Glossary · from SEC filings
What is a hyperscaler?
A hyperscaler is a company that builds and runs data centers at enormous scale, usually to rent computing to others through a public cloud. In AI the word has come to mean the handful of companies whose capital spending pays for most of the build-out: Microsoft, Amazon, Alphabet, Meta and Oracle.
What “hyperscaler” means
The term comes from computing architecture: a hyperscale system is one that can scale out by adding more machines rather than bigger ones. In business it describes the companies that operate data centers on that model across the world, measured in hundreds of facilities and gigawatts of power. Amazon Web Services, Microsoft Azure and Google Cloud are the original hyperscalers; Oracle has become one more recently by building very large AI clusters, and Meta builds at hyperscale for its own apps even though it doesn’t sell cloud services.
Suppliers use the word in their filings to describe their most important customers. Micron’s 10-K describes its Cloud Memory Business Unit as focused on “large hyperscale cloud customers”, and Nvidia now reports its data-center revenue split between hyperscalers and everyone else. When an AI supplier talks about demand, it is mostly talking about these companies.
Micron: Micron’s cloud memory unit
“CMBU is focused on memory solutions for large hyperscale cloud customers, and HBM for all data center customers.”
Micron 10-K, filed October 9, 2026
How much the hyperscalers spend
Over their latest four quarters, the five largest buyers spent $566B on capital projects, +81% on the four quarters before and 3.1 times what they spent two years earlier. Individually: Microsoft $136B, Amazon $173B, Alphabet $160B, Meta $104B, Oracle $75.7B. Those figures are cash capex only; several also sign large leases and finance leases that come on top.
The spending now absorbs 97% of their combined operating cash flow, up from about a third in 2023. That is why hyperscaler capex is the number the whole AI supply chain watches: chips, memory, servers, networking and power equipment are all, directly or indirectly, paid for from it. Our capex tracker follows it quarter by quarter.
How they earn it back
A hyperscaler earns its return by renting out capacity, through cloud services or AI models, or by using it in its own products. The cloud segments are where the filings show it: AWS $148B of revenue over the latest four quarters (+28%); Microsoft Cloud $229B of revenue over the latest four quarters (+28%); Google Cloud $93.2B of revenue over the latest four quarters (+76%); Oracle cloud infrastructure $22.1B of revenue over the latest four quarters (+94%). Meta has no cloud segment, so its return has to show up in advertising.
The gap between spending and revenue is the central question in AI investing. Cloud revenue is growing fast, but capex is growing faster, and the depreciation on all that equipment will keep rising for years. Whether the spending pays off depends on cloud revenue growing into it before the equipment wears out. Our AI bubble indicators track that gap.
Why the word shows up in suppliers’ results
Suppliers now report hyperscalers as a category of their own. Nvidia, the largest seller in the stack, splits its data-center revenue between hyperscalers and everyone else: in its quarter to 2026/07, hyperscale revenue was $48.7B, up from $24.2B a year earlier. Micron runs a business unit for large hyperscale cloud customers, and Western Digital reports that its Cloud end market, essentially the hyperscalers, accounted for 89% of its revenue.
For investors in those suppliers, the hyperscalers’ capex guidance is effectively their own revenue guidance, a quarter or two ahead. That is why a single sentence about capex on a hyperscaler’s earnings call can move the shares of companies several layers down the stack.
The five, one by one
Microsoft: $136B of capex against $196B of operating cash flow over its latest four quarters, leaving free cash flow of $60.0B, with $683B of contracted revenue not yet earned and $329B of data-center leases signed but not yet started. Company page.
Amazon: $173B of capex against $161B of operating cash flow over its latest four quarters, leaving free cash flow of −$11.6B. Company page.
Alphabet: $160B of capex against $180B of operating cash flow over its latest four quarters, leaving free cash flow of $19.7B, with $568B of contracted revenue not yet earned and $131B of data-center leases signed but not yet started. Company page.
Meta: $104B of capex against $133B of operating cash flow over its latest four quarters, leaving free cash flow of $28.6B and $356B of data-center leases signed but not yet started. Company page.
Oracle: $75.7B of capex against $46.9B of operating cash flow over its latest four quarters, leaving free cash flow of −$28.7B, with $664B of contracted revenue not yet earned and $260B of data-center leases signed but not yet started. Company page.
How long they say their servers last
Every hyperscaler chooses how many years to depreciate its servers over, and the choice moves reported profit by billions of dollars a year. The latest stated lives in their filings: Microsoft 2–6 years, Amazon 5–6 years, Alphabet 6 years, Meta 5–5.5 years, Oracle 1–6 years. Most lengthened those lives between 2022 and 2025, which lowered depreciation and raised profits; Amazon has since shortened its server life from six years to five.
The question for investors is whether AI hardware lasts as long as the older servers those lives were set for. Nvidia now releases a new generation of accelerators roughly every year. If the newer chips make older ones uneconomic sooner, the hyperscalers will either shorten their lives again, raising depreciation, or record write-downs. The useful-life calculator shows the effect for each company.
