Someone else's computers, by the hour
In 2005, a three-person team in Bengaluru with an idea for an online exam-practice service had to start by buying things. Two or three servers, a network switch, a rack in a data centre, a contract for bandwidth. The order took weeks; the money was spent before the first student arrived. If the service failed, the servers were sold at a loss. If it succeeded beyond all hopes — say, on the night before national exam results — it fell over, because nobody can buy servers in an evening.
The same team today types a command, or clicks a button, and a minute later has a virtual server in a data centre in Mumbai, billed by the second. On results night they run two hundred servers for six hours; the next morning they run four. They never see a machine. They also receive a bill that can surprise them, depend on a company whose rules they do not set, and remain responsible for much more of their system's security than many teams assume.
That is cloud computing: renting computing — processors, memory, storage, networks and the software built on them — over a network, on demand, from someone who runs it at enormous scale. This elective is about how it works underneath and how to build on it well. This first lesson answers four questions: where the idea came from, what exactly is being rented, how the physical world of regions and zones shows through, and when renting is actually the better deal.
From time-sharing to the public cloud
The idea is older than the personal computer. In 1961, at a lecture for MIT's centennial, John McCarthy suggested that computing might one day be organised as a public utility, as the telephone system was. The machines of the time made the suggestion natural: a computer filled a room and cost millions, so the question was how to share one. Time-sharing systems — CTSS at MIT from 1961, the Dartmouth Time-Sharing System from 1964 — let dozens of people at terminals use one computer at once, each feeling it was theirs. Multics, begun in 1964, was explicitly designed as a "computer utility", and commercial service bureaus sold computer time by the hour over telephone lines. Douglas Parkhill's 1966 book The Challenge of the Computer Utility already discussed public, private and community utilities, much as we classify clouds today.
Two inventions of that era still sit at the heart of every cloud. IBM's CP-67 and then VM/370 (1972) ran many complete virtual machines on one mainframe, each believing it owned the hardware. And the economics of sharing — buy one big machine, divide it among many users whose needs peak at different times — is exactly the argument this lesson makes with numbers below.
PCs and client–server systems pulled computing out of shared machines in the 1980s; the web pulled it back. Hosting companies rented servers, webmail and Salesforce (founded 1999) rented whole applications through a browser, and scientists built grids to share computers between universities. VMware brought virtual machines to ordinary x86 servers from 1999, and the open-source Xen hypervisor followed in 2003 (the next lesson explains how).
The modern public cloud is usually dated to 2006, when Amazon Web Services launched its object store S3 in March and rentable virtual machines, EC2, as a beta in August: an API call created a server, and you paid by the hour. Google App Engine (2008) ran customers' code without exposing servers at all. Microsoft announced Windows Azure in 2008 and opened it in 2010. Alibaba Cloud was founded in 2009 and grew into the largest provider in China. NASA and Rackspace started OpenStack in 2010, open-source software for building the same kind of cloud on your own hardware, and the US National Institute of Standards and Technology published its definition of cloud computing in 2011. Containers (Docker, 2013), Kubernetes (2014) and functions billed per request (AWS Lambda, 2014) changed what was rented, as later lessons show.
As of 2026 the market has a few very large global providers — Amazon Web Services, Microsoft Azure, Google Cloud and Alibaba Cloud, with Oracle, IBM, Huawei Cloud and Tencent Cloud among the others — many regional ones (OVHcloud and Scaleway in France, Hetzner and IONOS in Germany, Naver Cloud in South Korea and Sakura Internet in Japan, for instance), specialists in GPUs for AI, and a great many private clouds that organisations run on their own hardware with open-source software such as OpenStack, Apache CloudStack and Proxmox VE. None of them is the right answer for everyone; this elective describes them side by side.
What counts as cloud computing
The most widely cited definition is NIST's Special Publication 800-145 (2011). In our own words:
Definition— Cloud computing (after NIST SP 800-145)
Cloud computing is a way of providing computing resources — servers, storage, networks, platforms and applications — from a shared pool, over a network, so that customers can take and release them quickly, by themselves, with little effort from the provider. NIST names five essential characteristics:
- On-demand self-service: you obtain resources yourself, through a console or an API, without asking a person.
