the physical CPU is used. On a Dedicated Core Server, the CPU use is exclusive: IONOS CLOUD maintains dedicated resources for each customer, and you do not share your physical CPUs with other clients. On a vCPU Server, the CPU use is shared, and the CPU family cannot be chosen at creation or changed later.
The price list reflects that difference:
| Resource | vCPU Server | Dedicated Core Server |
|---|---|---|
| CPU, per hour | 0,012 € per vCPU | 0,036 € (AMD EPYC Milan) to 0,046 € (Intel Sierra Forest) per core |
| RAM, per GB per hour | 0,002 € | 0,0045 € |
| CPU family selectable | No | Yes, region-dependent |
| Covered by Cloud Savings Plans | No | Yes |
As shown above, for a 4-core, 16 GB server over 720 hours, a vCPU Server costs 57,60 €, while a Dedicated Core Server on an Intel family at 0,04 € per core costs 167,04 €. Shared CPU is the right default for tiers that tolerate contention: development, stateless web front ends, batch work. Dedicated cores are the right choice where predictable CPU or CPU-family control is part of the requirement (Unit 4.1).
One billing behaviour turns into a design rule. Because the resources are dedicated, a Dedicated Core Server switched off at the operating system level still incurs costs. Only a stop through the platform, in the DCD or with ionosctl compute server stop, deallocates the resources and stops billing for them, so scheduled shutdowns of non-production servers must call the platform, not the guest.
1.2 Over-Provisioning as a Control Cost
Regulated tiers often justify spending that a pure efficiency review would cut. FinCorp's payment-processing hosts sit within its compliance scope, so it runs them on dedicated cores, sized for the peak plus headroom, and does not autoscale them down to the last core. The excess is not waste; it is the price of predictable performance and a stable, auditable footprint. Record it as a control cost in the architecture decision record, with the control it supports, so that a later optimisation does not remove it by accident.
2. Storage: Pay for the Access Pattern
2.1 Block Storage Classes and the Performance Floor
The following table compares the three Block Storage classes. Prices are per GB per 30 days.
| Class | Price | Performance model | Typical use |
|---|---|---|---|
| Essential | 0,04 € | Static and independent of volume size: 1,100 IOPS regular, 2,500 IOPS burst, 200 MB/s sequential | Backups, archives, cold data |
| Balanced | 0,07 € | Scales with size: 40 read IOPS and 30 write IOPS per GB, up to 24,000 read IOPS per volume | Web servers, development |
| Performance | 0,15 € | Scales with size: 75 read IOPS and 50 write IOPS per GB, up to 45,000 read IOPS per volume | Databases, latency-sensitive applications |
Storage IOPS are free of charge, but on Balanced and Performance you buy them through capacity. The performance of those classes is directly related to volume size: IONOS CLOUD recommends volumes of at least 100 GB to get the full benefit of the SSDs, and volumes larger than 600 GB reach the maximum expected performance. A 40 GB Performance volume for a database is therefore a false saving. A 100 GB Performance volume costs 15 € per 30 days and delivers 7,500 read IOPS at 4 KB; reaching the 45,000 IOPS ceiling takes 600 GB.
2.2 Object Storage for Bulk and Archive
IONOS Cloud Object Storage costs 0,007 € per GB per 30 days, with no charge for API calls, writes or reads; data transmission is charged separately. Snapshots cost 0,04 € per GB per 30 days, the same as Essential. For archives, exported logs and datasets, Object Storage is roughly a sixth of the price of Essential Block Storage, and it is the target for the audit archive from Unit 2.3.
3. The Data Tier: Scale Up or Cache Out
In DBaaS for PostgreSQL, standby nodes do not serve read requests (Unit 1.3), so adding nodes buys availability, not read capacity. When reads saturate the primary, the two options are a larger instance or a cache in front of it. DBaaS for PostgreSQL and DBaaS for In-Memory DB share the same compute rates: 0,06 € per core per hour and 0,007 € per GB of RAM per hour.
The cost difference comes from where the capacity lands. Instance resources in a PostgreSQL cluster are set per instance, so scaling up a three-node cluster by 4 cores and 16 GB adds 0,352 € per hour on each of the three nodes, while only the primary serves the reads. A 2-core, 8 GB In-Memory DB instance costs 0,176 € per hour, half the increase on a single node, and takes repeated reads away from the primary. For read-heavy workloads with stable hot data, cache-out is the cheaper lever; for write-heavy workloads, it does not help, and scale-up remains the answer (Units 5.3 and 5.5).
4. Commitment: Cloud Savings Plans
4.1 How the Commitment Works
Cloud Savings Plans give lower hourly rates in exchange for a one-year or three-year commitment to a quantity of resources, not to specific machines. You commit to Cores and RAM separately, and a separate plan is created for each resource. Eligible products are Dedicated Core Servers, Dedicated Core Node Pools in Managed Kubernetes, and Nextcloud Workspace.
| Resource | Pay-as-you-go | 1-year plan | 3-year plan |
|---|---|---|---|
| 1 Dedicated Core, per hour | 0,036 € to 0,046 € by CPU family | 0,0340 € | 0,0240 € |
| 1 GB RAM, per hour | 0,0045 € | 0,0038 € | 0,0027 € |
The plan rate is one rate per dedicated core. The saving therefore depends on the CPU family the covered usage runs on: a plan saves more on an Intel Sierra Forest core than on an AMD EPYC Milan core.
