Skip to content

Posts tagged ‘Synergy Research Group’

Nov 12

Gartner sees 57% of servers in 2,880 large data centers by 2029

Gartner data center forecast 2029 hero graphic: installed servers by site class 2023 to 2029 with large data centers rising from 42% to 57% of all servers, the 2023 and 2025 editions of the large data center site forecast with 535 sites added for 2027, change in servers by site class 2024 to 2029 with large data centers adding 15.9 million and the other four classes losing 5.1 million, and key figures of 2,880 large sites, 17,140 servers per large site, 10.9 servers per rack and 980 large data centers in China
Four views of Gartner’s October 2025 data center sites forecast. Installed servers by site class, 2023 to 2029, with large data centers rising from 42% to 57% of all servers. The 2023 and 2025 editions of the large site count, 535 sites apart for 2027. Change in servers by site class, 2024 to 2029. Key figures on sites, servers per site, servers per rack and China against North America. Click any chart to open it full size. Sources linked below, analysis by softwarestrategiesblog.com.

Gartner’s data center forecast now puts 2,880 large data centers worldwide in 2029. By then it expects them to hold 57% of every server installed anywhere, compared to 44% in 2024. Two years earlier the same forecast put 1,955 large sites in 2027. The new edition puts 2,490 there, a 27% increase for the same year. Every other class of site, whether a single server in a closet or a 500-rack enterprise data center, loses servers through 2029.

The total site count, by comparison, barely moves. Gartner counts 3.46 million data center sites in 2024 and 3.74 million in 2029, a 1.6% compound annual growth rate. Of the 288,000 sites added, 261,000 are single-server sites. The growth that matters is in the large sites, which grow five times faster than the total, from 1,951 to 2,880 at an 8.1% CAGR. The servers move with them.

This post compares the two editions of Gartner’s data center sites forecast, published October 2023 and October 2025. It then places them beside the public counts from Synergy Research Group, JLL, the IEA and Cushman & Wakefield. Each source counts a different unit, so the comparison takes some care. Gartner counts sites and servers. Synergy counts hyperscale facilities and capacity share. JLL and Cushman & Wakefield count gigawatts, while the IEA counts terawatt-hours. All five still point the same way.

Gartner raised its large data center count 27% in two years

Gartner large data center site forecast, 2023 edition versus 2025 edition: the 2027 count rose from 1,955 to 2,490 large sites, up 535 or 27%, with 2,880 large sites by 2029
Large data center sites in the 2023 and 2025 editions of Gartner’s forecast. The 2027 figure rose 535 sites, or 27%, between editions. Sources linked below, analysis by softwarestrategiesblog.com.

Gartner’s 2023 edition of the forecast covered 2021 through 2027 and projected 1,955 large data centers when that window closed. The 2025 edition covers 2023 through 2029 and projects 2,490 large sites in 2027. The two editions overlap for five years, and the gap widens every year. For 2023 the difference is 57 sites. By 2025 it is 238, and for 2027 it is 535.

Gartner defines a large data center as a site with more than 500 racks and more than 15,000 square feet. Both editions use the same definition, so the revision reflects changed expectations. Gartner attributes the growth to AI workloads, from training to inference, through 2027, with the forecast stabilizing in later years.

More large sites, fewer servers in each one

Servers moved far less than sites did between the two editions. The 2023 edition projected 40.6 million servers in large data centers in 2027, and its 2025 successor projects 41.9 million, an increase of only 3.3%. Gartner added 535 large sites to its 2027 count but only 1.3 million servers to fill them.

Divide one by the other and the average large site in 2027 shrinks rather than grows. It falls from 20,758 servers in the 2023 edition to 16,833 in the 2025 edition, a 19% drop. The 2025 edition holds that average flat through the forecast, at 17,155 servers per large site in 2024 and 17,140 in 2029. Density per rack falls too, from 12.2 servers in 2024 to 10.9 in 2029. Rack counts in large sites rise 64% over the period, from 2.75 million to 4.52 million, while server counts rise 47%. An AI server takes more rack space and more power than the box it replaces. So a count of servers understates what is being built.