- Broad network access: resources are reached over the network with standard protocols, from many kinds of device.
- Resource pooling: the provider serves many customers (tenants) from the same physical resources, assigned and reassigned as demand changes; customers generally do not know or control exactly where their resources are, beyond a chosen region.
- Rapid elasticity: capacity can grow and shrink quickly, often automatically, so that it appears unlimited to the customer.
- Measured service: use is metered — CPU-seconds, gigabyte-months, requests — and that measurement drives billing and control.
A rented server in a colocation centre, set up by email over two weeks and paid for by the year, fails characteristics 1 and 4: it is hosting, not cloud. A company's own OpenStack installation that its teams use through an API meets all five: it is a cloud, a private one.
NIST also distinguishes service models by how much of the stack the provider runs:
| Model | You rent | You still manage | Examples (as of 2026) |
|---|---|---|---|
| IaaS — infrastructure as a service | virtual machines, disks, networks | the operating system and everything above it | Amazon EC2, Azure Virtual Machines, Google Compute Engine, Alibaba Cloud ECS, OpenStack Nova |
| PaaS — platform as a service | a platform that runs your code or data | your code, data and configuration | managed databases; app platforms such as AWS Elastic Beanstalk, Azure App Service, Google App Engine, Alibaba Cloud Serverless App Engine, or Cloud Foundry run privately |
| SaaS — software as a service | a finished application | your data, users and settings | webmail, online office suites, customer-relationship systems |
The boundaries have blurred since 2011. Containers as a service (managed Kubernetes, for instance) sits between IaaS and PaaS; functions as a service, or serverless, runs individual functions on demand and bills per request and per millisecond. The useful question is never "which acronym?" but "which layers do I operate, and which does the provider?".
And NIST names four deployment models: a public cloud, open to anyone who pays; a private cloud, run for one organisation (by it or by a contractor); a community cloud, shared by organisations with common needs, such as a group of universities or government agencies; and a hybrid cloud, a combination joined so that workloads and data can move between them. Using several public providers at once, multi-cloud, is a later term; the edge, multi-cloud and sovereignty lesson weighs it.
Why sharing makes capacity cheap
Why should renting from a provider ever be cheaper than owning, when the provider must also make a profit? Part of the answer is scale — buying servers, electricity and network capacity by the hundred thousand — but the deeper part is pooling. A single company must buy enough servers for its busiest hour, and they sit mostly idle the rest of the time. A provider serving thousands of companies only needs enough for the busiest hour of their combined load, and different customers are busy at different times: a school timetable service in Santiago peaks while a payroll system in Jakarta sleeps.
The simulation below gives each tenant a daily cycle in its own time zone, with random bursts on top, and compares two worlds: every tenant buys for its own peak, or one provider buys for the peak of the sum.
import numpy as np
rng = np.random.default_rng(2026)
HOURS = 24 * 60 # sixty days, hour by hour
t = np.arange(HOURS)
def tenant_loads(n):
"""n independent tenants: a daily cycle in each tenant's own time zone, plus bursts."""
phase = rng.integers(0, 24, size=(n, 1)) # where in the world the tenant is
daily = 1 + 0.8 * np.sin(2 * np.pi * (t - phase) / 24) # quiet at night, busy by day
bursts = rng.lognormal(mean=0.0, sigma=0.5, size=(n, HOURS))
return daily * bursts # work, in servers' worth
print("tenants | capacity needed: | average utilisation:")
print(" | each alone all pooled | each alone pooled")
for n in [1, 10, 100, 1000]:
load = tenant_loads(n)
alone = load.max(axis=1).sum() # every tenant buys enough for its own busiest hour
total = load.sum(axis=0)
pooled = total.max() # one provider buys enough for the busiest hour of the sum
used = total.mean()
print(f"{n:7d} | {alone:10.0f} {pooled:12.0f} | {used / alone:10.0%} {used / pooled:9.0%}")A tenant on its own keeps its servers about 14% busy on average: it pays for its rare busiest hour all the time. A thousand tenants, pooled, need 1230 servers instead of 7803, and keep them 92% busy. That gap is the margin the whole industry lives in.