The rules are strict, and each is a design input:
- Billing layer only. Plans operate on the billing layer, not the resource layer, so they do not reserve physical capacity.
- Flexible coverage. Plans apply across CPU models, sizes and regions, and continue to apply when you resize a VM or change its CPU family.
- Oldest first. Plan rates are applied starting with the oldest eligible VM in the contract; you cannot select which VMs are covered.
- Full billing. You are billed for the full plan quantity whether or not you use it, and unused resources do not roll over.
- Overflow at PAYG. Usage above the plan quantity is billed at the standard pay-as-you-go rate.
- No changes. Plans cannot be modified, cancelled or deleted after purchase; only the plan name can be edited. Payment is monthly; upfront payment is not offered.
- Expiry. You are notified 30 days before a plan expires. Plans do not renew automatically, and usage reverts to PAYG after expiry unless you have bought a follow-on plan.
Only the Contract Owner can create plans, and only for a contract in "Billable" status. Startup contracts do not qualify, and the contract's resource allocation limits cap the quantity you can configure.
4.2 Commit the Floor
The rules point to one discipline: commit only to the floor, the quantity that runs every hour of the term, and leave the peak on PAYG. Because several plans can stack and the oldest is applied first, grow the commitment in layers as the floor rises, instead of one large plan sized for a forecast. FinCorp's production floor is 48 dedicated cores and 192 GB RAM, running around the clock; it buys three-year plans for exactly that and keeps its autoscaled peak on PAYG.
5. Allocation: Showback and Chargeback
Three views serve different purposes:
- Cost & Usage in the DCD shows usage of all services at the currently applicable list prices, excluding discounts and VAT, grouped by product category, with VDC costs aggregated per VDC. Usage data is updated at least once a day, so the current period may lag by up to 24 hours, and reports cover up to the last six months. It is an estimate and may not match the final invoice.
- The invoice is the actual charge. Consumption at PAYG rates appears under its VDC, while consumption covered by a Savings Plan is moved out of its VDC and grouped under the plan's name.
- The Billing API returns invoice data and a resource-allocated usage distribution for integration into internal systems.
Use Cost & Usage for showback, where teams see their estimated consumption per VDC. Use the invoice or the Billing API for chargeback, where money moves between cost centres. Because Savings-Plan-covered usage leaves the VDC lines on the invoice, chargeback must redistribute the plan cost; FinCorp allocates it in proportion to each VDC's dedicated-core hours. The VDC-per-owner design from Unit 2.1 is what makes either view usable without manual tagging.
DCD Implementation Walkthrough
You will create a cost alert for FinCorp's production contract. It is the early-warning control that complements the monthly showback review.
Prerequisites: a Contract Owner or Administrator account. In the IONOS CLOUD multi-tenant model, the feature is available only to the contract owner of the respective multi-tenant contract. Decide the threshold first: the alert compares against the net amount of the contract invoice, excluding VAT.
Build goal: Create a cost alert against a budget threshold.
Steps (in the Data Center Designer):
- Go to Menu > Management > Cost alert.
- In the Cost alert window, select Create cost alert.
- Enter the threshold amount. FinCorp sets 80 % of its monthly production budget, so the warning arrives while there is still time to act.
- Enter the email address for the notification. Use a monitored team mailbox, not a person.
- Select Create cost alert. The alert is created and listed in the Cost alert window.
- To change the threshold or the address later, open Cost alert, select Edit, change the fields, and select Save.
Common mistakes:
- Expecting the alert to cap spending. It sends an email and does not execute any other operation, and in particular makes no changes to any service.
- Expecting repeated warnings. The email is sent only once, the first time the calculated invoice amount reaches the threshold, and is not repeated with later calculations.
- Setting the threshold at 100 % of the budget. Billing calculations run periodically, so an offset to the configured amount may be expected; leave headroom.
- Comparing the threshold with gross figures. The alert uses the net invoice amount, excluding VAT.
- Granting the feature to engineers. Users with the User role cannot access it; plan the recipient address instead.
For automation, the cost alert is part of the Billing API and is managed through the /{contractId}/ usage-alert endpoints. Accounts with multi-factor authentication must use a Bearer token from the Token Manager (Unit 2.2).
Summary
Cost architecture starts with the class choices: shared or dedicated CPU, the storage class and size that buy the needed IOPS, and cache-out before scale-up for read-heavy data tiers. Cloud Savings Plans then lower the rate for the always-on floor, under strict rules: billing layer only, oldest VM first, no changes after purchase. Allocation follows the contract and VDC boundaries from Unit 2.1, and a cost alert provides the early warning between reviews.
Key Points:
- Dedicated cores cost more than vCPUs and keep billing when the guest is shut down; deallocate through the DCD.
- On Balanced and Performance storage, IOPS scale with size; at least 100 GB is recommended, and 600 GB reaches the maximum.
- PostgreSQL standbys do not serve reads, so a cache is usually the cheaper read-scaling lever.
- Savings Plans commit to Cores and RAM for one or three years, cannot be changed, and should cover only the floor.
- A cost alert emails once at the net threshold and changes nothing; it is a warning, not a limit.
Important Terminology:
- Commitment floor: the resource quantity that runs every hour of a plan term, the safe size for a Savings Plan.
- Showback: reporting estimated consumption to teams without moving money.
- Chargeback: billing actual costs to cost centres, based on the invoice.