The total server base was revised down

Worldwide, the 2023 edition projected 82.6 million installed servers in 2027, while the 2025 edition projects 80.0 million, a reduction of 2.6 million. All of that reduction, and more, comes out of the smaller site classes. Enterprise data center servers for 2027 fell from 23.2 million to 21.3 million between editions, and rack and computer room servers fell from 6.6 million to 5.8 million. Gartner’s 2029 figure of 86.2 million servers implies a 2.8% CAGR from 2024. That is far slower than the 8.1% CAGR for large sites because the rest of the base is shrinking.

Only the 2023 edition broke servers into x86 and non-x86 units. It projected non-x86 servers growing from 2.0 million in 2023 to 11.9 million in 2027, with 10.5 million of them in large data centers. The 2025 edition drops that split, as well as the storage and data center switch counts the older edition carried. The dataset now tracks sites, servers and server racks.

Every other site class shrinks

Change in installed servers by data center site class 2024 to 2029 from Gartner: large data centers add 15.9 million servers while enterprise data centers lose 2.16 million, rack rooms 1.46 million, midsize 0.87 million and single server sites 0.64 million
Change in the installed server base by site class, 2024 to 2029. Large data centers add 15.9 million servers while the other four classes lose 5.1 million between them. Sources linked below, analysis by softwarestrategiesblog.com.

Between 2024 and 2029 the installed server base grows by 10.8 million units. Large data centers add 15.9 million, while the other four classes lose 5.1 million between them. Enterprise data centers are the 101 to 500 rack sites that most companies would call their own data center. They lose 2.16 million servers during the period, falling from 23.0 million to 20.8 million. Rack and computer room sites lose 1.46 million, midsize data centers 0.87 million and single-server sites another 0.64 million.

Enterprise data centers multiply while they empty. The number of enterprise sites rises from 6,937 to 7,570, a gain of 633 sites, while the servers inside them fall 9.4%. An average enterprise data center drops from 3,314 servers in 2024 to 2,752 in 2029, a 17% decline. The same pattern holds for midsize sites, which grow from 15,037 to 17,097 while their server base falls 12%.

Large data centers held 42% of installed servers in 2023 by the dataset’s own figures, 44% in 2024 and 48% in 2025. The share crosses 50% in 2026, then reaches 57% in 2029. Measured in racks rather than servers, large sites move from 43% of all server racks in 2024 to 56% in 2029.

China closes in on North America

Large data centers by region 2024 to 2029 from Gartner: China grows from 468 to 980 sites at a 15.9% CAGR, North America from 819 to 1,001, Europe from 387 to 479, with Emerging Asia/Pacific at 15.2% and Middle East and North Africa at 12.0%
Large data center sites by region in 2024 and 2029. China adds 512 large sites, North America adds 182. Sources linked below, analysis by softwarestrategiesblog.com.

North America has the most large data centers in every year of the forecast. It has 819 in 2024 and 1,001 in 2029, a 4.1% CAGR, but its lead narrows every year. China grows from 468 to 980 over the same period, a 15.9% CAGR, and adds 512 large sites against North America’s 182. In 2024 China had 57% as many large data centers as North America. In 2029 it has 98% as many.

Asia/Pacific accounts for two thirds of the revision

Most of the revision landed in one region. The 2023 edition grouped China with the rest of Asia/Pacific outside Japan and projected 563 large data centers for that region in 2027. Its 2025 successor splits the region three ways instead. For 2027 it projects 756 in China, 91 in Emerging Asia/Pacific and 69 in Mature Asia/Pacific, a combined 916. That is 353 more large sites than the prior edition. It accounts for 66% of the 535 sites Gartner added to its 2027 worldwide count. North America’s 2027 figure, by comparison, rose from 845 to 917 over the same two years.