The argument has limits that are worth knowing now. It needs loads that are roughly independent: when everyone wants the same thing at once — a pandemic moving every school and office online in March 2020, or a rush for scarce AI accelerators — the sum peaks with its parts, and providers ration capacity like anyone else. And it says nothing about who captures the saving; prices, discounts and contract terms decide that, as the section on economics shows.
Regions and availability zones
"The cloud" is a figure of speech. Underneath are buildings full of servers, with their own power supplies, cooling and fibre, in particular places, subject to particular laws, floods and power cuts. Every large provider organises them the same way.
Definition— Region and availability zone
A region is a geographic area — a metropolitan area such as Frankfurt, Mumbai, São Paulo or Johannesburg — where a provider runs several data centres. An availability zone (or simply zone) is one or more data centres within a region with independent power, cooling and networking, physically separated from the other zones — typically by kilometres to tens of kilometres — so that a fire, flood or power failure in one is unlikely to affect another, yet close enough that a round trip between zones takes only a millisecond or two.
The names differ a little — Amazon Web Services and Azure say availability zones, Google Cloud and Alibaba Cloud say zones — and so do the details: not every Azure region has zones, some providers add small "local" or "edge" zones in cities without a full region, and each publishes its own map (AWS, Azure, Google Cloud, Alibaba Cloud). A private cloud usually has fewer: perhaps two data centres in one city and a backup site elsewhere.
Regions and zones are how geography enters every design decision:
- Failure domains. A service running in one zone goes down with that zone. Running copies in two or three zones of a region survives the loss of one, at a small cost in latency between zones. Surviving the loss of a whole region needs copies in another region, at a much larger cost in latency and complexity (the storage and databases lesson covers how data is kept in step).
- Latency. Users in Nairobi are better served from a region in Africa or the Middle East than from one in Oregon; the speed of light in fibre adds about a millisecond of round trip per 100 km of cable. The cloud networking lesson computes latency budgets.
- Law. Data stored in a region is subject to that country's laws, and many laws restrict where personal data may go: the EU's GDPR, Brazil's LGPD, China's Personal Information Protection Law and India's Digital Personal Data Protection Act, among others. Choosing a region is often a legal decision before it is a technical one.
- Price and availability. The same machine type costs different amounts in different regions, and new hardware (the newest GPUs especially) reaches some regions long before others.
Owning or renting: the economics
The classic way to run computers is capital expenditure (capex): buy servers up front, then depreciate them — count a share of their price as a cost each month — over their useful life, typically four to six years. The cloud turns this into operating expenditure (opex): pay each month for what you used. Accountants care about the difference; engineers should care about three consequences.
- Elasticity. You pay for capacity only while you hold it. A batch job that needs 1000 server-hours costs about the same on one server for 1000 hours as on 1000 servers for one hour; the Berkeley report linked below called this cost associativity, and it is why a researcher can rent a large cluster for an afternoon.
- No stranded investment. A failed product leaves no servers behind; a successful one does not wait weeks for hardware.
- A price per unit that is higher than owning, when the owned unit is busy. The provider must pay for idle capacity, for staff, for research and for profit, so an hour of a rented server costs more than an hour of an owned one that is fully used.
So the question is not "is the cloud cheaper?" but "how busy would my own servers be?". The block below works it out for one server, with illustrative 2026 prices (examples, not quotes: real prices vary by provider, region, machine type and year), and then for a month of real-looking demand.
# Illustrative 2026 figures for ONE server (64 cores, 512 GiB) -- examples, not quotes.
PRICE = 18_000 # purchase price, USD
LIFE_MONTHS = 60 # depreciated over five years
POWER_KW = 0.7 # average draw, including its share of cooling
TARIFF = 0.15 # USD per kWh
SPACE = 150 # USD per month: rack space and network in a colocation centre
PEOPLE = 250 # USD per month: its share of the staff who run the fleet
HOURS = 730 # hours in an average month
own = PRICE / LIFE_MONTHS + POWER_KW * HOURS * TARIFF + SPACE + PEOPLE
print(f"owning one server costs {own:,.0f} USD a month, busy or idle")
for label, rate in [("on demand", 6.00), ("1-year commitment", 3.90), ("3-year commitment", 2.60)]:
rent_full = rate * HOURS
breakeven = own / rent_full
print(f"renting it {label:18} {rate:4.2f}/h = {rent_full:5,.0f} a month if always on;"
f" owning wins above {breakeven:4.0%} of hours")
