Who Are the Hyperscalers? The Companies Paying for the AI Build

EXPLAINED · Players · Last reviewed: 22 Oct 2026 · Next review: after the third-quarter 2026 results from Alphabet, Microsoft and Amazon (28 to 29 Oct 2026, US time)

The hyperscalers are the few companies that build and run data-centre computing at the largest scale: Amazon, Microsoft, Google, Meta and Oracle, with Alibaba leading in China.

Why an investor cares: almost every dollar in the AI build starts as a hyperscaler’s capital budget. When these few companies raise or cut spending, chip makers, memory suppliers, data-centre landlords and utilities feel it a quarter or more later. If the word itself is new, start with what a hyperscaler is; this page is about who they are and what role each plays.

Who are the hyperscalers?

The usual list in 2026 is six companies: Amazon, Microsoft, Alphabet’s Google, Meta and Oracle in the United States, and Alibaba in China. What joins them is scale. Each builds its own data centres, designs or buys servers by the hundreds of thousands, and spends tens of billions of dollars a quarter doing it.

There is no official list. The label comes from industry use, not a regulator, and it moves. Oracle was a software company with a small cloud a few years ago; it now delivers more than 300,000 GPUs in a quarter. Other large Chinese platforms, such as Tencent, Huawei and ByteDance, also run very large data-centre fleets, but report less about them.

The six do not all play the same part. Sort them by how their spending gets paid back and four roles appear: the rental clouds, the self-builder, the contract cloud, and the China cloud.

Latest reported quarter for each company, in their own words and definitions. Second Order table from company results.

Who are the big three hyperscalers?

Amazon Web Services, Microsoft Azure and Google Cloud. They rent computing, storage and software to millions of companies, so every server they buy has to be rented out to pay for itself. That makes their cloud revenue and their signed book the numbers to watch.

Amazon is the largest by cloud sales. AWS sold $42.2 billion of services in the April to June 2026 quarter, up 37%, which Amazon calls a $169 billion annual run rate. The bill for that growth shows in cash: Amazon bought $54.2 billion of property and equipment in the quarter, and its free cash flow over the past 12 months turned negative, an outflow of $7.6 billion.

Microsoft reports a $678 billion commercial remaining performance obligation, up 84% in a year: contracts signed but not yet delivered. Its additions to property and equipment were $35.8 billion in the quarter, about double a year earlier (our arithmetic), and $115.9 billion over its fiscal year to June. Our RPO explainer shows how that signed book turns into revenue.

Google Cloud is the fastest-growing of the three: revenue of $24.8 billion, up 82%, with a cloud backlog of $514 billion. Alphabet spent $44.9 billion in the quarter, about 60% of it on servers, and lifted its 2026 guide to $195–205 billion. Google also plays a second role. It designs its own AI chip, the TPU, and started delivering TPU systems into customers’ own data centres in the second quarter, so it is now a chip seller as well as a buyer.

Is Meta a hyperscaler? The one that does not rent out capacity

Meta. It builds data centres only for its own apps and AI models. It has no outside cloud customers and no cloud backlog to report, so its payback runs through advertising.

That makes Meta the purest test of AI spending. In the July to September quarter it spent $34.9 billion, including principal payments on finance leases, and it now expects $135–145 billion for 2026, a narrower range than the $130–145 billion it gave in July. Against that, ad impressions rose 10% and the average price per ad rose 14%, while free cash flow was negative $2.1 billion for the quarter. Our read: Meta has to show the return inside its own ad numbers, because there is no rental line to point to.

Why is Oracle counted as a hyperscaler now?

Because of its AI contracts. Oracle Cloud Infrastructure grew 121% to $7.4 billion in the quarter to August 2026, and the company booked more than $30 billion of new AI cloud contracts in that quarter, lifting its remaining performance obligations to $664 billion.

Oracle is also the clearest case of building ahead of the cash. Its capital expenditures were $28.5 billion in the quarter, more than three times the $8.5 billion a year earlier (our arithmetic), and its free cash flow was negative $5 billion. To fund the build it sold $20 billion of shares through an at-the-market programme. It delivered 850 MW of extra data-centre capacity and more than 300,000 GPUs in the quarter.

Our read: Oracle sits between a hyperscaler and a neocloud. It has a broad cloud and a large software business like the first, but its AI growth rests on a few very large contracts, more like the second. Neocloud vs hyperscaler sets out the difference.

Who is the hyperscaler in China?

Alibaba Cloud is the largest. Alibaba’s AI Cloud and Compute Services revenue was RMB48.4 billion, about US$7.1 billion, in the April to June 2026 quarter, up 45%. Alibaba cites the research firm Omdia ranking it first in China’s AI cloud market for 2025 with a 38.1% share.