Europe grows from 387 large sites to 479, a 4.3% CAGR. Japan grows from 100 to 132, while Emerging Asia/Pacific more than doubles from 54 to 110 at a 15.2% CAGR. The Middle East and North Africa grows from 25 to 43 at 12.0%. Gartner names the Middle East and Africa the fastest-growing region across all site classes, at a 6.5% CAGR. It credits government initiatives, foreign direct investment and partnerships with hyperscalers. In the dataset, Middle East and North Africa rack and computer room sites more than double, from 15,082 to 33,862. That is the fastest growth of any site class in any region.

On sites above the single-server class, North America still leads, with 165,950 in 2024 against China’s 73,143. China leads on total sites because it has 940,909 single-server sites against North America’s 728,078.

Five counts measure five different things

Data center forecasts compared: Gartner 3.74 million sites and 2,880 large data centers by 2029, Synergy Research 1,360 hyperscale data centers and 67% capacity share by 2031, JLL 200 GW by 2030, IEA 950 TWh by 2030, Cushman & Wakefield 31.7 GW under construction
Five sources count the data center build-out in five different units. Latest and outer-year figures from each. Sources linked below, analysis by softwarestrategiesblog.com.

Gartner counts 2,107 large data centers in 2025, while Synergy Research Group counts 1,360 hyperscale data centers at the end of the same year. The two figures measure different things. Gartner counts any site above 500 racks, whoever runs it, which brings in colocation campuses and the largest enterprise and government sites. Synergy, in contrast, counts facilities operated by the hyperscale cloud and internet companies it tracks. So the 747 sites between the two figures are a rough estimate of how many large data centers are not hyperscale operated.

Synergy’s capacity shares also line up with Gartner’s server shares. Synergy put hyperscale operators at 37% of worldwide data center capacity in mid-2023, 41% in mid-2024 and 48% at the end of 2025, with 67% forecast for 2031. Gartner’s shares for 2023, 2024 and 2025 are 42%, 44% and 48%. The two firms measure different units, one in megawatts of critical load and one in installed servers. Even so, they land within five points of each other in every overlapping year.

Gartner share of servers in large data centers, 42% in 2023 to 57% in 2029, compared with Synergy Research hyperscale share of capacity, 37% in 2023 to 48% in 2025 and 67% by 2031, and enterprise on-premise share falling from 56% in 2018 to 19% by 2031
Gartner’s server share in large data centers and Synergy Research’s hyperscale capacity share, with Synergy’s enterprise on-premise share falling from 56% to 19%. Sources linked below, analysis by softwarestrategiesblog.com.

Enterprise on-premise capacity, in Synergy’s count, has fallen from 56% of the total in 2018 to 32% at the end of 2025. It forecasts 19% by 2031, roughly two points a year. Gartner’s enterprise data center class loses 2.16 million servers over five years, according to the 2025 dataset. Both firms see the enterprise data center shrinking as a share while the buildings themselves remain.

Gigawatts and terawatt-hours tell the same story

JLL’s 2026 Global Data Center Outlook puts worldwide capacity at 103 GW and projects 200 GW by 2030, which works out to a 14% CAGR. Roughly 100 GW of new supply arrives during those five years across colocation, hyperscale and on-premise sites. JLL expects AI workloads to represent half of all data center capacity by 2030, up from about a quarter in 2025. The IEA’s Key Questions on Energy and AI puts data center electricity consumption at 485 TWh in 2025, up 17% in a year. It then projects 950 TWh in 2030, around 3% of global electricity demand. AI-focused data centers grew their electricity use 50% in 2025 and triple it by 2030 in the IEA’s projection.

Cushman & Wakefield’s 2026 Global Data Center Market Comparison reports 31.7 GW under construction worldwide in 2025, up from 12.5 GW in the prior edition. During the same period, planned capacity across the Americas rose from 46.1 GW in 2024 to 191.3 GW at the end of 2025. Synergy’s pipeline of future hyperscale data centers stood at 915 facilities in August 2026, up from almost 800 in April.