# A month of hourly demand for a ticketing service, in servers (weekday peaks and one sale).
demand = []
for day in range(30):
for hour in range(24):
busy = 8 <= hour < 22
need = (12 if busy else 3) + (4 if busy and day % 7 < 5 else 0)
if day == 17 and 10 <= hour < 14: # tickets for a big concert go on sale
need = 60
demand.append(need)
rate = 6.00
peak = max(demand)
own_monthly = own * 30 * 24 / HOURS # this "month" has 720 hours
def cost(k):
"""Own k servers; rent on demand whatever the demand exceeds them by, hour by hour."""
return k * own_monthly + rate * sum(max(0, d - k) for d in demand)
best = min(range(peak + 1), key=cost)
print(f"\npeak demand {peak} servers, average {sum(demand) / len(demand):.1f}")
print(f"own enough for the peak ({peak}): {cost(peak):8,.0f} USD")
print(f"rent everything, hour by hour: {cost(0):8,.0f} USD")
print(f"own {best:2d}, rent the rest: {cost(best):8,.0f} USD")
for k in [best - 1, best, best + 1]:
share = sum(d >= k for d in demand) / len(demand)
print(f" the {k}th server is needed in {share:5.1%} of hours")With these numbers an owned server costs 777 USD a month, and renting the same capacity on demand costs 4380 USD a month if it runs all the time. Owning wins for any server that would be busy more than 18% of the hours; with a three-year commitment to the provider, the threshold rises to 41%. For the ticketing service, owning enough servers for the concert sale and renting everything on demand cost about the same, roughly 46,000 and 44,000 USD — but owning the steady part and renting only the peaks costs about 13,000.
Theorem— How many servers to own
Let be the number of servers needed in each of hours, let owning a server cost for the period, and let renting one cost per hour. If we own servers and rent the shortfall each hour, the total cost is minimised by owning the -th server exactly when it would be needed in at least a fraction of the hours.
Proof
Adding the -th owned server costs and saves one rented server in every hour with , so , where . As grows, can only stay the same or fall, so these differences never decrease: is convex in . It therefore decreases while and increases once (when exactly, both choices cost the same), and the minimum is at the largest with .
∎
Here : the 16th server is needed in 42.8% of hours, so own it; the 17th only during the four-hour sale, 0.6% of hours, so rent it. The same rule decides how much capacity to commit to a provider for a discount and how much to leave on demand, which is how the cost and FinOps lesson will use it.
That is why the industry's real decisions vary. Start-ups, research groups, retailers with seasonal peaks and anyone uncertain about demand mostly rent. Organisations with large, steady workloads sometimes find owning cheaper and have moved some of them back to their own hardware ("repatriation"), while keeping the cloud for bursts and experiments. Hybrid arrangements — owned or committed capacity for the base, on-demand for the peaks — are common. Data has gravity: once many terabytes live in one provider, moving them costs time and transfer fees, which is one reason the EU's Data Act requires providers serving EU customers to phase out charges for switching between them by January 2027.
The professional view
Choosing the cloud is the start of a long list of decisions, and professionals approach them in a fairly standard way:
- Start from the workload. How steady is the load, how sensitive is the data, which laws apply, what latency do users need, and what does an hour of downtime cost? The answers choose the region, the service model and the redundancy.
- Prefer managed services where they fit. A managed database costs more per hour than one on a virtual machine, but includes backups, patching and failover; the price is less control and more dependence on one provider.
- Design for failure from the start: spread across zones, automate recovery, and test it (Lumenary CS's reliability engineering).
- Treat infrastructure as code and cost as a metric: everything created by reviewed, versioned code (infrastructure as code), and spending watched like latency (cost and FinOps).
Each large provider publishes a "well-architected" guide organising these questions into pillars such as reliability, security, cost, performance and operations: AWS Well-Architected, the Azure Well-Architected Framework, Google Cloud's Well-Architected Framework, and Alibaba Cloud's equivalent. They overlap heavily, and reading one with the others' names in mind is a good way to see which advice is general and which is specific to a product. For private clouds, the OpenStack and Apache CloudStack projects document how the same building blocks are run on your own hardware.
