AI Wiki

Hyperscaler

AI EnergyAI HardwareAI InfrastructureData Centers
20 min read
Updated
Suggest editHistoryTalk39 citationsFact-checked
RawGraph

A hyperscaler is a company that builds and operates computing infrastructure at a scale far beyond a conventional enterprise IT estate: globally distributed fleets of data centers holding millions of servers, connected by privately owned networks, and increasingly filled with hardware the operator designed itself. IBM's reference glossary defines hyperscale as "a distributed computing environment and architecture that is designed to provide extreme scalability to accommodate workloads of massive scale", and applies the word hyperscaler to the data centers themselves, which it describes as considerably larger than traditional on-premises facilities; it also cites an IDC threshold of at least 5,000 servers and 10,000 square feet.[1] Common industry usage, including that of the market research firms that size the segment, instead attaches the label to the operating companies, and defines the group by naming its members rather than by applying a server count.

The list usually starts with the three companies that sell most of the world's cloud computing capacity, Amazon Web Services, Microsoft Azure and Google Cloud, and then adds operators that run comparable fleets primarily for their own products. Meta is the largest of those captive operators. In China the corresponding operators are Alibaba Cloud, Tencent and ByteDance. Synergy Research Group, whose hyperscale tracker is widely cited for the category, builds its data from the footprints of 19 major cloud and internet service firms spanning SaaS, IaaS, PaaS, search, social networking, e-commerce and gaming, and ranks Amazon, Microsoft and Google first by capacity, followed by Meta, Alibaba, Tencent, Apple and ByteDance.[2][3]

The defining feature of the group since 2023 has been the scale of its capital spending on AI infrastructure. Guidance issued by Amazon, Alphabet, Microsoft and Meta during 2026 implies combined capital expenditure above $700 billion for the calendar year,[11][17][22][25] roughly double the combined $360 billion those four reported for their most recent full years on the differing bases each uses.[12][21][23][24] The resulting build-out has turned electricity procurement into a central constraint on the industry.[37]

What counts as a hyperscaler

Hyperscale is best understood as a set of operating characteristics rather than a size cutoff. A hyperscaler owns or leases its own buildings rather than renting racks; designs its own server, rack and network hardware instead of buying reference systems; runs a private global backbone rather than buying transit; and amortizes engineering effort across a fleet large enough to justify in-house design teams. Google Cloud reports 43 regions, 130 zones, more than 200 network edge locations and about 10 million kilometers of terrestrial and subsea fiber.[4] AWS reports 123 Availability Zones across 39 geographic regions with nearly 20 million kilometers of fiber-optic cabling, plus announced plans for two further regions in Saudi Arabia and Chile.[5] Microsoft advertises more than 80 Azure regions and more than 500 data centers.[6]

Synergy counted 1,360 hyperscale data centers in operation at the end of the fourth quarter of 2025, up from 1,136 at the end of 2024, and estimated that hyperscale operators accounted for 48 percent of the worldwide capacity of all data centers. Almost 60 percent of that hyperscale capacity sat in own-built, owned facilities, with the balance leased. Non-hyperscale colocation held another 20 percent and enterprise on-premises facilities the remaining 32 percent, a reversal from 2018 when on-premises accounted for 56 percent. Synergy projects hyperscale share reaching 67 percent by 2031.[2][3]

A distinction worth keeping is between commercial and captive hyperscalers. Amazon, Microsoft, Google, Alibaba and Tencent sell infrastructure to third parties, and Synergy also counts ByteDance among the fastest-growing tier-two cloud providers; Meta and Apple operate at hyperscale mainly to serve their own products, so their fleets show up in data center capacity rankings but not in cloud market share tables. Oracle is sometimes grouped with the leaders, but as of the first quarter of 2026 Synergy placed it among the tier-two cloud providers rather than in the leading group.[7]

Cloud market position

Enterprise spending on cloud infrastructure services reached $128.6 billion in the first quarter of 2026, a 35 percent year-on-year increase and the ninth consecutive quarter of accelerating growth, taking trailing-twelve-month revenues to $455 billion. Worldwide market shares for the quarter were 28 percent for Amazon, 21 percent for Microsoft and 14 percent for Google; in public IaaS and PaaS alone the three held 67 percent.[7]