Set against those figures, Gartner’s 8.1% CAGR for large sites is the most conservative growth rate in the group. JLL’s capacity doubles in five years, and so does the IEA’s electricity use. Synergy’s hyperscale capacity doubles in three. Gartner’s large site count grows 48% in five years, and its large site server base grows 47%. The difference is the unit rather than the direction. A site count cannot capture a 100 MW campus replacing a 10 MW building on the same ledger line. A server count cannot capture an 8-GPU system replacing a 2-socket x86 box.

How to read the Gartner data center forecast

  • Treat the site count as a floor. Gartner’s 2,880 large sites in 2029 says how many buildings cross the 500-rack line. It says nothing about the power or compute inside them, which is where JLL, the IEA and Synergy see the growth doubling.
  • Watch servers per site, not sites. The 2025 edition cut the expected server count of a 2027 large site by 19% while raising the number of sites 27%. Fewer, denser, more power-hungry systems per site is also the pattern every other source describes.
  • Expect the enterprise class to keep emptying. Gartner has enterprise data centers losing 2.16 million servers, while Synergy has on-premise capacity falling two points a year. New enterprise sites get built, and the workloads still leave.
  • Read the China number with its definition. Gartner’s 980 Chinese large sites in 2029 includes every operator above 500 racks. Synergy’s hyperscale count still puts the United States at well over half of worldwide capacity. Both can be true at the same time.
  • Match the unit before comparing growth rates. An 8.1% CAGR in sites, a 14% CAGR in gigawatts and a doubling of hyperscale capacity in three years describe the same build-out at three different resolutions.
  • Expect the next edition to move again. The 2023 edition undercounted 2027 large sites by at least 535 within two years. So far, nothing in the public pipelines from Synergy or Cushman & Wakefield points to the 2025 edition proving high.

Related comparisons

This post is the third in a series that places published forecasts side by side. AI infrastructure spending forecasts compared (October 22, 2026) covers what the build-out costs. AI revenue gap forecasts compared across Gartner, Goldman, Bain and JPMorgan (October 29, 2026) covers what it has to earn back. This post covers where it gets built. Gartner’s $4.05T AI forecast and Forrester’s Predictions 2027 (November 2, 2026) covers when the spending slows.

Key takeaways

  • Large data centers hold 57% of the world’s servers by 2029. Gartner’s dataset puts sites with more than 500 racks at 44% of installed servers in 2024 and 57% in 2029, when they hold 49.4 million of 86.2 million servers.
  • The large site count reaches 2,880 in 2029. That is up from 1,951 in 2024, an 8.1% compound annual growth rate, while total sites grow 1.6% a year to 3.74 million.
  • Gartner raised its 2027 large data center count by 535 sites in two years. The October 2023 edition had 1,955 large sites for 2027, while the October 2025 edition has 2,490, a 27% increase. The 2027 total server base fell from 82.6 million to 80.0 million at the same time.
  • Every other class of site loses servers. Large data centers add 15.9 million servers between 2024 and 2029, while enterprise data centers lose 2.16 million, rack rooms 1.46 million, midsize sites 0.87 million and single-server sites 0.64 million.
  • China nearly matches North America in large sites by 2029. Gartner’s dataset has China growing from 468 large data centers in 2024 to 980 in 2029, a 15.9% CAGR, while North America grows from 819 to 1,001 and keeps the lead by 21 sites.
  • Hyperscale operators already hold 48% of worldwide capacity. Synergy Research Group counted 1,360 hyperscale-operated data centers at the end of 2025, with a pipeline of 915 more as of August 2026 and a 67% capacity share forecast for 2031.
  • Data center electricity use doubles to 950 TWh in 2030. The IEA projects that figure from 485 TWh in 2025, about 3% of global electricity demand, with AI-focused sites tripling their consumption.

