Cloud Computing

Cloud Computing Fundamentals: Service Models, Deployment Models and Shared Responsibility

Beginner10 min readReviewed by on Updated

Cloud computing is infrastructure you rent through an API and pay for by usage. After reading this guide you will be able to explain the five NIST characteristics, place any service on the IaaS–PaaS–SaaS spectrum, describe regions and availability zones, and say precisely which security and operational duties stay with you under the shared-responsibility model.

What you will learn

  • Define cloud computing using the five essential characteristics from NIST SP 800-145
  • Distinguish IaaS, PaaS, SaaS, containers-as-a-service and function-as-a-service by what you still manage
  • Explain public, private, hybrid and multi-cloud deployment models and when each is chosen
  • Describe how regions, availability zones and edge locations shape availability and latency
  • Apply the shared-responsibility model to decide who patches, encrypts and backs up what

The word "cloud" gets attached to everything from a laptop backup service to a global content network, so it helps to start from a definition that engineers can test against. The one most people fall back on comes from the United States National Institute of Standards and Technology (NIST) in Special Publication 800-145, published in 2011 and still the reference. This guide walks through that definition, then explains the service models, deployment models and global infrastructure that every provider shares, and finishes with the shared-responsibility model — the part that determines what will be your fault when something breaks.

The five characteristics that make something "cloud"

NIST defines cloud computing as a model for on-demand network access to a shared pool of configurable computing resources that can be provisioned and released rapidly with minimal management effort. Five essential characteristics unpack that sentence:

  1. On-demand self-service. You create a virtual machine, a queue or a database yourself, through a console, a command-line tool or an API, without a ticket to a human operator.
  2. Broad network access. Resources are reachable over standard network protocols from any kind of client.
  3. Resource pooling. The provider serves many customers from the same physical pool — this is multi-tenancy — and you generally do not know or control which physical machine you land on beyond a coarse location such as a region.
  4. Rapid elasticity. Capacity can grow and shrink quickly, sometimes automatically, and from your point of view appears unlimited.
  5. Measured service. Usage is metered (CPU-seconds, gigabyte-months, requests, bytes transferred) and that meter drives the bill.

The test is useful in reverse. A virtualisation platform run by your own operations team behind a change-request process has resource pooling but not self-service or elasticity; NIST would call it virtualised infrastructure, not a private cloud. The distinction matters because most of the benefits people expect from "moving to the cloud" — speed, pay-per-use, automatic scaling — come from the operating model, not from the hardware being somewhere else.

Elasticity is not the same as scalability

Scalability is the ability of a system to handle more load when you give it more resources. Elasticity is the ability to acquire and release those resources automatically as load changes. A monolith that only runs on one large machine can be scalable (buy a bigger machine) but is not elastic. Cloud platforms make elasticity cheap to implement, but only for systems designed so that adding a third identical instance actually helps — stateless request handling, shared session storage, queues to absorb bursts. That design work is yours; the platform only provides the knobs.

Service models: what you stop managing

The classic taxonomy has three layers, and two more have become standard since the NIST paper was written. The clearest way to compare them is to ask, at each level, what you still manage.

ModelYou provideProvider managesTypical examples
IaaS — Infrastructure as a ServiceOS configuration, runtime, application, data, scaling rulesPhysical hosts, hypervisor, storage hardware, network fabricAmazon EC2, Azure Virtual Machines, Google Compute Engine
CaaS — Containers as a ServiceContainer images, orchestration manifests, application, dataNodes (optionally), control plane, image registryAmazon EKS/ECS, Azure Kubernetes Service, Google GKE
PaaS — Platform as a ServiceApplication artifact, configuration, dataOS, runtime, load balancing, patching, scalingAzure App Service, AWS Elastic Beanstalk, Heroku-style platforms
FaaS — Function as a Service ("serverless")Function code, event wiringEverything else, including capacity planningAWS Lambda, Azure Functions, Google Cloud Functions
SaaS — Software as a ServiceYour data and user configurationThe entire applicationMicrosoft 365, Salesforce, GitHub

Two rules of thumb follow. First, the further right you move, the less operational work you carry and the less freedom you have: on PaaS you cannot install a kernel module, and on FaaS your code has a maximum execution time and no persistent local disk. Second, managed data services (a hosted PostgreSQL, a managed message broker) behave like PaaS for that one component even when the rest of your system is on IaaS. Real architectures mix the layers freely; the label applies per component, not per company.

Deployment models: who shares the pool

NIST's four deployment models describe who the infrastructure is operated for, and multi-cloud has become a fifth in practice.