Common misconceptions
Misconception
"The cloud is always cheaper." It is cheaper for capacity you would otherwise leave idle, and usually more expensive for capacity that is busy all the time. Whether it is cheaper overall depends on utilisation, on discounts, on staff time, and on costs that are easy to forget, such as data transfer out of the provider.
Misconception
"The provider takes care of security and backups." The provider secures the layers it runs. Your data, your identities and permissions, your configuration and — on IaaS — your operating system are yours to secure, and protecting data against deletion or encryption by an attacker needs backups that you design.
Misconception
"Running in three zones means we cannot go down." Zones protect against failures that stay inside one zone. Bugs, bad deployments, expired certificates and regional control-plane outages can cross zone boundaries, and they then dominate your downtime.
Misconception
"The cloud is virtual, so it has no physical footprint." It is buildings full of servers drawing electricity and water. The International Energy Agency estimated that data centres used about 415 TWh of electricity in 2024, around 1.5% of the world's consumption, and projected that this could roughly double by 2030, mostly because of AI. The sustainable cloud lesson returns to this.
Exercises
1. What the cloud is
Warm-upHumanA network of clinics in Accra moves its appointment-booking system to a cloud. The team uploads its application code; the provider runs it, patches the operating system and the language runtime, and adds instances when traffic grows. The team never logs in to a virtual machine, but it still writes the code and decides who may see the appointment data. Which service model is this?
2. What the cloud is
Warm-upHumanWhich of these are among the five essential characteristics of cloud computing in NIST's definition (SP 800-145)? Select all that apply.
3. What the cloud is
CoreHumanA small online bookshop in Lima runs its website on virtual machines it rents from a public cloud (IaaS), using a Linux image it chose when it created them. A serious vulnerability is announced in the SSH server that ships with that Linux distribution. Who is responsible for patching it on the bookshop's virtual machines?
4. What the cloud is
CoreHumanOwning a server costs a company 900 USD a month, all in, whether or not it is used. Renting the same capacity on demand costs 4.50 USD per hour. Above how many hours of use per month is owning cheaper? (These are illustrative figures.)
5. What the cloud is
Warm-upHumanA research group in Nairobi must run a simulation that needs 1200 server-hours of work, split into many independent pieces. The provider bills per second of server time, so the bill is the same whether they use few servers for a long time or many for a short time. If they use 400 servers at once, how many minutes does the run take (ignoring start-up time)?
6. What the cloud is
CoreHumanA service runs in two zones and stays up as long as at least one zone is up. Each zone is unavailable 0.1% of the time, independently of the other. But both zones also depend on one regional service, which is unavailable 0.05% of the time, and when it is down the whole service is down. Which is closest to the service's availability?
7. What the cloud is
CoreComputationalWrite best_owned(demand, own_cost, rent_rate). demand lists how many servers are needed in each hour of a period; owning one server for the whole period costs own_cost; renting one for an hour costs rent_rate. The company owns servers and rents, each hour, however many more that hour needs. Return (k, cost): the number of servers to own that minimises the total cost, and that total cost. If several tie, return the smallest.
The starter always owns enough for the peak.
Further reading and resources
- The NIST Definition of Cloud Computing (SP 800-145) · Peter Mell and Timothy Grance · 2011Two pages that gave the field its shared vocabulary.
- Above the Clouds: A Berkeley View of Cloud Computing · Michael Armbrust and colleagues · 2009An early, clear analysis of why the economics of the cloud work, and of the obstacles.
- AWS Global Infrastructure
- Azure availability zones overview
- Google Cloud: regions and zones
- Alibaba Cloud ECS: regions and zones
- AWS shared responsibility model
- Shared responsibility in the cloud (Microsoft Azure)
- Shared responsibilities and shared fate on Google Cloud
- ECS shared security responsibility model (Alibaba Cloud)
- OpenStackOpen-source software for building a cloud on your own hardware.
- Energy and AI · International Energy Agency · 2025Estimates of data centres' electricity use, today and to 2030.
- Regulation (EU) 2023/2854 (the Data Act)Among other things, rules on switching between cloud providers.