Cloud computing industry
The cloud computing industry comprises the companies that supply computing capacity, storage, networking and software as a metered service delivered over the internet, rather than as equipment that a customer buys and operates itself. It is among the largest and fastest-growing sectors of the technology economy, with global spending measured in hundreds of billions of dollars annually, and it has become the infrastructure on which most modern software, and increasingly most artificial intelligence, is built.
The industry's defining characteristic is the conversion of computing from a capital expense into an operating one. An organisation that once bought servers, housed them in a building it cooled and powered, and employed people to maintain them, now rents capacity by the hour or the request from a provider operating at a scale no individual customer could match. The economic consequence is that the cost of starting a software business fell by orders of magnitude, which is the single most important reason the number of such businesses rose so steeply after 2006.
The market is highly concentrated. Amazon Web Services, Microsoft Azure and Google Cloud Platform together account for roughly two-thirds of global spending on cloud infrastructure, with Alibaba Cloud, Oracle, IBM and Tencent Cloud holding most of the remainder. Above this infrastructure layer sits a much larger and more fragmented population of software companies delivering applications on a subscription basis.
Service models
The industry is conventionally divided into three layers, distinguished by how much of the stack the provider manages.
Infrastructure as a service supplies the raw components: virtual machines, block and object storage, virtual networks and load balancers. The customer manages the operating system and everything above it. Amazon's Elastic Compute Cloud, launched in 2006, established the category.
Platform as a service supplies a managed environment in which a customer deploys code without managing servers at all. Databases, message queues, container orchestration and, in the most abstracted form, serverless functions that run in response to events and are billed by the millisecond, belong here.
Software as a service supplies a finished application accessed through a browser and licensed by subscription. Salesforce, launched in 1999, is generally identified as the first major example, and the category now includes most business software: electronic mail, document collaboration, customer relationship management, human resources, accounting, communications and design.
The boundaries have blurred considerably. Every large infrastructure provider sells managed databases, machine learning platforms and analytics services that are plainly platform offerings, and the strategic direction of the industry is upward: providers compete to supply higher-level services, because these carry better margins and because a customer using a managed database is harder to move than one using a virtual machine.
Deployment models
Public cloud describes capacity shared among many customers on infrastructure the provider owns.
Private cloud describes infrastructure dedicated to a single organisation, whether operated by that organisation or hosted for it, and adopted where regulation, latency or existing investment make shared infrastructure unsuitable.
Hybrid cloud combines the two, typically with some workloads retained in a company's own data centres and others in public capacity. Most large enterprises operate this way, and the major providers all sell hardware and software intended to extend their environments into customer premises.
Multicloud describes the deliberate use of more than one provider, adopted to reduce dependence, to obtain the best service in each category or, frequently, as the accidental result of acquisitions and of individual teams choosing their own tools.
Edge computing places capacity close to where data is produced, in local facilities, in telecommunications networks or in devices, for applications where the delay in reaching a distant data centre is unacceptable.
Sovereign cloud describes capacity operated under legal and technical arrangements intended to satisfy a government's requirements about jurisdiction and access, a category that has grown rapidly as the legal conflicts described below have intensified.
History
Antecedents
The idea is older than the term. John McCarthy suggested in 1961 that computation might one day be organised as a public utility, and the time-sharing systems of the 1960s and 1970s, in which many users shared a single expensive mainframe through terminals, embodied the principle. The economics reversed with the minicomputer and then the personal computer, which made it cheaper for each organisation, and eventually each person, to own computing outright.
Several intermediate forms appeared in the 1990s. Application service providers hosted software for customers over networks, mostly unsuccessfully, because the underlying infrastructure was not elastic and the economics did not work. Grid computing distributed computation across many machines for scientific purposes. Virtualisation, which allows one physical server to present itself as several independent machines, was developed for mainframes in the 1960s and brought to commodity hardware by VMware from 1999; it is the technology without which the modern industry would not exist, because it is what allows a provider to sell fractions of a machine to unrelated customers.
The founding of the modern industry, 1999 to 2010
Salesforce launched in 1999 selling customer relationship management software over the web, under the slogan "no software", and demonstrated that businesses would trust important data to an application they did not host.