Public cloud. The provider owns the infrastructure and serves the general public. Tenants share hardware, isolated by virtualisation and network controls. This is what "the cloud" means in most conversations and the model the rest of this subject assumes.

Private cloud. The infrastructure is operated exclusively for one organisation, on-premises or hosted by a third party. It can be genuine cloud (self-service, elastic, metered) when built on a platform such as OpenStack or VMware with proper automation, but it is frequently just virtualisation with a portal. Regulated industries choose it for data-residency or latency reasons; the trade-off is that the organisation pays for peak capacity up front and staffs the platform team itself.

Hybrid cloud. Two or more distinct clouds — commonly a private data centre and a public provider — bound by technology that lets workloads and data move between them: a private network link, a shared identity provider, a common orchestration layer. The typical drivers are gradual migration, keeping a mainframe or a latency-sensitive system on-premises, or "bursting" batch work into the public cloud when the local capacity is exhausted. Hybrid is where most large enterprises actually live, and the network and identity plumbing between the two sides is where most of the difficulty sits.

Community cloud. Shared by several organisations with common concerns — a consortium of hospitals, a government agency cluster. Rare in practice, but the model exists.

Multi-cloud. Using more than one public provider. Sometimes deliberate (avoid lock-in, satisfy a regulator asking for exit plans, use one provider's strength in analytics and another's in identity), often accidental (one team chose AWS, an acquired company runs on Azure). Genuine portability across providers is expensive because managed services differ in behaviour, not just in API; most multi-cloud strategies settle for a common identity layer, common observability and per-workload placement rather than running the same workload everywhere.

Regions, zones and edges

Every large provider organises its footprint in the same three tiers, even if the names vary.

A region is a geographic area — a metropolitan area or a country — containing a cluster of data centres. Regions are isolated from each other: a control-plane failure in one does not cascade to another, and data does not leave a region unless you move it. You choose a region for three reasons: proximity to users (latency), legal constraints on where data may reside (sovereignty), and service availability, because new services and instance types arrive in the largest regions first.

An availability zone (AZ) is one or more physically separate data centres inside a region, each with its own power, cooling and network, connected to its siblings by high-bandwidth, low-latency private links. AWS regions typically contain three or more zones; Azure has been retrofitting zones into its regions and now offers them in most major ones. The design contract is that a failure affecting one zone should leave the others running. Everything in the high-availability discipline builds on that contract: put one instance in each of two or three zones behind a load balancer and the loss of a data centre becomes an event you read about in a report rather than an outage.

Edge locations are small points of presence, far more numerous than regions, that host content-delivery caches, DNS resolvers and, increasingly, small compute runtimes. They exist to shorten the round-trip to the user; they do not run your databases.

Region eu-west-1 (Ireland)
├── Availability Zone a  ── data centre(s), own power/cooling/network
├── Availability Zone b  ── data centre(s), own power/cooling/network
└── Availability Zone c  ── data centre(s), own power/cooling/network
        ▲ private low-latency links between zones (synchronous replication feasible)

Region eu-central-1 (Frankfurt)   ── a second region for disaster recovery;
                                     replication is asynchronous over the backbone

Edge locations (hundreds)          ── CDN caches and DNS close to users

One consequence catches newcomers: zone names are per-account. The zone your account calls a may be a different physical building from the one another account calls a; providers shuffle the mapping to spread load. Use the zone identifier the provider exposes when you need to coordinate across accounts.

The shared-responsibility model

The most important diagram in cloud security has two rows. The provider is responsible for the security of the cloud — the buildings, the hardware, the hypervisor, the managed-service software. You are responsible for security in the cloud — everything you configure or deploy. The line between the rows moves with the service model.

ConcernIaaS (VM)PaaS (managed app / database)SaaS
Physical security, hardware, hypervisorProviderProviderProvider
Operating-system patchingYouProviderProvider
Runtime and middleware patchingYouProviderProvider
Network controls (firewall rules, private subnets)YouShared — you configure what is exposedProvider
Identity and access managementYouYouYou
Application code and its vulnerabilitiesYouYouProvider
Data classification, encryption choices, backupsYouYouYou
Client devices and end-user accountsYouYouYou

Three rows never move to the provider: identity, data and the endpoints your users log in from. This is why the large public cloud breaches of the past decade were almost all misconfigurations — a storage bucket made world-readable, an access key committed to a public repository, a database with no authentication reachable from the internet — rather than failures of the underlying platform. The provider hands you encrypted disks and fine-grained access policies; whether you use them is up to you.

Two practical habits follow from the model. Read the responsibility matrix of every managed service you adopt, because "managed" varies: a managed Kubernetes control plane does not patch your worker nodes unless you enable node auto-upgrade, and a managed database applies engine patches only inside the maintenance window you define. And treat provider compliance certifications as covering their half only: an ISO 27001 certificate for the platform says nothing about whether your application handles personal data lawfully.

