What the cloud is

The cloud is a data centre you rent over the internet

Youa browser, a script, a toola request over the interneta server, running in about a minuteThe provider’s data centresowned, powered, cooled and staffed by themComputeprocessors, memoryStoragedisks and filesNetworkaddresses and linksServicesdatabases, queues, AIYou get a slice of each for as long as you want it,and the bill counts what you used.Cloud computing is renting computing over the internet: on demand, and paid by use.The hardware is ordinary: racks of servers, disks and cables in a building. What changed is who owns it,and that you get it by sending a request instead of signing a purchase order.

Cloud computing delivers computing as a service: someone else owns the machines, and you rent what you need, when you need it.

What the cloud is · Going deeper · 1 of 6

From an experiment in 2006 to the default

20062010201520202025Each step moved one more job from the customer to the provider.AWS S3, then EC2storage in March,servers in AugustGoogle App Engineyou bring code,Google runs itMicrosoft Azuregenerally availableNIST definitionfive traits, threemodels, four waysKubernetes, Lambdacontainers at scale;code with no serverPer-second billingAWS, Google andAzure virtual machinesAI as a servicemodels rentedby the tokenIn under twenty years, renting went from an experiment to the default way to run software.By 2025 the market was about 100 billion US dollars a quarter (Synergy Research). The idea has not changed since 2006:someone else runs the building, and you call an API.

Amazon rented out storage in March 2006 and servers that August. Google and Microsoft followed, and each later step, from platforms to functions to AI models, handed the provider one more job.

What the cloud is · Going deeper · 2 of 6

Five traits make a service cloud

On-demandself-serviceYou start it yourself.No ticket, no phone call.Broad networkaccessReached over standardnetworks, from any device.ResourcepoolingMany customers shareone pool of machines.RapidelasticityGrows and shrinkswith demand, fast.MeasuredserviceUse is metered, shownto you, and billed.A service is cloud when it has all five. If one is missing, it is hosting.This is the definition from NIST, the US standards body (Special Publication 800-145, 2011). Exams quote italmost word for word, and every later page of this lesson is one of these five at work.

The NIST definition: on-demand self-service, broad network access, resource pooling, rapid elasticity, measured service.

What the cloud is · Going deeper · 3 of 6

NIST’s whole model: 5 traits, 3 service models, 4 deployment models

5essential characteristicswhat makes it cloudon-demand self-servicebroad network accessresource poolingrapid elasticitymeasured service3service modelshow much the provider runsIaaSinfrastructure: you rent machinesPaaSplatform: you bring only codeSaaSsoftware: you just use it4deployment modelswhose it is, who shares itpublicopen to anyoneprivateone organisationcommunitya group with shared ruleshybridtwo or more, linkedNIST’s whole model is 5, 3 and 4: five traits, three service models, four deployment models.The next lesson is the three service models. Community cloud is rare outside government and research; exams from AWS,Microsoft and Google ask about public, private and hybrid, and add multi-cloud, which NIST did not name.

The same two-page standard also names the service models (IaaS, PaaS, SaaS) and the deployment models (public, private, community, hybrid).

What the cloud is · Going deeper · 4 of 6

A rented server is not cloud

Rented server (traditional hosting)CloudGet a serveremail sales, wait daysone request, a minuteChange its sizea new contractany timeShrink or stoppay until the term endsstop paying when you stopThe billa fixed fee per monthmetered by the secondWho else uses the hardwareoften just youa shared poolRenting alone is not cloud. Cloud is self-service, elastic, pooled and metered.A server rented by the month from a hosting company is someone else’s computer too, but it fails four of the five traits.Exam questions test this boundary: "which characteristic lets a team create a database at 3 am with no approval?"is on-demand self-service.

Traditional hosting fails self-service, elasticity, pooling and metering. The five traits are how you tell the two apart.

What the cloud is · Going deeper · 5 of 6

Public, private, hybrid, multi-cloud: where the machines sit

PublicProvideryours in purpleThe provider’s hardware, sharedwith other customers. The default.PrivateYour data centreCloud software on hardware usedby one organisation only.HybridYoursProviderYour site and a public cloud,joined by a private link.Multi-cloudProvider AProvider BTwo or more public providers,by choice or by acquisition.Where the machines sit, and who shares them, is a separate choice from using cloud at all.Most organisations are hybrid in practice: new systems go to a public cloud while older ones stay where they are.Hybrid means the two are connected and work as one system. Multi-cloud means more than one provider.

