How Does Cloud Computing Store Data? | How It Works

12 min read

On this page

What is the short answer?

When the student saves a file "to the cloud", it travels over the internet to a data centre, a building full of computers run by a provider. Software there splits the file into pieces, writes each piece to drives on different machines and records where every piece went. To open the file later, the system looks up that record and joins the pieces again.

Definition: The US National Institute of Standards and Technology (NIST) defines cloud computing as a model for on-demand network access to a shared pool of configurable computing resources, such as servers and storage, that can be rapidly provisioned and released with minimal management effort.

NIST's definition, from September 2011, names three service models: Infrastructure as a Service (IaaS), renting processing, storage and networks; Platform as a Service (PaaS), deploying applications onto the provider's tools; and Software as a Service (SaaS), using the provider's applications, such as web-based email. Its deployment models are private, community, public and hybrid clouds.

What is inside a data centre?

A data centre holds rows of servers, computers without screens, stacked in metal frames called racks and linked by network switches. The documentation of Hadoop, an open-source data platform, notes that bandwidth within a rack is greater than between racks, so storage software takes racks into account.

A University of Illinois course text describes three kinds of storage:

  • Hard disk drives (HDDs) hold bits magnetically on rigid platters spinning on a spindle, read by a moving arm.
  • Solid-state drives (SSDs) use flash memory with no moving parts. They read quickly but wear out with repeated erasing.
  • Magnetic tape is much slower to reach but cheap for what it holds; the text calls it the dominant form of long-term backup storage.

All of this needs steady power and cooling. The US Department of Energy (DOE) says data centres can use 10 to 50 times the energy per floor space of a typical commercial office building.

What happens to a file when it is saved to the cloud?

A NIST storage guide of October 2020 names three traditional types of storage. Block storage gives a computer numbered chunks of raw space, like an empty drive. File storage keeps named files in folders, as on a laptop. Object storage keeps each item whole, with a description, in a flat store.

In the documentation of Amazon S3, one provider's object store, an object is a file plus metadata, name and value pairs such as the date last modified. Each object sits in a container called a bucket and has a key, its unique identifier within that bucket.

Large files are also cut up. The Hadoop Distributed File System (HDFS), open-source software from the Apache Software Foundation, stores each file as a sequence of blocks, typically 128 megabytes (MB) each. A central server, the NameNode, holds the metadata and decides which machines store each block.

How do copies stop data from being lost?

Drives fail, so pieces are kept more than once, often by replication. For three copies, the HDFS documentation puts the first on the writer's own machine or one in its rack, and the other two on two machines in a different rack.

Erasure coding uses less space. The data is cut into pieces, and extra parity pieces are calculated from them. UC Berkeley's course notes on Reed-Solomon codes show that a message of n pieces, sent as n + k, can be rebuilt from any n of them. HDFS's default policy uses 6 data and 3 parity pieces. Its documentation says erasure coding gives the same fault tolerance as replication with a storage overhead of no more than 50%, against 200% for three copies, at the cost of extra computing and network work.

Durability is whether data survives at all; availability is whether it can be reached right now. Amazon states that S3 Standard is designed for 99.999999999% durability and 99.99% availability of objects over a year.

Copies do not stop attacks: the UK National Cyber Security Centre (NCSC) reports ransomware compromising connected cloud backups. It describes a 3-2-1 rule, at least 3 copies on 2 devices with 1 offsite, and an offline rule: at least one backup disconnected at any time.

How is stored data kept private?

Encryption scrambles data so that only someone with the right key can read it. The NCSC's cloud security principles, last reviewed in June 2023, cover two states of data:

  • In transit: data crossing networks should be protected against tampering and eavesdropping, using standard protocols such as TLS and IPsec.
  • At rest: providers should encrypt all customer data stored in the service, and protect the keys through their full lifecycle.

The NCSC notes that if data has been appropriately encrypted, deleting its decryption key makes it inaccessible, a method called crypto-shredding.

Under the shared responsibility model the NCSC describes, the provider almost always keeps its servers running and physically protected. The customer always decides which data to store and must configure its services securely. The split shifts with the service model: the customer carries most under IaaS and least under SaaS.

Where does the data physically sit, and which rules apply?

Large providers group data centres into regions. Amazon describes each Availability Zone as one or more data centres with redundant power and networking, many kilometres from the other zones in its region but within 100 kilometres (km). Its S3 objects stay in their region unless the customer moves or copies them.