Amazon Web Services launched its first services in 2004 and 2006, beginning with the Simple Storage Service in March 2006 and the Elastic Compute Cloud in August of the same year. The origin is frequently misdescribed: the business did not arise from selling spare capacity left over from retail operations but from Amazon's own internal need for standardised infrastructure services, which the company concluded could be sold to others. The pricing model, by the hour with no commitment, was the decisive innovation, because it meant a start-up could obtain the same infrastructure as a large corporation and pay only for what it used.
Google launched App Engine in 2008 and Microsoft announced Azure in the same year, though both took several years to become serious competitors in infrastructure. Heroku, Rackspace and a number of others occupied the intervening ground. The open source project OpenStack, launched in 2010, attempted to provide an alternative that any organisation could run, and became widely used in telecommunications and in private deployments without displacing the commercial providers.
Enterprise adoption, 2011 to 2019
The decade that followed was characterised by the migration of established organisations, which was slower and more difficult than the early adoption by start-ups.
The obstacles were principally organisational. Existing applications had been written for fixed infrastructure and did not benefit from elasticity without rewriting. Security and compliance functions objected, in most cases before the providers had obtained the certifications that would satisfy them. Procurement processes designed for capital purchases did not fit a metered service. And the people who operated the existing data centres had little incentive to advocate their closure.
Providers addressed these through certification, through regional data centres allowing customers to specify where data resided, and through professional services and partner networks. The technical development that mattered most was the container, popularised by Docker from 2013 and orchestrated by Kubernetes, released by Google in 2014 and given to the Cloud Native Computing Foundation. Containers package an application with its dependencies so that it runs identically anywhere, which made moving workloads between environments practicable and, by design or otherwise, reduced the degree to which a customer was tied to a single provider.
Microsoft's position changed decisively under Satya Nadella, appointed chief executive in 2014, who redirected the company around cloud and subscription software and used its existing enterprise relationships, its licensing agreements and its familiarity to large information technology departments as the route to market. Azure grew from a distant competitor to a clear second.
The pandemic and the artificial intelligence period, 2020 to 2026
The coronavirus pandemic accelerated adoption sharply, as organisations moved to remote working and discovered that capacity they could increase in hours was worth more than capacity they owned.
The period after 2022 has been dominated by artificial intelligence. Training large models requires enormous quantities of specialised processors, principally graphics processing units from Nvidia, connected by very fast networks and housed in facilities drawing power measured in hundreds of megawatts. Almost no organisation can assemble this independently, so the providers have become the intermediaries through which nearly all large-scale model development occurs.
The consequences for the industry have been substantial. Capital expenditure by the largest providers rose to levels comparable with the largest infrastructure programmes in any sector, exceeding a hundred billion dollars annually in aggregate. Access to processors became a competitive weapon, and several providers designed their own, Google with its Tensor Processing Units, Amazon with Trainium and Inferentia, Microsoft with Maia. Partnerships with model developers became strategically central, most prominently Microsoft's investment in OpenAI, Amazon's and Google's in Anthropic, and the arrangements made by Oracle and others to supply capacity to model developers at scale.
Electricity became a binding constraint for the first time in the industry's history. Data centre power demand rose steeply enough to affect national electricity planning in several countries, to delay grid connections by years in some regions, and to prompt providers to contract directly for nuclear and renewable generation.
Market structure
The infrastructure providers
Amazon Web Services is the largest by revenue and has been since the market began. It is the most profitable part of Amazon by a wide margin, generating the majority of the parent company's operating income for several years while representing a far smaller share of its revenue, and it has effectively funded the retail business's investment. Its advantages are the breadth of its service catalogue, the length of time its services have been operated at scale and the size of the ecosystem built around it.
Microsoft Azure is second and has grown fastest among the three for most of the past decade. Its advantage is distribution: Microsoft already sells software to virtually every large organisation in the world, its sales force and licensing agreements reach procurement departments directly, and its hybrid products appeal to customers who intend to retain their own data centres. Its position in artificial intelligence, through the OpenAI relationship, has been the principal driver of recent growth.