Custom chips: the hyperscalers’ answer to Nvidia
Each hyperscaler that sells cloud services also designs its own AI chips: Google its TPUs, Amazon its Trainium and Inferentia, and Microsoft and Meta chips of their own, usually with a design partner such as Broadcom or Marvell. Amazon’s filings describe customer commitments that include obligations tied to the performance of AWS’s own chips, a sign of how central these designs have become.
Custom chips matter to investors on both sides. For the hyperscalers they are a way to lower the cost of AI computing and reduce dependence on one supplier. For Nvidia they are the main long-term competitive threat, and for Broadcom and Marvell they are a large and growing business. See the chips layer.
Hyperscalers, neoclouds and colocation
Hyperscalers are not the only places AI computing is rented. Neoclouds such as CoreWeave rent GPU capacity only, often to the hyperscalers themselves and to AI labs, and depend on borrowed capital in a way the hyperscalers don’t. Colocation companies such as Equinix and Digital Realty rent the building, power and cooling, and let customers bring their own servers. See neocloud for how the GPU clouds differ.
The hyperscalers sit at the top of the money flow in the AI stack: they pay the GPU clouds, the server makers, the chip designers and the power companies, and they are paid by software companies and end customers. That position makes them both the biggest customers and the biggest risk in the system: if they slow spending, every layer below them feels it within a quarter or two.
What to watch
Three lines in each hyperscaler’s quarterly filing tell you most of what matters: capex and any change in capex guidance, the growth and operating margin of its cloud segment, and leases signed but not yet started. Microsoft, Meta, Oracle and Alphabet together disclose hundreds of billions of dollars of such leases, commitments that don’t appear in capex until the leases begin.
It also helps to read what each hyperscaler says about next year. Capex guidance is not in the 10-Q itself but in the earnings release and call, and changes to it move the whole stack. When a hyperscaler raises its guidance, the effect runs down to Nvidia, Broadcom, Micron, Dell, Vertiv and the power companies within a quarter or two; when it cuts, the same chain runs in reverse. The filings then confirm, a quarter later, whether the spending happened as guided.
The company pages for Microsoft, Amazon, Alphabet, Meta and Oracle set out each one’s figures and what has to go right.
Related terms
More from the glossary: Neocloud, High-bandwidth memory (HBM), Remaining performance obligations (RPO), AI data center, TPU vs GPU, CoWoS (advanced packaging), Vendor financing, and every other term you need to read AI companies’ filings.
Free · quarterly · from SEC filings
Get the AI Stack Report every quarter
After each Big Tech earnings season we update the map from the new filings and publish what changed: who spent, who got paid, who kept it and what the filings reveal about the loop. Free, in Rob May’s Investing in AI newsletter.
Form not showing? Subscribe on Substack. No spam; unsubscribe any time.
Questions about hyperscalers
What is a hyperscaler?
A company that builds and runs data centers at global scale, usually to rent computing through a public cloud. In AI the term usually means Microsoft, Amazon, Alphabet, Meta and Oracle, whose capital spending pays for most of the AI build-out.
Who are the hyperscalers?
Amazon (AWS), Microsoft (Azure) and Alphabet (Google Cloud) are the original three. Oracle has become one by building large AI clusters, and Meta builds at hyperscale for its own apps although it doesn’t sell cloud services.
Is Nvidia a hyperscaler?
No. Nvidia sells chips and systems to the hyperscalers; in its filings, hyperscalers are its largest group of customers. It does own some cloud capacity, mostly bought back from AI cloud partners for its own research.
How much do hyperscalers spend on AI?
The five largest spent more than half a trillion dollars on capital projects over their latest four quarters, nearly all of their operating cash flow. Not all of it is AI, but AI data centers account for most of the increase.
What is the difference between a hyperscaler and a neocloud?
A neocloud rents GPU capacity only, is usually much smaller and funds itself with borrowed money and long customer contracts. Hyperscalers sell a full range of cloud services and fund their spending largely from their own cash flow.
Is Meta a hyperscaler?
It builds and runs data centers at hyperscale and is one of the five largest spenders on AI infrastructure, but it doesn’t rent capacity to others through a cloud business. Its return has to come through its own apps and advertising.
Do hyperscalers make their own chips?
Yes. Google designs TPUs, Amazon designs Trainium and Inferentia, and Microsoft and Meta design chips of their own, usually with partners such as Broadcom or Marvell. They still buy large numbers of Nvidia GPUs.
Why do hyperscalers matter to AI investors?
Because their capital spending is most of the demand for AI chips, memory, servers, networking and power equipment. Their capex plans set the revenue outlook for much of the AI supply chain, and their cloud revenue shows whether that spending is earning a return.
What is a hyperscale data center?
A very large data center, typically tens to hundreds of megawatts, designed to add capacity by adding standardized servers. AI data centers are pushing individual campuses to gigawatt scale.
Research and education, not investment advice: nothing here is a recommendation to buy or sell any security. Figures come from companies’ SEC filings as reported; companies named were chosen because they show how the stack works, not as a view on their shares.