How I built this analysis

The Gartner figures come from the two forecast documents and their accompanying Excel datasets. I used the pivot tables in each dataset to total sites, installed servers and server racks by site class and region, and to calculate shares, differences between editions, compound annual growth rates and servers per site and per rack. The 2023 edition uses seven regions and the 2025 edition uses nine, so the Asia/Pacific comparison combines China, Emerging Asia/Pacific and Mature Asia/Pacific from the 2025 edition to match the older grouping.

Two figures in Gartner’s summary differ from the dataset. The summary says large data centers held 41% of servers in 2024, while the dataset gives 41.5% for 2023 and 44.4% for 2024. I use the dataset values throughout. The summary also gives the Middle East and Africa a 6.5% CAGR, while the dataset’s two sub-regions combined grow at 5.7%. I report Gartner’s stated figure and attribute it, since the two cannot be reconciled from the dataset. Synergy, JLL, IEA and Cushman & Wakefield figures come from the public releases linked below, while the charts are my own work.

More analysis from this blog

This post is my personal analysis of Gartner, Synergy Research Group, JLL, IEA and Cushman & Wakefield research and does not represent my employer.

Sources

Gartner

  • Gartner, Forecast: Data Center Sites, 2023-2029, Worldwide (G00808671), 17 October 2025, including the accompanying dataset.
  • Gartner, Forecast: Data Center Sites, Worldwide, 2021-2027, 2023 Update (G00799095), 31 October 2023, including the accompanying dataset.

Public research and market data

Nov 17

1 Comment

Google Cloud technologies overview, 2026 edition, with a 2027 outlook

Updated 12 November 2026. This post first appeared on 17 November 2010 as a short note on Google App Engine and the Google Labs cloud APIs. It is the most visited page on this site this month. So the text below replaces the original with a Google Cloud technologies overview for 2026 and a look at 2027. The 2010 version is summarized in the first section for anyone who arrived looking for it.

Google Cloud technologies overview 2026 hero graphic: quarterly revenue from $15.2 billion in Q3 2025 to $30.7 billion in Q3 2026, contracted backlog from $155 billion to $561 billion, a table mapping the 2010 Google cloud stack of App Engine, Google Storage, Prediction API, BigQuery and Google Labs to their 2026 names, and key figures of $30.7 billion revenue, $561 billion backlog, 36.2% operating margin and 16% market share
Google Cloud by the numbers in 2026. Quarterly revenue and contracted backlog for the five quarters through Q3 2026, what each piece of the 2010 stack is called today, and the four figures that define the business at the end of 2026. Click any chart to open it full size. Sources linked below, analysis by softwarestrategiesblog.com.

Google Cloud reported $30.7 billion of revenue for the third quarter of 2026, up 103% from $15.2 billion a year earlier. Operating income reached $11.1 billion, a 36.2% margin, compared to $3.6 billion and 23.7% a year earlier. Contracted backlog stood at $561 billion at the end of September, up from $155 billion twelve months before. Alphabet expects to recognize just over 55% of it as revenue within 24 months. The business this post described in 2010 did not disclose revenue at all. In 2026 it is the fastest growing of the three largest clouds and Alphabet’s second largest source of operating income.

This Google Cloud technologies overview covers four things. It starts with what the 2010 stack looked like and what each piece is called today. It then works through the financial record from the first disclosure in 2020 to the Q3 2026 results. The third section compares Google Cloud with AWS and Microsoft Azure using Synergy Research Group’s Q3 2026 market data. After that comes the platform announced at Google Cloud Next in April 2026. The final section sets out what the backlog, capex guidance and competitor commentary already say about 2027.

What the 2010 Google cloud stack became

In November 2010 Google’s cloud effort was a developer evangelism program. The original post pointed readers to the Google Labs catalog and to a slide deck by Google developer advocate Chris Schalk. The deck covered Google Storage for Developers, the Prediction API and BigQuery, all three of them in preview. App Engine, launched in 2008, was the production product at the time. The post contrasted Google’s incremental approach with how often Amazon Web Services changed its platform, pricing and services. Google did not report cloud revenue then, since it had no cloud segment to report.