OperatorLatest reported cloud revenueGrowthPeriod
AWS$37.6 billion (segment sales)28% year over yearQuarter ended 31 March 2026[8]
Microsoft Cloud$54.5 billion29% year over yearQuarter ended 31 March 2026[9]
Google Cloud$24.8 billion82% year over yearQuarter ended 30 June 2026[10]
AWS (full year)$128.7 billion20% year over yearCalendar 2025[11]
Google Cloud (full year)$58.7 billion36% year over yearCalendar 2025[12]

AWS remains the revenue leader and reported segment operating income of $14.2 billion in the first quarter of 2026 on $37.6 billion of sales.[8] Google Cloud's growth accelerated sharply through the first half of 2026, and Alphabet disclosed a revenue backlog of $519.5 billion in remaining performance obligations as of 30 June 2026, of which $513.9 billion related to Google Cloud.[13] Alphabet's segment description also changed during 2026 to say that Google Cloud generates product revenues primarily from the sale of TPU systems, language absent from the equivalent description in its 2025 annual report, which listed only consumption-based fees and subscriptions.[12][13]

Custom silicon and system design

Every large hyperscaler now designs AI accelerators for its own fleet, and the three commercial leaders also design general-purpose server CPUs, partly to reduce dependence on NVIDIA and partly because owning the design lets them tune memory, interconnect and packaging to their own workloads.

OperatorAI acceleratorServer CPUNotes
AWSTrainium, InferentiaGravitonTrainium2 fully subscribed with 1.4 million chips landed; Trainium3 UltraServers scale to 144 chips and 362 MXFP8 PFLOPs[11][14]
GoogleTPU (Ironwood, then TPU 8t and TPU 8i)AxionIronwood pods reach 9,216 liquid-cooled chips and 42.5 FP8 exaflops, spanning nearly 10 MW[15]
MicrosoftMaia 100, Maia 200CobaltMaia 200 live in Iowa and Arizona data centers; Cobalt deployed in nearly half of Azure datacenter regions[16][17]
MetaMTIAnone announcedMTIA serves ranking and recommendation inference, primarily ads workloads[18][19]

Amazon reported that its chips business, covering Graviton, Trainium and the Nitro system, exceeded a $20 billion annual revenue run rate in the first quarter of 2026 while growing at triple-digit percentages.[8] Project Rainier, the cluster Amazon describes as the world's largest operational AI compute cluster with more than 500,000 Trainium2 chips, is used by Anthropic to train Claude.[11] Amazon has also said OpenAI committed to consume roughly two gigawatts of Trainium capacity beginning to ramp in 2027, and that Anthropic will secure up to five gigawatts of current and future Trainium generations.[8]

Google unveiled its eighth TPU generation at Cloud Next in April 2026, splitting the line into a training part (TPU 8t) and an inference part (TPU 8i), the latter claimed to deliver 80 percent better performance per dollar, alongside a custom-built system for connecting supercomputers called the Virgo Network.[20] Microsoft says Maia 200 offers over 30 percent improved tokens per dollar compared with the latest silicon in its fleet.[17] The second-generation MTIA, announced in April 2024, is a TSMC 5nm part running at 1.35 GHz with a 90W thermal design point, 256 MB of on-chip memory and 128 GB of LPDDR5; Meta reported going from first silicon to production models in 16 regions in under nine months.[19] By late 2025 Meta said a training chip for ranking and recommendations was ramping into production, with further parts in development.[18]

Custom silicon is only part of the picture. Hyperscalers also design racks, cooling and power distribution, and publish some of those designs through the Open Compute Project. Meta's Catalina rack houses 72 NVIDIA Blackwell GPUs across two racks drawing roughly 140 kW, cooled with air-assisted liquid cooling because Meta's older buildings lack facility water; a six-rack pod delivers 360 PFLOPS of FP16 compute.[18] At those densities, operators and analysts alike describe new capacity in megawatts and gigawatts rather than floor area.[2][17][18]

The capital expenditure race

The four large US operators have raised capital expenditure guidance repeatedly through 2025 and 2026. Definitions differ between companies, so the figures below are not strictly comparable.