A deployment model says whose hardware it is and who shares it. Most organisations end up hybrid.

What the cloud is · Going deeper · 6 of 6

Why anything stays on premises, and how the cloud reaches it

“On premises” means in a building you control. Four reasons keep systems there, and each provider sells a way to extend its cloud to them.

Reasons to keep a system where it is

  • Latency A machine on a factory floor or a trading system needs an answer in a few milliseconds, and the nearest cloud region is too far away.
  • Law and contracts Some data must stay inside a country, a building or a certified facility (data residency).
  • Hardware already paid for A data centre bought two years ago is a sunk cost. Moving before it wears out means paying twice.
  • Licences and old systems Software licensed per physical processor, or a mainframe nobody dares to touch, can cost more to move than to keep.

The provider comes to you

AWS OutpostsAWS-owned racks in your building, run from the AWS console
Azure Local, Azure ArcAzure services on your hardware; Arc manages servers anywhere
Google Distributed CloudGoogle’s stack on site, including fully disconnected sites
For exams: hybrid needs a connection between the two sides, either a VPN over the internet or a dedicated private line (AWS Direct Connect, Azure ExpressRoute, Google Cloud Interconnect). Lesson 6 covers both.

Owning servers

Owning means buying for the peak and paying for the idle

paid for, switched on, idleMonTueWedThuFriSatSun1000Friday salecapacity you bought: sized for the peak, plus headroomdemand19%of this capacity is usedacross the week.12 to 18%is what typical companyservers run at, measuredacross real data centres.NRDC, 2014A server you own costs the same whether it is busy or idle, so you pay for the peak all year.You must also guess the peak months ahead. Guess high and money sits idle. Guess low and the sale that shouldhave been your best day turns customers away.

Owned capacity is fixed: sized for the busiest hour, and idle for most of the others.

Owning servers · Going deeper · 1 of 2

Guess too low and the best day becomes the worst

MonTueWedThuFriSatSun1000capacity you bought: 60Above the linerequests queue, time outor get an error page.It happens at the worstmoment: the hour with themost customers watching.Guessing low is the expensive mistake: lost sales on the busiest day, and no way to add servers in time.A new physical server takes weeks to arrive. So owners over-buy on purpose, which is how utilisation ends up near 15%.Both errors come from one cause: capacity is fixed long before demand is known.

Under-buying fails at the moment of highest demand, and physical servers cannot arrive in time to fix it. Fear of this is why owners over-buy.

Owning servers · Going deeper · 2 of 2

The server is only the first line of what a server costs

The server itselfthe only line on the purchase orderSpacea room, racks, physical securityPower and coolingfor every hour, busy or idleNetworkswitches, cabling, internet linksSoftwareoperating system and other licencesPeopleinstalling, patching, on call at nightSpares and replacementfailed parts now, a new fleet in 3 to 5 yearswhat the quote showswhat you also payCompare the whole cost of owning, not the price of the box: this total is called TCO.TCO is total cost of ownership. Providers publish calculators for it, and exam questions expect you to countthe hidden lines: power, space, staff and replacement, not only hardware.

Total cost of ownership counts every line: space, power, cooling, network, licences, people and replacement. Compare that total with the cloud bill.

Elasticity

Renting follows demand: capacity rises and falls with the load

MonTueWedThuFriSatSun1000demandcapacity you rentscale outscale in61%of what you pay for is used:it was 19% when you owned it.Steps of 10Servers are added andremoved by a rule, withnobody watching.Elasticity: capacity follows demand in both directions, automatically.Scaling out for the sale is half of it. Scaling back in afterwards is the half that saves money,and it is the half that owned hardware can never do.

Elasticity is capacity that grows and shrinks automatically to match demand, so you stop paying for idle machines.

