A hyperscaler is a company that owns and operates computing infrastructure at such scale that it can add servers, storage and network capacity almost without limit as demand rises, and rent that capacity to others on demand. In everyday use the word points to three firms: Amazon Web Services (AWS), Microsoft Azure and Google Cloud. Wider definitions add Oracle, Alibaba, Meta and Apple. The model exists because a few operators building identical facilities by the hundred can buy chips, power and land more cheaply, per unit, than any single customer running its own servers.

How scale is defined

No single threshold exists. IDC, the market research firm, describes a hyperscale data centre as one with at least 5,000 servers and 10,000 square feet of floor space, according to IBM. Synergy Research Group uses a different test. It classes a company as a hyperscale operator by the size of its cloud, search, social media, e-commerce or gaming business, not by building size, and tracks around 20 such firms.

The output of those firms is measured in megawatts of critical IT load. At the end of 2024, Synergy counted 1,136 hyperscale data centres, roughly double the number five years earlier, and found that Amazon, Microsoft and Google alone held 59% of hyperscale capacity. By the end of 2025, hyperscale operators held 48% of all data centre capacity worldwide, against 20% for colocation providers and 32% for companies running their own facilities, according to Synergy's April 2026 release. The firm forecasts the hyperscale share reaching 67% by 2031.

How the model works

A hyperscaler builds standardised facilities grouped into regions, each split into availability zones with separate power and cooling, so one failure should not take down a whole region. Servers are commodity hardware, often custom designed: Google runs its own Tensor Processing Units (TPUs), and Amazon its Trainium and Inferentia chips. Software abstracts the hardware, so a customer rents a virtual machine, a database or an artificial intelligence (AI) model by the second, the gigabyte or the token rather than buying a server.

The services sit in layers. Infrastructure as a service (IaaS) rents raw compute and storage; platform as a service (PaaS) adds managed databases, analytics and machine learning tools; software as a service (SaaS) sits on top. Gartner put the IaaS market at $171.8 billion in 2024, with Amazon at 37.7%, Microsoft at 23.9%, Google at 9.0%, Alibaba at 7.2% and Huawei at 4.1%. Synergy measures a broader category, cloud infrastructure services, and estimated spending of $143 billion in the second quarter of 2026 alone, up 43% year on year, with Amazon at 28%, Microsoft at 20% and Google at 15%.

Pricing follows the same logic of scale. Customers typically pay little or nothing to move data in and pay egress fees to move it out. That asymmetry is central to the lock-in debate covered below.

Origin and evolution

The adjective came from data centre engineering. An early recorded use is a February 2008 Dell article titled "Creating a Hyper-efficient Hyper-scale Data Center", tied to its Data Center Solutions unit, which sold stripped-down servers to large internet companies. The underlying business is older. AWS launched its Simple Storage Service in March 2006 and Elastic Compute Cloud in August 2006, Google App Engine arrived in 2008 and Microsoft's Windows Azure became commercially available in February 2010. Google's BigQuery, now used across advertising analytics, was announced on May 19, 2010.

Through the 2010s "hyperscaler" shifted from describing buildings to describing companies. The AI boom then changed what the companies build. Graphics processing units (GPUs) and custom accelerators draw far more power per rack than general servers, and campuses are now planned in gigawatts. Microsoft's Mount Pleasant site in Wisconsin and Meta's Prometheus and Hyperion clusters, the latter planned to reach 5 gigawatts, illustrate the change.

The scale of spending

Capital expenditure is the clearest measure. Alphabet spent $91.4 billion in 2025, according to its fourth-quarter release, up from about $31 billion in 2022. In July 2026 it raised 2026 guidance to $195 billion to $205 billion after second-quarter capex of $44.9 billion. Amazon's purchases of property and equipment reached $131.8 billion in 2025; Andy Jassy, chief executive, said on the July 2026 call that the company expected about $220 billion in 2026. Microsoft spent $41.0 billion in its fiscal fourth quarter to June 2026 and guided to about $175 billion for calendar 2026, a figure lowered partly by how leases are classified. Meta narrowed its 2026 range to $130 billion to $145 billion, and Oracle expects fiscal 2027 capex of $90 billion to $95 billion.

Taken together, the five plan to spend roughly $800 billion in 2026, as of October 2026. For comparison, Synergy's trailing 12-month cloud infrastructure market was about $500 billion. Cloud revenue is rising too: Google Cloud grew 82% to $24.8 billion in the second quarter of 2026, AWS grew 37% to $42.2 billion, and Microsoft said Azure passed $100 billion in annual revenue.

Why it matters for marketing

Much of the advertising stack runs on hyperscaler infrastructure. Ad tech historically relied on colocation and owned servers to keep auction latency low, but AWS has become central infrastructure for advertising beyond Amazon's own business. In October 2025 it launched RTB Fabric, a dedicated network for real-time bidding claiming single-digit millisecond latency and up to 80% lower networking costs, with TripleLift, Viant and Yieldmo among early partners. Others hedge. Magnite runs connected television transactions on cloud infrastructure and most other traffic in its own data centres, while PubMatic states in its annual report that owning hardware saves money compared with public cloud.

Measurement is the second link. Clean rooms such as Google's Ads Data Hub, built on BigQuery, Amazon Marketing Cloud, described by Amazon as built on AWS, and AWS Clean Rooms, generally available since March 21, 2023, run on hyperscaler stacks. Query, storage and egress pricing therefore becomes part of the cost of measuring a campaign.