Company2025 reported capital expenditure2026 guidanceBasis
Amazon$131.8 billionabout $200 billionPurchases of property and equipment, calendar year[21][11]
Alphabet$91.4 billion$195 billion to $205 billionPurchases of property and equipment, calendar year[12][22]
Microsoft$64.6 billion (fiscal year to 30 June 2025)roughly $190 billion for calendar 2026Additions to property and equipment; guidance includes finance leases[23][17]
Meta$72.2 billion$125 billion to $145 billionIncludes principal payments on finance leases, calendar year[24][25]
AlibabaRMB126.1 billion (about $18.3 billion)not disclosedFiscal year to 31 March 2026[26]

The revisions came fast. Alphabet guided to $175 billion to $185 billion for 2026 at its fourth-quarter 2025 results in February, lifted the range to $180 billion to $190 billion in April, then to $195 billion to $205 billion on 22 July 2026, with CFO Anat Ashkenazi attributing the last increase to "an acceleration in the delivery of capacity to meet growing demand" and describing the company as still operating in "a supply-constrained environment".[22][27] Meta raised its 2026 range from $115 billion to $135 billion up to $125 billion to $145 billion in April, citing higher component pricing and additional data center costs.[24][25] Microsoft's roughly $190 billion calendar-2026 figure includes about $25 billion attributable to higher component prices alone.[17]

Memory pricing is the component cost most often identified behind those revisions, amid a supply crunch driven by AI demand.[30] Microsoft guided fourth-quarter fiscal 2026 capital expenditure above $40 billion, of which roughly $5 billion was attributed to component pricing, and told investors it expected to remain capacity constrained at least through 2026.[17]

Synergy put hyperscale operator capital expenditure at $127 billion in the second quarter of 2025, up 72 percent year on year, and noted that capital intensity across the group rose from under 9 percent of revenues in 2021 to over 16 percent in the first half of 2025, a level comparable to telecommunications operators.[28]

Financing has begun to shift as a result. Alphabet's free cash flow turned negative in the second quarter of 2026, at minus $5.9 billion, as quarterly capital expenditure reached $44.9 billion.[10] In June 2026 the company took in $49.6 billion of net equity proceeds: $30.5 billion from common stock, comprising an underwritten public offering completed on 4 June and a concurrent $10.0 billion private placement to an affiliate of Berkshire Hathaway, and $19.1 billion from an issue of 6.25 percent mandatory convertible preferred stock the following day. Alphabet said the proceeds were for general corporate purposes "including capital expenditures to scale AI infrastructure and global compute". It also issued senior unsecured notes for net proceeds of $20.3 billion during the quarter and entered a $40 billion at-the-market equity distribution program.[13] Amazon's trailing-twelve-month free cash flow fell to $1.2 billion by the first quarter of 2026, driven primarily by a $59.3 billion year-on-year increase in property and equipment purchases.[8] Meta moved part of one build off its own balance sheet: in October 2025 it formed a venture to co-develop a Louisiana data center campus, contributing $4.3 billion of assets, taking a 20 percent membership interest, receiving a $2.55 billion distribution, and committing to a pro rata share of roughly $27 billion in estimated development costs while leasing the resulting properties back.[29]

Depreciation and accounting scrutiny

Because AI hardware is expensive and short-lived, the useful lives hyperscalers assign to servers materially affect reported profits, and the assumptions have moved in both directions. Alphabet depreciates servers and network equipment generally over six years and data center buildings over seven to 40 years.[12] Meta increased the estimated useful lives of most servers and network assets to 5.5 years effective 1 January 2025.[29] Amazon went the other way: after extending server lives from five to six years effective January 2024, it cut a subset of servers and networking equipment back from six to five years effective January 2025, citing "the increased pace of technology development, particularly in the area of artificial intelligence and machine learning".[21] Microsoft said that in the quarter ended 31 March 2026 roughly two thirds of its capital expenditure went to short-lived assets, primarily GPUs and CPUs, with the remainder in long-lived assets it expects to support monetization over 15 years and beyond.[17] Rising depreciation is already visible in margins: Microsoft's gross margin fell to 67.6 percent in the quarter ended 31 March 2026, its narrowest since 2022, on data center depreciation.[30]

Power, siting and nuclear procurement

The International Energy Agency estimated that data centers consumed about 415 TWh of electricity in 2024, roughly 1.5 percent of the world total, with the United States accounting for 45 percent, China 25 percent and Europe 15 percent. Its base case projects consumption more than doubling to around 945 TWh by 2030 and reaching roughly 1,200 TWh by 2035.[31] Because hyperscale capacity is geographically concentrated, that demand lands on a small number of grids, and securing firm generation has become a procurement priority for every large operator. See AI energy consumption for the broader picture.