Elasticity · Going deeper

Scale up or scale out: a bigger machine or more machines

Scale up (vertical): a bigger machine2 CPUs8 CPUs64 CPUs✓no change to the software✗a restart to resize, and a largest size exists✗still one machine: if it fails, everything is downScale out (horizontal): more machinesload balanceradd one moreidentical servers✓no ceiling, and one failure removes only a share✗the software must run as many copies at onceClouds are built for scaling out: many small, identical, replaceable machines behind a load balancer.Scalable means it can grow to meet load. Elastic means it also shrinks again by itself. A system can bescalable without being elastic; the saving comes from elastic.

Vertical scaling swaps in a larger machine and has a ceiling. Horizontal scaling adds identical machines behind a load balancer and has none, which is why clouds favour it.

Capital and operating expense

A purchase becomes a bill: capital expense turns into operating expense

Own: capital expense (CapEx)the purchase, before the first customerpower, space and staff, every monthmonth 012✗a guess, locked in for 3 to 5 years✓cheapest if it is fully used the whole timeRent: operating expense (OpEx)each month’s bill is that month’s usemonth 112✓nothing upfront, and you can stop any day✗the bill moves, so someone must watch itRenting turns one large purchase into a running bill that follows use.Finance calls the first capital expense and the second operating expense. The bars are shapes, not prices:what matters is when you pay, and whether the amount can change.

Owning is CapEx: pay first, use later. Renting is OpEx: pay as you go, for what you used.

Capital and operating expense · Going deeper · 1 of 4

You pay for servers multiplied by time, in any shape

One server for 100 hours100 hours wide, 1 server tall100 server-hours × $0.096 = $9.60The answer arrives in about 4 days.100 servers for one hour1 hour wide,100 servers tall100 server-hours × $0.096 = $9.60The answer arrives in 1 hour.You pay for area: servers multiplied by time. The same area costs the same in any shape.So speed is free. A job that would take four days on one machine can finish in an hour on a hundred, for the same $9.60.Billing is per second on the large providers, usually with a one-minute minimum; a ten-minute test of this server costs under 2 cents.Example price: a general-purpose virtual machine with 2 CPUs and 8 GB of memory, on demand, about $0.096 an hour (AWS list price, 2025).

One server for 100 hours costs the same as 100 servers for one hour. Per-second billing makes finishing sooner free.

Capital and operating expense · Going deeper · 2 of 4

Pooling makes it cheaper: customers’ peaks do not line up

A: a shop, busy at 9 am100B: a game, busy at 3 pm100C: reports, run at 9 pm100Each owns its own peak: 100 + 100 + 100 = 300300130what three owners buyone shared pool needs 130midnightnoonmidnightPeaks that do not line up can share machines: 130 does the work of 300.This is resource pooling, and it is why a provider can charge less than your own hardware would cost and still profit.With millions of customers the pooled curve is smoother still. Three example customers are drawn; the sum is computed from them.

Resource pooling: one shared pool needs far less hardware than its customers would need separately.

Capital and operating expense · Going deeper · 3 of 4

Economies of scale: every unit gets cheaper

Electricity bought for every 1 kWh that reaches the computers (PUE)1 kWh of computingcooling and lossesAverage data centre1.56 kWh from the gridGoogle’s fleet1.09 kWh from the gridUptime Institute global survey, 2024, and Google’s published fleet average, 2024.Their own chipsAWS Graviton, Azure Cobalt,Google AxionBulk buyingservers by the hundredthousand, built to orderAutomationthousands of serversper engineerHigh utilisationpooling keeps machinesbusy day and nightScale makes each unit cheaper: less wasted power, cheaper parts, fewer people per machine.AWS calls this "benefit from massive economies of scale" and it shows in the price list: AWS has announced well overa hundred price cuts since 2006. You get the efficiency of a giant without being one.

A provider wastes less electricity, designs its own chips, buys in bulk and automates. Economies of scale turn that into lower prices.

Capital and operating expense · Going deeper · 4 of 4

Multi-tenancy: shared hardware, separate machines

Customer Atheir appsoperating systemCustomer Btheir appsoperating systemCustomer Ctheir appsoperating systemWalls enforced by thehypervisor and the chip:A cannot see B’s memory,disk or network traffic.Hypervisor: shares out the CPU, memory and diskOne physical server, owned by the providervirtualmachineshardwareMulti-tenancy: strangers share one physical machine, and each sees only its own virtual one.Sharing is what makes pooling possible. Its side effect is the noisy neighbour: another tenant’s heavy use can slow yours.When rules or licences forbid sharing, every provider also rents single-tenant hardware (dedicated hosts) at a higher price.