Almost every element of that stack still exists under a different name. Google Storage for Developers became Cloud Storage. BigQuery left preview in 2011 and now anchors what Google calls the Agentic Data Cloud. The Prediction API was shut down in 2018. Its role passed to Cloud Machine Learning Engine, then to Vertex AI in 2021, and in April 2026 to the Gemini Enterprise Agent Platform. Google Labs closed in 2011, and Google Cloud Next is where new services launch instead. App Engine, in fact, remains available inside Google Cloud. Google itself became a subsidiary of Alphabet in 2015. Alphabet first reported Google Cloud as a segment in February 2020, disclosing $8.9 billion of 2019 revenue. The 2018 figure was $5.8 billion and the 2017 figure $4.1 billion.

Two businesses that did not exist in 2010

The first is security. Google Cloud bought Mandiant in September 2022 for $5.4 billion. It then closed its $32 billion acquisition of Wiz on 11 March 2026, the largest purchase in Alphabet’s history. The second is custom silicon sold to customers rather than used internally. The Tensor Processing Unit did not reach outside developers until 2018. In 2026 Alphabet counts more than 100 significant TPU customers and sells TPU systems for installation in customers’ own data centers.

Google Cloud revenue went from $4 billion a year to $30.7 billion a quarter

Google Cloud annual revenue 2017 to 2026 from Alphabet filings: $4.1 billion in 2017, $8.9 billion in 2019, $13.1 billion in 2020, $26.3 billion in 2022, $43.2 billion in 2024, $58.7 billion in 2025, $75.5 billion in the first nine months of 2026 and a $122.8 billion Q3 2026 run rate, with operating income of $1.7 billion in 2023, $6.1 billion in 2024, $13.9 billion in 2025 and $26.5 billion in the first nine months of 2026
Google Cloud segment revenue by year, from Alphabet’s 10-K and 10-Q filings. The segment lost money every year through 2022, earned $1.7 billion in 2023 and $26.5 billion in the first nine months of 2026. Sources linked below, analysis by softwarestrategiesblog.com.

Google Cloud revenue has grown every year since Alphabet began disclosing it. Segment revenue was $13.1 billion in 2020 and $19.2 billion in 2021. It then reached $26.3 billion in 2022, $33.1 billion in 2023 and $43.2 billion in 2024. Alphabet’s 2025 Form 10-K puts 2025 revenue at $58.7 billion. The first nine months of 2026 added $75.5 billion, more than the whole of 2025. At the Q3 2026 rate the segment is on a $123 billion annual run rate. That is roughly fourteen times the $8.9 billion it reported for 2019, before the segment turned a profit.

Profitability arrived later than scale. The segment reported operating losses in every year through 2022. It then earned $1.7 billion in 2023, $6.1 billion in 2024 and $13.9 billion in 2025. The first nine months of 2026 produced $26.5 billion, almost double the full year before, according to Alphabet’s Q3 2026 Form 10-Q. Operating margin moved from 5.2% in 2023 to 14.1% in 2024, 23.7% in 2025 and 36.2% in the latest quarter.

Growth accelerated for five straight quarters

The acceleration is recent and steep. Google Cloud grew 34% in Q3 2025 and 48% in Q4 2025. It grew 63% in Q1 2026, 82% in Q2 2026 and 103% in Q3 2026. Quarterly revenue went from $15.2 billion to $30.7 billion over those five quarters. Alphabet attributes the growth to Google Cloud Platform, and within it to AI infrastructure and enterprise AI solutions. Core GCP products grew as well, though more slowly. Eighteen Google Cloud product lines each generate more than $1 billion a year. About 90% of the Fortune 100 use Gemini Enterprise, according to the Q3 2026 earnings call.