The most visible response has been long-term nuclear contracting.

BuyerCounterpartyAssetContracted capacityTerms
MicrosoftConstellation EnergyCrane Clean Energy Center, a restart of Three Mile Island Unit 1about 835 MW20-year PPA signed September 2024; return to service targeted for 2027[32][33]
AmazonTalen EnergySusquehannaup to 1,920 MW at full contract quantityPPA through 2042, expanded June 2025, full volume no later than 2032[34]
MetaConstellation EnergyClinton Clean Energy Centerfull plant output plus a 30 MW uprate20-year PPA beginning June 2027[33]
GoogleKairos Powerfleet of small modular reactorsup to 500 MWFirst reactor targeted for 2030, further deployments through 2035[35]

The Crane restart also drew federal support: in November 2025 the US Department of Energy issued a guarantee for up to $1 billion for a loan to finance it, maturing in October 2055.[36] Alongside the Talen agreement, AWS announced a $20 billion investment in Pennsylvania, which it described as the largest private-sector investment in the state's history.[34]

Individual sites have grown to the point where they are planned as power projects. Meta's Prometheus cluster is a one-gigawatt build spanning five or more data center buildings in a single region, supplemented by weatherproof tents and adjacent colocation space; its Hyperion site is expected to begin coming online in 2028 and to scale to five gigawatts.[18] Microsoft said it added another gigawatt of capacity in the quarter ended 31 March 2026 and remained on track to double its overall footprint within two years.[17]

Supply-side constraints are now partly political. Synergy's July 2026 assessment found total US data center capacity on course to double within three years, and hyperscale capacity within two, drawing on a known pipeline of almost 1,500 large data centers worldwide, almost half of them in the United States; the US portion alone accounts for around 45 gigawatts of IT capacity to be added by 74 different companies. Synergy added that constrained power availability and rising local concerns are crimping many new plans, while expecting developers to keep finding ways around those obstacles.[37]

Hyperscalers, neoclouds and colocation

Three business models are frequently conflated. Colocation providers lease space, power and cooling to tenants who supply their own equipment, rather than selling compute capacity itself. Neoclouds sell GPU capacity as a service but focus narrowly on accelerated compute rather than the broad service catalog of a general-purpose cloud. Hyperscalers do both and more, and they are also colocation customers: Synergy puts almost 60 percent of hyperscale capacity in own-built facilities, leaving just over 40 percent in leased space.[2]

Neocloud revenues reached $9 billion in the fourth quarter of 2025, up 223 percent year on year, and exceeded $25 billion for full-year 2025; Synergy forecasts the segment approaching $400 billion by 2031 at a 58 percent compound annual growth rate, and identifies CoreWeave, Crusoe, Core Scientific, Lambda, Nebius and Nscale as the leading providers, with CoreWeave the most direct challenger to the incumbents.[38] By the first quarter of 2026 five neoclouds ranked among the top thirty cloud providers and the category held about 5 percent of the overall cloud market, with a substantially larger share of AI-specific segments.[7]

The boundaries are blurring in both directions. In May 2026 Blackstone announced a joint venture with Google to build a TPU-based cloud, backed by an initial $5 billion equity commitment and an expected 500 MW of capacity coming online in 2027, with Google supplying TPUs, software and services.[39] In the other direction, Amazon signed an agreement for Meta to deploy tens of millions of Graviton cores for CPU-intensive agentic workloads, and Meta is already an Amazon Bedrock customer.[8]

Criticism

The main line of criticism concerns whether the demand justifying the spending is durable, and whether parts of it are circular. Large blocks of hyperscale AI capacity are contracted to a small number of model developers that are also equity investees of their infrastructure suppliers. Amazon has said Anthropic will secure up to five gigawatts of Trainium capacity, and in the first quarter of 2026 Amazon recognized $16.8 billion of pre-tax gains on its Anthropic investments.[8] Alphabet's second-quarter 2026 other income and expense of $98.0 billion likewise included $99.0 billion of net gains on equity securities; the filing attributed those primarily to unrealized gains on SpaceX and an unnamed private company, while CNBC described the line as reflecting stakes including Anthropic and SpaceX.[13][22] Synergy's counterargument, made in September 2025, is that generative AI was already producing an incremental $50 billion per quarter in hyperscaler revenue on top of prior trends, and that capital intensity in the mid-teens is normal for infrastructure industries.[28]