Rules on where data may go differ by place:

  • European Union. Under the General Data Protection Regulation (GDPR), Regulation (EU) 2016/679, personal data may flow freely to a country the European Commission finds gives adequate protection. Its list includes Japan, the United Kingdom and, since 10 July 2023, the United States, for firms in the EU-US Data Privacy Framework. Other transfers can rely on tools such as standard contractual clauses.
  • India. The Reserve Bank of India directed on 6 April 2018 that payment system data be stored in a system only in India, giving providers six months.
  • Singapore. Parliament passed a Digital Infrastructure Bill on 7 October 2026, law firm Baker McKenzie reported. Data centres with a critical IT load of at least 3 megawatts (MW) will need a licence and must meet energy-efficiency requirements; major cloud providers will be licensed too.

How much energy and water do data centres use?

The IEA, an intergovernmental energy agency, estimated in its 2026 report Key Questions on Energy and AI that data centres used 485 terawatt-hours (TWh) in 2025. It projects about 950 TWh by 2030, roughly 3% of world electricity demand, Eco-Business reported in May 2026.

  • United States. A Lawrence Berkeley National Laboratory report for the DOE, released in December 2024, found data centres used 176 TWh in 2023, about 4.4% of US electricity. It projected 6.7 to 12% by 2028.
  • Ireland. The Central Statistics Office (CSO) reported in July 2026 that data centres used 23% of metered electricity in 2025, up from 5% in 2015. Their use rose 10% in a year, to 7,663 gigawatt-hours (GWh).

Efficiency is measured as power usage effectiveness (PUE), first published by The Green Grid, an industry body, in 2007: the total energy a data centre uses divided by the energy its computing equipment uses. A PUE of 1.0 would mean every unit reached the computers.

Cooling also uses water: the Berkeley report estimated that US data centres consumed 64 billion litres directly for cooling in 2023, University of Wisconsin-Milwaukee researchers reported.

What do supporters and critics argue?

  • Data localisation. Supporters include the Reserve Bank of India, a regulator, which cited in 2018 the need for "unfettered supervisory access" to payment data. Critics include the Information Technology and Innovation Foundation (ITIF), a think tank, which counted 144 localisation measures in 62 countries in July 2021, up from 67 in 35 countries in 2017. It estimated that stricter data rules cut trade output and productivity and raise prices.
  • Efficiency or growth? Those who stress efficiency cite the IEA's finding that energy per AI task has fallen dramatically. Those who stress growth cite the same report, which found these gains outpaced by rising demand. The Uptime Institute, a data centre advisory body, reported in July 2025 that average PUE had changed little for a sixth year.
  • A few large providers. The UK Competition and Markets Authority (CMA), a competition authority, decided on 31 July 2025 that barriers to switching providers, or using several, harm competition. It estimated that Microsoft and Amazon Web Services (AWS) each held 30 to 40% of UK IaaS supply in 2024. AWS called the proposed action "unwarranted", and Microsoft said the panel's report "misses the mark again", The Register, a technology news site, reported. In March 2026 the CMA said both had set out steps on egress fees, charges for moving data out.

How does this work with real numbers?

Both examples use illustrative inputs and decimal units: 1 terabyte (TB) = 1,000 gigabytes (GB) = 1,000,000,000,000 bytes.

How much space do replication and erasure coding take?

  • Three-way replication of an illustrative 1 TB: 3 × 1 TB = 3 TB stored, an overhead of 200%. Any 2 copies can be lost.
  • Erasure code, 6 data + 3 parity pieces: each piece holds 1 TB ÷ 6 ≈ 166.7 GB. Nine pieces store 9 × 166.7 GB ≈ 1.5 TB, an overhead of 50%. Any 6 of the 9 rebuild the data, so any 3 can be lost.

The erasure code stores half as much and survives one more loss, but rebuilding a lost piece means reading 6 others, not copying 1.

How long does 1 TB take to upload?

Take an illustrative connection of 100 megabits per second (Mbit/s), ignoring overheads.

  • Bytes to bits: 1 TB × 8 bits per byte = 8,000,000,000,000 bits = 8,000,000 megabits.
  • Time: 8,000,000 ÷ 100 = 80,000 seconds ≈ 22.2 hours.
  • Double the speed: at 200 Mbit/s, 8,000,000 ÷ 200 = 40,000 seconds ≈ 11.1 hours, half the time.

Dividing bytes by a speed in bits would give an answer 8 times too small.

Where can the student go deeper in World 101?

This page follows one file from a laptop into a data centre and back. The World 101 subject Computing Fundamentals covers the wider field: the algorithms, hardware and software behind every computer, and how machines process information, from a single laptop to the large systems that link thousands of machines.