Backlog grew faster than revenue. Remaining performance obligations are the contracted revenue not yet recognized. They went from $155 billion in Q3 2025 to $240 billion in Q4 2025 and $462 billion in Q1 2026. They then reached $514 billion in Q2 2026 and $561 billion in Q3 2026. At the Q2 call Alphabet said just over 50% of backlog would convert to revenue within 24 months. By the Q3 10-Q that figure was just over 55%. So about $309 billion of contracted revenue sits inside the next two years before any new signings.

Google Cloud gains share while AWS loses it

AWS, Microsoft Azure and Google Cloud compared for Q3 2026: revenue of $46.0 billion, $33.2 billion and $30.7 billion, year over year growth of 39%, 48% and 103%, and Synergy Research market share of 26%, 19% and 16% against 29%, 20% and 13% a year earlier
The three largest cloud providers in Q3 2026. AWS remains the largest, Google Cloud grows more than twice as fast as either rival, and Synergy Research has Google Cloud gaining three points of share in a year while AWS loses three. Sources linked below, analysis by softwarestrategiesblog.com.

Synergy Research Group puts worldwide cloud infrastructure services revenue at $167.7 billion for Q3 2026, up 55% from a year earlier. It was the twelfth consecutive quarter in which the growth rate increased. Trailing twelve-month revenue, as a result, reached $558 billion. Within that market Synergy has Amazon at 26% share, down from 29% a year earlier. Microsoft holds 19% against 20%, and Google 16% against 13%. Oracle, Alibaba, SpaceX and xAI, CoreWeave, Salesforce and Anthropic each hold between 2% and 4%.

The company results line up with the shares. AWS reported $46.0 billion of Q3 2026 revenue, up 39%. Its operating income was $17.8 billion, so its run rate is $184 billion at a 38.7% margin. Microsoft disclosed Azure revenue in dollars for the first time at $33.2 billion, up 48%, a $133 billion run rate. Google Cloud’s $30.7 billion, up 103%, sits $2.5 billion behind Azure and $15.3 billion behind AWS for the quarter. A year earlier the gaps were larger on every measure. Google Cloud was growing 34% then, while AWS grew 20% and Azure 40%.

Google Cloud’s margin now sits between its two rivals. AWS earned a 38.7% operating margin in Q3 2026 while Google Cloud earned 36.2%. Microsoft, by comparison, does not disclose Azure operating income. For comparison, Synergy’s Q3 2022 shares were 34% for Amazon, 21% for Microsoft and 11% for Google. Google has gained five points of share in four years while Amazon has lost eight.

The 2026 platform is built around agents and TPUs

Google Cloud Next 2026 ran from 22 to 24 April in Las Vegas. It drew more than 32,000 attendees and produced 260 announcements. The center of them was the Gemini Enterprise Agent Platform, the renamed and extended Vertex AI. It adds Agent Studio for building agents and an Agent Registry for cataloging them. Agent Identity gives each agent its own credentials, Agent Gateway enforces policy on every call, and Agent Observability traces what agents do. The point of the five components is that an enterprise can build, authorize, route and audit agents in one place rather than stitching controls together from separate products.

Infrastructure announcements followed the same theme. TPU 8t is the eighth generation training chip. It scales to 9,600 chips and 2 petabytes of high bandwidth memory in a single superpod. Google rates it at three times the compute of Ironwood, twice the performance per watt and 2.7 times the training performance per dollar. TPU 8i is the inference part, with 1,152 chips per pod and an 80% improvement in inference performance per dollar compared to the prior generation. Google also described Virgo, a custom network for connecting TPU supercomputers into larger clusters. The Agentic Data Cloud rounds it out. Its Cross-Cloud Lakehouse is standardized on Apache Iceberg, so BigQuery can query data where it sits, including data left in AWS.

The money behind the platform is unusually visible. Alphabet spent $51.4 billion on capital expenditure in Q3 2026. It held its 2026 guidance at $195 billion to $205 billion. Total Alphabet revenue reached $127.6 billion in the quarter, up 25%, with $40.6 billion of operating income. The Gemini app also passed one billion monthly users during the quarter.