AI is the third. Generative tools for creative, bidding and search are trained and served on hyperscaler compute, and the advertising giants are themselves the largest hyperscalers. Sundar Pichai said Google was "supply-constrained", naming memory, wafer supply, power and permitting as limits.

Concentration, lock-in and energy

Critics raise three problems. The first is concentration. When a latent race condition in the DNS management system for DynamoDB broke AWS's us-east-1 region between October 19 and 20, 2025, according to AWS's own post-event summary, Snapchat, Fortnite and several UK banks were disrupted. Nine days later an Azure Front Door configuration error caused a separate global disruption.

The second is lock-in. Regulators have focused on egress fees, committed-spend discounts and software licensing. The UK Competition and Markets Authority (CMA) concluded its market investigation on July 31, 2025, finding competition was not working well. In March 2026 it opted for a strategic market status (SMS) investigation into Microsoft alone, opened on May 14, 2026, after AWS offered commitments on egress and interoperability. The EU Data Act, applicable since September 12, 2025, bans switching charges from January 12, 2027. Google, AWS and Microsoft each waived exit egress fees in early 2024. Sovereignty is a related dispute: Microsoft France's Anton Carniaux told the French Senate in June 2025 that he could not guarantee that data would be shielded from US authorities.

The third is energy. Data centres used about 415 terawatt-hours in 2024, roughly 1.5% of global electricity, and the International Energy Agency (IEA) projects about 945 terawatt-hours by 2030. Projects worth $156 billion were blocked or delayed by local opposition in 2025.

Financing is a fourth, newer concern. S&P cut Oracle to BBB- in July 2026, and UBS estimates cited by PPC Land suggest OpenAI and Anthropic could account for 48% of Google Cloud revenue in 2027. Bulls see contracted backlog; sceptics see circular dependence on a few AI laboratories.

Not the same as

Cloud provider is the broader category. Every hyperscaler is one, but a regional host renting a few thousand servers is not a hyperscaler.

Colocation operators such as Equinix and Digital Realty rent space, power and cooling in which customers install their own equipment. They often lease capacity to hyperscalers but do not run the cloud services themselves.

Neoclouds are specialist GPU clouds such as CoreWeave, Crusoe, Nebius and Nscale. CoreWeave, founded in 2017 as a cryptocurrency miner, listed on Nasdaq on March 28, 2025, with Microsoft accounting for 62% of its 2024 revenue. Synergy counted nine neoclouds among the top 40 cloud providers in mid-2026.

Hyperscale data centre describes a building; hyperscaler describes the company.

Recent developments

On June 25, 2026, the European Commission reached a preliminary view that AWS and Azure should be designated gatekeepers under the Digital Markets Act (DMA), following investigations opened on November 18, 2025. No final decision had been published as of October 2026. Microsoft and OpenAI rewrote their agreement in April 2026, letting OpenAI serve products on any cloud. Today, Meta's infrastructure chief Santosh Janardhan put 2026 AI spending at "well over $100 billion".

Timeline

  • March 2006 - AWS launches Simple Storage Service; Elastic Compute Cloud follows in August 2006.
  • February 2008 - Dell publishes "Creating a Hyper-efficient Hyper-scale Data Center", an early use of the term.
  • February 2010 - Windows Azure becomes commercially available.
  • May 19, 2010 - Google announces BigQuery.
  • May 24, 2017 - Google announces Ads Data Hub, built on Google Cloud infrastructure.
  • October 5, 2023 - Ofcom refers the UK cloud market to the CMA.
  • January 11, 2024 - The EU Data Act's reduced switching charge period begins; Google waives exit egress fees the same day.
  • March 2024 - AWS (March 5) and Microsoft waive egress fees for departing customers.
  • March 19, 2025 - Synergy counts 1,136 hyperscale data centres at the end of 2024.
  • March 28, 2025 - CoreWeave begins trading on Nasdaq.
  • April 2025 - IEA publishes Energy and AI, projecting about 945 TWh of data centre demand by 2030.
  • July 31, 2025 - CMA concludes its cloud market investigation.
  • September 12, 2025 - EU Data Act becomes applicable.
  • October 19-20, 2025 - AWS us-east-1 outage.
  • November 18, 2025 - European Commission opens three DMA cloud market investigations.
  • March 11, 2026 - Google completes its $32 billion acquisition of Wiz.
  • May 14, 2026 - CMA opens an SMS investigation into Microsoft's business software ecosystem.
  • June 25, 2026 - Commission reaches a preliminary view that AWS and Azure should be gatekeepers.
  • July 2026 - Second-quarter results: Alphabet raises 2026 capex guidance to $195 billion to $205 billion; Synergy reports a $143 billion quarterly cloud market.
  • January 12, 2027 - Switching charges banned under the EU Data Act.

Summary

Who. Amazon Web Services, Microsoft Azure and Google Cloud are the core hyperscalers, with Oracle, Alibaba, Meta and Apple included by broader definitions. Regulators including the European Commission and the UK CMA oversee them; ad tech companies, advertisers and AI laboratories are among their customers.

What. Operators of very large, standardised data centre fleets that rent computing, storage, networking and AI capacity on demand and expand it rapidly.

When. Public cloud began with AWS in 2006, the term appeared in data centre engineering by 2008, and the AI buildout from 2023 pushed combined capex for the five largest towards roughly $800 billion in 2026.

Where. In more than 1,100 hyperscale data centres worldwide, with the United States holding about 54% of capacity and new campuses spreading across Europe and Asia.

Why. Scale lowers unit costs for computing, which made hyperscalers the default infrastructure for software, advertising and AI. The same scale drives concerns about concentration, lock-in, outages, energy use and financing.