Two other criticisms are more concrete. First, depreciation schedules on assets that may become obsolete faster than their assigned useful lives can flatter current earnings, and the operators do not agree on those lives: Amazon assigns five years to a subset of its servers and networking equipment, while Alphabet depreciates servers and network equipment generally over six years and Meta uses 5.5 years for most such assets.[12][21][29] Second, siting and grid impact are contested locally, and Synergy's own analysis identifies power availability and public opposition as material constraints on the announced pipeline.[37]

See also

References

  1. ^IBM, "What is hyperscale?" ibm.com/...hyperscale
  2. ^Synergy Research Group, "Hyperscale Operators to Account for 67% of all Data Center Capacity by 2031", 7 April 2026. srgresearch.com/...ll-data-center-capacity-by-2031
  3. ^Synergy Research Group, "Hyperscale Data Center Count Hits 1,136; Average Size Increases; US Accounts for 54% of Total Capacity", 19 March 2025. srgresearch.com/...counts-for-54-of-total-capacity
  4. ^Google Cloud, "Global Locations: Regions and Zones" (page updated 23 July 2026). cloud.google.com/...locations
  5. ^Amazon Web Services, "Global Infrastructure". aws.amazon.com/...global-infrastructure
  6. ^Microsoft, "Azure global infrastructure". azure.microsoft.com/...global-infrastructure
  7. ^Synergy Research Group, "Cloud Market Annual Revenue Run Rate Topped Half a Trillion Dollars in Q1 as Growth Surge Continues", 29 April 2026. srgresearch.com/...in-q1-as-growth-surge-continues
  8. ^Amazon.com, Inc., "Amazon.com Announces First Quarter Results", 29 April 2026 (Form 8-K, Exhibit 99.1). sec.gov/...amzn-20260331xex991
  9. ^Microsoft Corporation, FY26 Q3 earnings press release, quarter ended 31 March 2026. microsoft.com/...press-release-webcast
  10. ^Alphabet Inc., "Alphabet Announces Second Quarter 2026 Results", 22 July 2026 (Form 8-K, Exhibit 99.1). sec.gov/...googexhibit991q22026
  11. ^Amazon.com, Inc., "Amazon.com Announces Fourth Quarter Results", 5 February 2026 (Form 8-K, Exhibit 99.1). sec.gov/...amzn-20251231xex991
  12. ^Alphabet Inc., Annual Report on Form 10-K for the year ended 31 December 2025, filed 5 February 2026. sec.gov/...goog-20251231
  13. ^Alphabet Inc., Quarterly Report on Form 10-Q for the quarter ended 30 June 2026, filed 23 July 2026. sec.gov/...goog-20260630
  14. ^Amazon Web Services, "AWS Trainium" product page. aws.amazon.com/...trainium
  15. ^Google, "Ironwood: The first Google TPU for the age of inference", 9 April 2025. blog.google/...ironwood-tpu-age-of-inference
  16. ^Microsoft Azure Blog, "Microsoft Azure delivers purpose-built cloud infrastructure in the era of AI", 15 November 2023. azure.microsoft.com/...astructure-in-the-era-of-ai
  17. ^Microsoft Corporation, FY26 Q3 earnings conference call transcript, 29 April 2026. microsoft.com/...earnings-fy-2026-q3
  18. ^Meta Engineering, "Meta's Infrastructure Evolution and the Advent of AI", 29 September 2025. engineering.fb.com/...olution-and-the-advent-of-ai
  19. ^Meta AI, "Our next generation Meta Training and Inference Accelerator", 10 April 2024. ai.meta.com/...ining-inference-accelerator-AI-MTIA
  20. ^Google, "7 highlights from Google Cloud Next '26", 24 April 2026. blog.google/...google-cloud-next-26-recap
  21. ^Amazon.com, Inc., Annual Report on Form 10-K for the year ended 31 December 2025, filed 6 February 2026. sec.gov/...amzn-20251231
  22. ^CNBC, "Alphabet earnings takeaways: Q2 revenue beats, GOOGL stock sinks on 2026 capex hike", 22 July 2026. cnbc.com/...google-earnings-q2-goog-live-updates