Security is Google Cloud’s $37 billion acquisition story

Google Cloud entered security by acquisition. Mandiant, bought for $5.4 billion in 2022, brought incident response and threat intelligence. Wiz was announced in March 2025 for $32 billion and closed on 11 March 2026 for $29.5 billion after adjustments. It brought a multicloud security platform that Google has kept as a separate brand since the deal closed. Wiz still works across AWS, Azure and Oracle as well as Google Cloud. The purchase added $22.7 billion of goodwill to the Google Cloud segment. That is what Alphabet paid for the customer base, the engineers and the market position rather than the code.

The combined products now ship under Google AI Threat Defense. It pairs Gemini models with Wiz, Mandiant intelligence and the CodeMender code repair agent. Wiz, for its part, introduced Red, Blue and Green agents for offensive testing, defensive response and remediation. Google’s security blog describes the goal as an agentic security operations center. In it, agents triage alerts, investigate and recommend responses while analysts approve the steps. For readers of this site, that matters for one reason. The controls Google ships for its own agents, Agent Identity and Agent Gateway in particular, are the same controls enterprises need for agents built on any platform.

What 2027 already looks like

Three numbers already shape 2027, starting with the backlog. Just over 55% of $561 billion converts within 24 months, so roughly $309 billion of revenue is contracted before Google Cloud sells anything new. Spread evenly, that alone is about $155 billion a year, above the current $123 billion run rate. The second is capital spending, which is still rising. Alphabet told investors on the Q3 2026 call to expect capex to increase significantly in 2027. The 2026 range is $195 billion to $205 billion, and the Cloud segment carries much of the depreciation. Operating margin rose from 23.7% to 36.2% in a year while that spending ramped. The question for 2027 is whether margin keeps rising as the depreciation catches up.

The third is the competition’s own guidance. Amazon CEO Andy Jassy told analysts he expects AWS to grow more than 40% in 2027 and 2028. To keep gaining share, Google Cloud has to grow faster than that from a base now two thirds the size of AWS. Synergy’s data has the entire market compounding at over 50%. The test of 2027 is whether the three leaders can all grow above 40% at the same time. That has never happened at this scale.

Two other threads run into 2027. TPU systems are now sold into customers’ own data centers, with Oracle and Crux AI named as early buyers. That turns Google’s silicon into a product line rather than an internal advantage. Wiz plus Mandiant gives Google Cloud a security business that competes with CrowdStrike and Palo Alto Networks as much as with AWS. Neither chips for sale nor security acquisitions were part of the plan when App Engine was the product. The 2010 version of this post did not see either one coming.

Key takeaways

  • Google Cloud doubled in a year. Q3 2026 revenue of $30.7 billion was up 103% from $15.2 billion. Growth also accelerated in each of the last five quarters.
  • The backlog is the forecast. Contracted backlog reached $561 billion, with just over 55% converting within 24 months. That puts about $309 billion of revenue under contract for the next two years.
  • Margins moved from 5% to 36% in three years. Operating margin went from 5.2% in 2023 to 36.2% in Q3 2026, on $11.1 billion of quarterly operating income.
  • Google Cloud is the share gainer. Synergy Research puts it at 16% of a $168 billion quarterly market, up from 13%. AWS, in contrast, fell from 29% to 26%.
  • The 2010 stack survived under new names. Google Storage is Cloud Storage and BigQuery anchors the Agentic Data Cloud. The Prediction API’s role, in turn, now belongs to the Gemini Enterprise Agent Platform.
  • Security is a $37 billion acquisition story. Mandiant at $5.4 billion in 2022 and Wiz at $32 billion in 2026 gave Google Cloud a security business that sells across every major cloud.
  • 2027 is a capex and margin test. Alphabet expects capex to rise significantly from the 2026 range of $195 billion to $205 billion. AWS guides to more than 40% growth, so share gains require Google Cloud to grow faster than that.

Sources

This analysis is my own and reflects public filings, earnings calls and vendor announcements. It does not represent the views of my employer.