Its spending rose faster than its revenue: capital expenditures of RMB67.7 billion, about US$10.0 billion, up 75%, which Alibaba put down partly to higher prices for chip components. It also designs its own chips through its T-Head unit. US export rules limit which Nvidia chips Chinese clouds can buy, which is one reason Chinese hyperscalers lean on their own silicon; our China door tracks that side.

Add up the four largest US names for April to June 2026 and the total is about $166 billion in one quarter (our sum: $54.2 billion, $44.9 billion, $35.8 billion and $31.1 billion). Each company counts a slightly different line: Amazon’s is gross purchases, Microsoft’s leaves out finance leases, Meta’s includes their principal. Treat the total as a scale, not a precise figure.

Who gets paid when the hyperscalers spend?

The suppliers, first, and long before the hyperscalers bill. Chip makers are paid when servers are bought; who buys AI GPUs shows how much of Nvidia’s revenue the hyperscalers account for, and who makes custom AI chips covers the design partners behind their in-house chips.

Landlords are paid next. Hyperscalers build many of their own campuses, but they also lease halls from data-centre landlords, often signing years before the servers move in; how a data-centre lease works explains the rent, and hyperscale vs colocation the two kinds of building. In Asia the same buyers lease capacity in Singapore, where new data-centre capacity is rationed, and across the causeway in Johor.

Utilities and builders are paid to deliver power and finished halls. Our Wire on Amazon’s €4.8 billion Frankfurt plan is one example from this week of where the money lands on the ground.

Where is the stuck step?

At the powered hall. The hyperscalers have the money and, on their own numbers, more signed demand than they can serve. What they cannot buy quickly is a finished building with grid power, which is why how hyperscalers buy power now matters as much as how they buy chips.

Alphabet said it again on its July call: it is still in a supply-constrained environment, and it planned to expand its use of third-party capacity in the third quarter as a bridge while it builds its own. In plain words, the largest buyers are renting from others because their own halls are not ready. That is a large part of why neoclouds and landlords exist at today’s scale.

The signed books make the same point. Microsoft’s $678 billion, Google Cloud’s $514 billion and Oracle’s $664 billion add up to about $1.86 trillion of contracted work (our sum; each company defines its book differently). That work only turns into revenue as capacity switches on. How long it takes to build an AI data center shows why that clock runs in years, not quarters. The layer view is on our Hyperscalers stack and AI Data Centers stack.

Who is not a hyperscaler?

Several kinds of company get the label by mistake. They are better seen as suppliers or customers of the hyperscalers.

Our classification. Roles only, not a ranking.

The neoclouds are the closest cousins. They rent GPU hours, often to the hyperscalers themselves, and depend on a handful of large contracts; GPU cloud providers maps who does what, and the Neoclouds stack tracks their funding.

What would prove this wrong?

• Spending guides get cut. If the third-quarter results lower 2026 or 2027 spending plans, the buyers have more capacity than demand and the wait is no longer power.
• Backlogs stall. If RPO and cloud backlog stop growing while capacity comes online, the signed book was the limit, not the halls.
• Third-party renting stops. If Alphabet and others stop leasing outside capacity, their own halls have caught up with demand.

What to check next

Meta reported third-quarter results on 21 October 2026 (US time): $34.9 billion of spending in the quarter and a 2026 guide narrowed to $135–145 billion. Alphabet and Microsoft report on 28 October and Amazon on 29 October. In each, check three lines: the full-year spending guide, the signed book, and any comment on capacity being short. Hyperscaler capex lines to watch lists where each company puts them, and AI capex explains what the money buys.

Common misconceptions

• People assume there is an official list of hyperscalers. Actually it is an industry label based on scale, and it changes; Oracle joined it through AI contracts.
• People assume all hyperscalers rent out capacity. Actually Meta uses everything it builds for itself and is paid back through ads.
• People assume hyperscalers own every hall they use. Actually they also lease from landlords and rent from neoclouds; Alphabet said it would expand third-party capacity as a bridge.
• People assume a bigger backlog means revenue soon. Actually it is billed as capacity switches on, which waits on halls and power.

FAQ

Related: hyperscaler; neocloud vs hyperscaler; who buys AI GPUs; remaining performance obligation; AI capex; sovereign AI; AI data center.

Stack doors

Layer: Hyperscalers. Also: Neoclouds, Semiconductors, AI Data Centers. Markets: US, Singapore, Malaysia, China.

Get the Wire — short daily notes when hyperscaler spending, backlogs and capacity move.

Not investment advice. Do your own research.

References