A hypervisor slices one physical server into isolated virtual machines for different customers.

Agility

Minutes, not months: a new server is one request away

Owntypical for a new physical server; it varies by companyApprove the budget2 weeksOrder, build and deliver6 weeksRack, cable, install2 weeks10 weeks to the first loginRentOne request. About a minute.Drawn to the same scale it would be far thinner than this line: 10 weeks is 100,800 minutes.When a server takes a minute, trying an idea costs almost nothing, and so does deleting it.That is what providers mean by agility: you can run ten experiments this afternoon and keep the one that worked,instead of defending one big purchase for the next three years.

Agility: when resources arrive in a minute, experiments are cheap and so is undoing them.

Agility · Going deeper · 1 of 3

Four ways in, one API behind them

Console: click in a web pagebest for looking around and learningCommand line (CLI)aws ec2 run-instances --instance-type t3.microSDK: from your own programec2.run_instances(InstanceType="t3.micro")Infrastructure as code: a file that describes itresource "aws_instance" "web" { ... }The APIone front door,signed requestsa serverEvery way in ends at the same API, so anything you can click you can also script and repeat.Azure and Google work the same way: az vm create and gcloud compute instances create are the same request with another accent.Because the request is text, it can be reviewed, versioned and run a thousand times without a mistake.

The web console, the command line, an SDK and an infrastructure-as-code file all send the same signed request to the provider’s API.

Agility · Going deeper · 2 of 3

The six advantages on the AWS exam are these mechanisms, renamed

What AWS calls itWhat is actually happening1Trade fixed expense for variable expenseNo purchase; the bill follows usepage 42Benefit from massive economies of scalePooling and bulk buying lower the unit pricepage 73Stop guessing capacityElastic capacity follows demandpage 34Increase speed and agilityA server in a minute, not ten weekspage 55Stop spending money running and maintaining data centersThe building, power and hardware are theirspage 16Go global in minutesData centres on every continent, one request awaylesson 3The six advantages on the AWS exam are the mechanisms of this lesson under official names.Questions describe a situation and ask which advantage it shows. Find the mechanism first, then the phrase. A company thatno longer forecasts hardware for a seasonal peak has "stopped guessing capacity", whatever else the answer options say.

Each official advantage is one mechanism: variable cost, scale, elasticity, speed, no data centre, global reach.

Agility · Going deeper · 3 of 3

The same ideas in Microsoft’s and Google’s words

The ideaAWS (CLF-C02)Microsoft (AZ-900)Google (Cloud Digital Leader)Pay only for what you usevariable expenseconsumption-based modelOpEx, pay as you goCapacity follows demandstop guessing capacityscalability and elasticityelasticity, scalabilitySurvives failureshigh availabilityhigh availability, reliabilityreliabilityFast to start and changespeed and agilityagility, manageabilityagility, time to marketCheaper at scaleeconomies of scaleeconomies of scaletotal cost of ownershipReach users anywherego global in minutesgeo-distributionglobal reachThree exams, one set of ideas: learn the mechanism once and translate the wording.AZ-900 adds predictability (of cost and performance), security and governance as named benefits. Google’s exam framesthe same points as reasons for digital transformation. Wording follows the published exam outlines; check the currentoutline before you sit an exam, because the phrases are revised.

AZ-900 speaks of the consumption-based model, high availability, scalability, reliability and predictability. Google’s exam speaks of agility, total cost of ownership and transformation.

When the cloud costs more

Renting wins when demand moves; a flat load can be cheaper to own

how much of the month the machine is busy0%100%costrent on demandown, or commit for yearsrenting is cheaperowning is cheaperThe lines are illustrative.The crossing point is real,and differs for every workload.Renting wins when demand moves. A flat, always-busy load can be cheaper to own or to commit to.On-demand prices include a premium for the freedom to stop. If you never stop, you are paying for an option you do not use.Providers answer this themselves: commit to one or three years and the price falls by up to about 70% (lesson 9).

On-demand prices include a premium for the freedom to stop. Steady workloads should commit, or may be cheaper owned.