Three related subjects go further. Networks & the Internet explains the network of networks that carries a file to a data centre and back. Cybersecurity looks at protecting systems and data from misuse and unauthorised access, and at restoring them when defences fail. The Energy Transition covers the shift in how the world supplies and uses energy, towards low-carbon electricity, which shapes how data centres are powered.

Sources

  • NIST, SP 800-145, September 2011, and glossary; SP 800-209, October 2020.
  • University of Illinois, CS 340 text "Storage", 10 December 2024; UC Berkeley, CS 70 coding notes, Summer 2018.
  • Apache Hadoop 3.3.5, HDFS Architecture and Erasure Coding pages, 15 March 2023.
  • Amazon S3 User Guide, read 9 November 2026.
  • UK NCSC: offline backups blog, 13 August 2019; Cloud Security Principles and shared responsibility guidance (10 May 2022), reviewed 7 June 2023.
  • European Commission, adequacy decisions (Regulation (EU) 2016/679; Data Privacy Framework, 10 July 2023), read 9 November 2026.
  • Reserve Bank of India, Storage of Payment System Data, 6 April 2018.
  • Singapore MDDI, 8 September 2026; Baker McKenzie, 30 October 2026 (passed 7 October 2026).
  • US DOE, Lawrence Berkeley report release, 20 December 2024 (2023 and 2028 figures); Data Centers and Servers page, read 9 November 2026.
  • IEA, Key Questions on Energy and AI, 2026 (2025 and 2030 figures), via Eco-Business, 11 May 2026.
  • CSO, Data Centres Metered Electricity Consumption 2025, July 2026 (2015 figure).
  • The Green Grid, PUE pages (metric published 2007); Uptime Institute survey, 29 July 2025.
  • P. McCauley and M. Scanlan, The Conversation, 19 August 2025 (2023 water figure).
  • ITIF, cross-border data flows report, 19 July 2021 (2017 and 2021 counts).
  • CMA, cloud market investigation, 31 July 2025, and release, 31 March 2026; The Register, 31 July 2025.

Common errors and misconceptions

  • Misconception: Data in the cloud has no physical home. Correct: It sits on drives in data centres, in named regions whose rules differ.
  • Misconception: A file saved to the cloud is kept as one copy. Correct: Its pieces are replicated or erasure coded across machines and racks.
  • Misconception: Replication is the same as a backup. Correct: It guards against failed hardware, but the NCSC reports ransomware reaching connected backups and describes keeping one offline.
  • Misconception: Durability and availability mean the same. Correct: Durability is whether data survives; availability is whether it can be reached right now.
  • Misconception: A 100 Mbit/s connection uploads 100 MB a second. Correct: Speeds count bits and sizes count bytes, so it moves 12.5 MB a second.

Key takeaways

  • NIST defines cloud computing as on-demand network access to a shared pool of computing resources, offered as infrastructure, platform or software services.
  • A saved file is split into blocks or kept as an object with metadata and a key, and a record notes where each piece is stored.
  • Three-way replication triples the space used; an erasure code of 6 data and 3 parity pieces uses 1.5 times the space.
  • Durability means data survives and availability means it can be reached now; the NCSC describes a 3-2-1 rule and offline backups.
  • Encryption protects data in transit and at rest, but the customer still decides what to store and how to configure it.
  • Data centres used 485 TWh worldwide in 2025 by the IEA's estimate, and 23% of Ireland's metered electricity, the CSO reported.

Test yourself

What happens to a file when it is saved to a cloud service?

It travels over the internet to a data centre, is split into pieces stored on several machines, and its locations are recorded as metadata.

How does an erasure code of 6 data and 3 parity pieces survive losses?

Any 6 of the 9 pieces are enough to rebuild the data, so up to 3 pieces can be lost.

How long would an illustrative 1 TB take to upload at 100 Mbit/s?

1 TB is 8,000,000 megabits; 8,000,000 ÷ 100 = 80,000 seconds, about 22.2 hours, ignoring overheads.

What does a PUE of 1.5 tell the student?

The data centre uses 1.5 units of energy in total for every unit its computing equipment uses.

What is crypto-shredding?

Deleting the decryption key for appropriately encrypted data, which the NCSC says makes that data inaccessible.

What did the UK CMA find in its cloud investigation of 2025?

That barriers to switching or using several providers harm competition; it estimated Microsoft and AWS each held 30 to 40% of UK IaaS supply in 2024.

Go deeper in World 101: Computing Fundamentals

Write about it

Turn a question from this chapter into a paper.

The Writing Studio has a structure guide, a rubric and citation tools for 15 formats, from essays to research papers. A free account keeps up to five drafts. Learnacy Essential adds feedback from an AI writing coach and submission to The OYI Review.

Publication depends on review and is not guaranteed.