Paying for usage: CapEx, OpEx and the real economics

On-premises infrastructure is a capital expenditure: you buy servers sized for the peak you expect in three years, depreciate them, and run them at low average utilisation. Cloud is an operational expenditure: you pay monthly for what you used. The shift is not automatically cheaper. It is cheaper when load is variable, when you would otherwise over-provision heavily, when you can switch resources off outside working hours, or when the alternative was a team of people racking hardware. It is more expensive when load is flat and predictable and the organisation already has the data centre and the staff, and it can be dramatically more expensive when nobody switches unused resources off.

The metering that makes cloud flexible also makes cost an engineering concern rather than a procurement one. Every architectural decision — how much data crosses region boundaries, whether logs are retained forever, whether a batch job runs on the cheapest interruptible capacity — now shows up on an invoice within days. The cost-management guide covers the discipline in depth; the fundamental point here is that in the cloud, engineers spend money by writing code.

Common mistakes

Treating a lift-and-shift as "being in the cloud". Copying a virtual machine to a provider and running it exactly as before gains you a new bill and none of the elasticity. Until the workload can scale horizontally and be recreated from automation, it has moved, not transformed.

Assuming the provider patches your VMs. On IaaS the operating system is yours. Unpatched images and forgotten test machines are a standard finding in cloud security audits. Automate patching or, better, rebuild instances from fresh images on a schedule.

Confusing high availability with disaster recovery. Spreading instances across zones handles a data-centre failure; it does nothing for a bad deployment that corrupts the database or a regional outage. Backups and, where justified, a second region are separate decisions — covered in high availability and disaster recovery.

Using the root or owner account for daily work. Every provider offers a single all-powerful account identity. Lock it with hardware multi-factor authentication, store the credentials offline and create scoped identities for people and workloads.

Ignoring data-transfer boundaries. Traffic inside a zone is usually free, across zones costs a little, across regions or out to the internet costs meaningfully more. Architectures that ship large datasets between regions "because the analytics team is over there" surprise their owners at the end of the month.

Key takeaways

  • Cloud computing is defined by an operating model — self-service, elasticity, pooling, metering, network access — not by whose building the servers stand in.
  • Service models form a spectrum from IaaS to SaaS; each step right removes operational duties and control in equal measure, and the label applies per component.
  • Public, private, hybrid and multi-cloud describe who shares the pool; hybrid is where most large organisations sit, and its hard parts are identity and networking.
  • Regions isolate failures and satisfy data-residency rules; availability zones inside a region are the unit of high availability; edges bring caches closer to users.
  • Under shared responsibility the provider secures the platform, while identity, data, application code and configuration always remain yours.
  • Pay-per-use turns architecture choices into cost choices, so cost awareness is part of engineering, not a finance afterthought.

Sources and further reading

Frequently asked questions

Is hosting on a rented dedicated server the same as using the cloud?
No. A dedicated server ordered by e-mail and delivered in three days fails the NIST test on on-demand self-service, rapid elasticity and measured service. Cloud computing means you can create and destroy resources yourself through an API within minutes, scale them up and down, and be billed only for what you consumed. The physical machine is rented in both cases; the operating model is what differs.
What is the practical difference between IaaS and PaaS?
With IaaS you receive a virtual machine, a disk and a network and you own everything above the hypervisor: operating system patches, runtime, middleware, backups and scaling. With PaaS you hand over an application artifact and the provider runs the operating system, runtime and scaling for you. PaaS removes work but also removes control, so software that needs kernel modules, custom daemons or exotic runtimes usually stays on IaaS.
Does the cloud provider back up my data automatically?
Only where the service explicitly says so, and never for data inside your own virtual machines. Managed databases usually offer automated snapshots with a retention window you configure; object storage offers versioning and replication you must turn on. Under the shared-responsibility model, data, its classification and its recoverability remain your responsibility regardless of the service model.
Why do providers split a region into several availability zones?
An availability zone is one or more data centres with independent power, cooling and networking, placed far enough apart that a fire or power failure in one should not affect another, yet close enough for low-latency synchronous replication. Spreading instances across zones lets you survive a data-centre failure without paying the latency and complexity cost of running across two regions.
Is serverless really without servers?
There are servers; you simply never see them. A function-as-a-service platform receives your code and an event trigger, allocates an execution environment when the event fires, runs your code, and reclaims the environment afterwards. You pay per invocation and per unit of execution time and never patch or size a host, which is the actual promise behind the name.