When the cloud costs more · Going deeper · 1 of 3

Data in is free, data out is billed

You and your usersProvideryour datadata in: freedata out to the internet: about $0.09 per GB1 TB outabout $85100 TB outabout $7,800Uploading is free and downloading is billed, so data is easy to move in and costly to move out.The charge is called egress. Example list price: AWS, first 10 TB a month to the internet, 2025; the first 100 GB a month is free andthe rate falls with volume. Azure and Google price it the same way. It matters most for video, backups and analytics exports.Since 2024 the three large providers waive egress fees for customers who are leaving altogether, but not for everyday traffic.

Egress is the charge for data leaving the provider. It is small per gigabyte and large at scale.

When the cloud costs more · Going deeper · 2 of 3

Lock-in: the more a service does for you, the harder it is to leave

easy to movehard to moveVirtual machinesLinux runs anywhereContainers, Kubernetesthe same on every providerManaged open sourcePostgreSQL, MySQL, KafkaProvider-only servicesDynamoDB, BigQueryyou operate morethe provider operates more✓Portable: you can leave, and you do the patching, scaling and backups yourself.✓Provider-only: far less work and often better, and your code now speaks one provider’s API.Lock-in is a price you choose, service by service, in exchange for less work.The services that save the most effort are the ones that exist on one provider only. Using them is usually right; do it knowingly,and keep the data in formats you can export. Skills are a form of lock-in too: a team that knows one provider is slow on another.

Portable building blocks mean more work for you. Provider-only services mean less work and a harder exit.

When the cloud costs more · Going deeper · 3 of 3

What else you give up, and what to do about it

Renting moves some risks to the provider and creates a few new ones. None is a reason to avoid the cloud; each is a thing to plan for.

  • Control of the hardware You cannot walk in, choose the disk model or delay the provider’s maintenance. In exchange, you never replace a failed disk.
  • Shared fate When a provider region has a bad day, thousands of companies do at once. On 20 October 2025 a DNS fault in a single AWS region disrupted services worldwide for hours. Spreading a system across zones and regions is lesson 10.
  • A bill that can surprise Anyone with access can start a hundred servers, and a forgotten test keeps charging. Budgets and alerts are lesson 9.
  • Shared responsibility The provider secures the building and the hardware. What you put on it, and who you let in, is still yours to secure. That split is the next lesson.

Leaving happens, rarely

$74.6Msaved by Dropbox over two years after moving most storage from AWS to its own data centres (2018 filing)
Exabytesis the scale at which that paid off: a huge, steady, predictable load
Mostworkloads are neither that large nor that steady, and stay
Moving work back from the cloud is called repatriation. It makes sense for the far right of the chart above: large, flat and well understood.

The providers

Three providers hold about 63% of the market

Amazon Web Servicessince 200630%Microsoft Azuresince 201020%Google Cloudsince 200813%Everyone elseAlibaba, Oracle, IBM and many more37%Three providers hold about 63% of the market, and they sell the same building blocks.Shares of worldwide cloud infrastructure spending, mid 2025 (Synergy Research); bars are to scale. The services differ in nameand detail, not in kind, which is why this course teaches the idea first and each provider’s name for it second.

AWS, Microsoft Azure and Google Cloud lead, and the three sell the same building blocks under different names.

The providers · Going deeper

Same building blocks, three sets of names

The building blockAWSAzureGoogle CloudA virtual machineEC2Virtual MachinesCompute EngineObject storageS3Blob StorageCloud StorageA managed SQL databaseRDSAzure SQL DatabaseCloud SQLCode with no server to manageLambdaAzure FunctionsCloud Run functionsManaged KubernetesEKSAKSGKEA private networkVPCVirtual NetworkVPCWho can do whatIAMMicrosoft Entra ID, RBACCloud IAMLearn the left column. The other three are vocabulary.Each row is a later lesson in this course: compute (4), storage (5), networking (6), databases (7) and identity (8).The exam courses under this foundation, such as AWS Cloud Practitioner and AZ-900, then drill one column.

A virtual machine is EC2 on AWS, Virtual Machines on Azure and Compute Engine on Google Cloud. The idea is the thing to learn.

Lesson complete

You finished “What the cloud is”

Nicely done.

Open in a new tab