Google Cloud Platform is third and was the last of the three to treat enterprise sales seriously. Its strengths are in data analytics, in machine learning infrastructure including its own processors, and in Kubernetes, which it created. It reached profitability considerably later than its competitors.
Alibaba Cloud leads in China, where the American providers have limited presence for regulatory reasons, and competes across Asia. Tencent Cloud and Huawei Cloud are the other substantial Chinese providers.
Oracle entered late and has grown rapidly by concentrating on database workloads, where its existing customers are captive, and by supplying very large quantities of capacity to artificial intelligence developers. IBM repositioned around hybrid arrangements following its 2019 acquisition of Red Hat for about US$34 billion.
A further group supplies specific niches: Cloudflare from the network edge, DigitalOcean and Linode for smaller customers, Snowflake and Databricks for data platforms operating across the major clouds, and CoreWeave and similar firms supplying graphics processors to model developers.
Economics
The industry's economics are distinctive in several respects.
Capital intensity is extreme and rising. A single large data centre costs billions of dollars, and the largest providers each spend tens of billions annually on facilities, servers and networks. This is the principal barrier to entry and the reason the market concentrated so quickly.
Operating leverage is correspondingly high. Once the facility exists, additional customers are served at low marginal cost, which is why gross margins in infrastructure are well above those in any hardware business and why the largest provider is also the most profitable.
Utilisation determines profitability. Capacity that is idle earns nothing and depreciates anyway, so providers price aggressively to fill it, and offer substantial discounts for commitments and for tolerance of interruption.
Pricing is a competitive instrument that has moved almost entirely downward for the underlying commodities, storage and computing time, while the value has migrated to higher-level managed services where price comparison is harder. Egress charges, levied on data leaving a provider's network, became a subject of regulatory attention because they raise the cost of moving to a competitor, and were reduced or removed by several providers in response.
Reserved and committed spending arrangements, in which a customer commits to a level of expenditure over several years in exchange for discounts, are now the normal form of enterprise contract and produce the remaining performance obligation figures that investors watch as an indicator of future revenue.
Concentration and lock-in
Three companies supply roughly two-thirds of global cloud infrastructure, and the proportion has been stable or rising for a decade. Concentration of this degree in infrastructure on which most of the economy's software depends has attracted increasing scrutiny.
The mechanisms of lock-in are several and mostly not contractual. Data has mass: moving petabytes between providers is slow and, where egress is charged, expensive. Managed services differ between providers, so an application built on one provider's database, queueing and identity services must be substantially rewritten to run elsewhere. Skills are provider-specific, and an organisation that has trained its engineers on one platform faces a real cost in retraining. And discount structures reward concentration of spending with a single supplier.
Containers and Kubernetes reduce these frictions at the lowest layer without removing them at the higher layers where most of the value now sits, which is the reason providers have promoted them so readily.
Regulation and public policy
Competition
Competition authorities in several jurisdictions have examined the sector. The United Kingdom's Office of Communications referred the cloud services market to the Competition and Markets Authority in 2023, citing egress fees, committed spending discounts and technical barriers to switching. The European Commission has examined software licensing practices under which a customer running certain vendors' software on a competitor's infrastructure pays more than on the vendor's own, a practice complained of by European providers and the subject of settlements with some of them. In the United States the Federal Trade Commission has conducted market studies of the sector, including of the partnerships between cloud providers and artificial intelligence developers.
The question authorities have found difficult is that the concentration arises from genuine economies of scale rather than from exclusionary conduct, and that customers benefit from prices that have fallen consistently. The remedies proposed have accordingly concentrated on switching costs rather than on structure.
Data sovereignty and jurisdiction
The most consequential regulatory issue is jurisdictional. Data stored by an American provider may be subject to demands from American authorities regardless of where it physically sits, a position confirmed by the CLOUD Act of 2018. European regulators and courts have repeatedly held that this creates a conflict with the General Data Protection Regulation, most notably in the Schrems II decision of 2020, which invalidated the arrangement then governing transfers of personal data between the European Union and the United States.