  23. ^Microsoft Corporation, Annual Report on Form 10-K for the fiscal year ended 30 June 2025, filed 30 July 2025. sec.gov/...msft-20250630
  24. ^Meta Platforms, Inc., "Meta Reports Fourth Quarter and Full Year 2025 Results", 28 January 2026 (Form 8-K, Exhibit 99.1). sec.gov/...meta-12312025xexhibit991
  25. ^Meta Platforms, Inc., "Meta Reports First Quarter 2026 Results", 29 April 2026 (Form 8-K, Exhibit 99.1). sec.gov/...meta-03312026xexhibit991
  26. ^Alibaba Group Holding Limited, Fiscal Year 2026 Annual Report (Form 6-K, Exhibit 99.1), 18 June 2026. sec.gov/...d133513dex991.pdf
  27. ^CNBC, "Alphabet resets the bar for AI infrastructure spending", 4 February 2026. cnbc.com/...the-bar-for-ai-infrastructure-spending
  28. ^Synergy Research Group, "Justifying the Explosive Growth in Hyperscale CAPEX", 18 September 2025. srgresearch.com/...sive-growth-in-hyperscale-capex
  29. ^Meta Platforms, Inc., Annual Report on Form 10-K for the year ended 31 December 2025, filed 29 January 2026. sec.gov/...meta-20251231
  30. ^CNBC, "Microsoft calls for $190 billion in 2026 capital spending on soaring memory prices", 29 April 2026. cnbc.com/...microsoft-msft-q3-earnings-report-2026
  31. ^International Energy Agency, "Energy and AI", executive summary, April 2025. iea.org/...executive-summary
  32. ^Constellation Energy Corporation, "Constellation to Launch Crane Clean Energy Center, Restoring Jobs and Carbon-Free Power to the Grid", 20 September 2024 (Form 8-K, Exhibit 99.1). sec.gov/...ceg-202409208kexh991
  33. ^Constellation Energy Corporation, "Constellation Reports Second Quarter 2025 Results", 7 August 2025 (Form 8-K, Exhibit 99.1). sec.gov/...ceg-20250807991
  34. ^Talen Energy Corporation, "Talen Energy Expands Nuclear Energy Relationship with Amazon", 11 June 2025 (Form 8-K, Exhibit 99.1). sec.gov/...a20250611pressreleasebusin
  35. ^Google, "New nuclear clean energy agreement with Kairos Power", 14 October 2024. blog.google/...iros-power-nuclear-energy-agreement
  36. ^Constellation Energy Corporation, "Constellation Reports Fourth Quarter and Full Year 2025 Results", 24 February 2026 (Form 8-K, Exhibit 99.1). sec.gov/...ceg-20260224991
  37. ^Synergy Research Group, "Synergy Reports that U.S. Data Center Capacity Will Continue to Balloon Despite Increasing Headwinds", 23 July 2026. srgresearch.com/...on-despite-increasing-headwinds
  38. ^Synergy Research Group, "Neocloud Market Forecast to Approach $400B by 2031, Driven by Surging AI Infrastructure Demand", 2 April 2026. srgresearch.com/...urging-ai-infrastructure-demand
  39. ^Google, "Blackstone will create a new TPU cloud in a joint venture with Google", 19 May 2026. blog.google/...blackstone-tpu-cloud

Improve this article

Add missing citations, update stale details, or suggest a clearer explanation. Every suggestion is reviewed for sourcing before it goes live.

v1 · 3,900 words · full history

Fact-checks are independent of edits: a reviewer re-verifies the article against its sources and stamps the date. How we verify

Research and drafting on this wiki are AI-assisted, under named human editorial standards. How AI is used here

Reviewer note: Independent adversarial fact-check at creation (wanted175 campaign, 2026-07-24): every claim verified against primary sources by a dedicated verification agent; corrections applied before publication.

Cite this page: AI Wiki. "Hyperscaler." aiwiki.ai, updated 24 Jul 2026, fact-checked 24 Jul 2026. CC BY 4.0. https://aiwiki.ai/wiki/hyperscaler

Suggest edit