The commercial response has been the sovereign cloud: capacity operated within a jurisdiction, in several cases by a locally owned entity under licence, with technical controls intended to prevent access from outside. Whether such arrangements genuinely resolve the legal conflict is contested, and the growth of European providers positioning themselves on this basis, together with initiatives such as Gaia-X, reflects a policy preference in several European states for infrastructure not controlled from elsewhere.
Resilience
As more of the economy's critical functions moved onto a small number of platforms, financial regulators in particular began treating the providers as systemically important. The European Union's Digital Operational Resilience Act, applying from 2025, brings critical third-party technology providers to financial firms within the supervisory perimeter, and comparable regimes exist or are being developed in the United Kingdom and elsewhere.
The concern is straightforward. Outages affecting a single provider's region have simultaneously disrupted payments, airlines, broadcasters, retailers and government services, and the correlation of failure across apparently unrelated organisations is a systemic risk of a kind that did not exist when each ran its own systems.
Energy and environment
Data centres consumed a small but rising share of global electricity through the 2010s, held down by efficiency improvements that offset growth in demand for most of the period. The artificial intelligence build-out since 2022 broke that pattern: power demand has risen faster than efficiency has improved, and data centre load has become a material factor in electricity planning in Ireland, parts of the United States, the Netherlands and elsewhere, prompting moratoria on new connections in some jurisdictions.
The providers are among the largest corporate purchasers of renewable electricity in the world and have each set targets for carbon-free operation. The criticisms are that matching annual consumption with renewable purchases is not the same as running on clean power at every hour, which providers have begun to address through hourly matching commitments; that water consumption for cooling is substantial in regions where water is scarce; and that the growth in absolute demand may outpace any plausible rate of clean generation. Several providers have contracted for nuclear generation, including restarting retired plants and funding small modular reactor development.
Effects
On software
The industry made it possible to start a software company for a few hundred dollars a month, which changed the structure of venture financing, the number of companies founded and the speed at which products reach users. It also made possible applications whose demand is unpredictable or intermittent, since capacity can be acquired for an hour.
The architectural consequences were equally large. Software designed for elastic infrastructure is built differently: as many small services rather than one large program, stateless where possible so that instances can be created and destroyed freely, and designed to tolerate the failure of any individual component because at sufficient scale something is always failing.
On enterprise information technology
The corporate data centre has not disappeared but has ceased to be the default. The occupations associated with it, in facilities, hardware and system administration, have contracted, while the ones associated with cloud engineering, platform operations and cost management have grown. The category of financial operations, concerned with controlling cloud expenditure, did not exist before the industry and now supports a substantial software market of its own, because metered pricing turns out to produce bills that surprise organisations accustomed to fixed costs.
On geopolitics
Cloud infrastructure has become an element of state capability. The location of data centres, the nationality of the companies operating them, the legal regimes governing access to data and the availability of advanced processors are all matters of national policy. Export controls on high-performance processors imposed by the United States from 2022 restrict which capabilities may be supplied to particular countries, and providers have had to design offerings around them. The result is a partial fragmentation of what had been a genuinely global market into regional blocs, with China served almost entirely by domestic providers and Europe pursuing arrangements intended to reduce dependence.
Assessment
The cloud industry is the clearest example in the modern economy of a transformation that made something cheaper and more available while concentrating control of it.
The benefits are not seriously disputed. Computing that once required capital, premises, specialist staff and a planning horizon of years is now available immediately and billed by use. That change is responsible for a substantial part of the growth in software businesses since 2006 and for the fact that artificial intelligence development, which requires resources beyond almost any organisation's means, is possible outside a handful of states and corporations.
The difficulties follow from the same structure. Three companies operate the infrastructure on which most of the world's software runs, the barriers to becoming a fourth are measured in tens of billions of dollars a year, and the technical characteristics of the services make customers progressively harder to move. Regulators have found this hard to address because the concentration arises from efficiency rather than from misconduct, and because the customers who would be protected mostly do not wish to leave.
The constraint that may reshape the industry next is physical rather than commercial. For its first two decades the sector's growth was limited by capital and by demand; since 2022 it has been limited increasingly by electricity, by grid connections and by the supply of specialised processors. Those are constraints that money alone resolves slowly, and they have already begun to determine where capacity is built and which customers obtain it.