The AI stack map · from SEC filings
AI infrastructure stocks: who gets paid when AI gets built
In the four quarters to Q2 2026, Microsoft, Amazon, Alphabet, Meta and Oracle spent $566B on property and equipment, most of it on data centers for AI. This map of AI infrastructure stocks follows that money down the stack, layer by layer, using only what the companies report to the SEC. It also asks the question that decides whether any of it was a good investment: is the money coming back up?
- capital spending by the five largest buyers in the latest four quarters ($184B two years earlier)
- $566B
- capital spending by the five largest buyers in the latest four quarters ($184B two years earlier)
- of their operating cash flow went to capex in Q2 2026 (low: 35% in Q3 2023)
- 97%
- of their operating cash flow went to capex in Q2 2026 (low: 35% in Q3 2023)
- of the extra gross profit suppliers earned over two years went to chips and memory
- 83%
- of the extra gross profit suppliers earned over two years went to chips and memory
- quarterly revenue at the four big clouds in Q2 2026, +40% in a year
- $132B
- quarterly revenue at the four big clouds in Q2 2026, +40% in a year
The map
Read the map from the top. Customers pay for software and cloud services. The clouds turn that revenue, and a great deal of their own cash, into capital spending. That spending becomes revenue for the companies that build the systems, wire them together and power them. Those companies buy chips and memory, and the chipmakers in turn pay foundries, which buy tools. Every number on the map is the layer’s own: total revenue over the latest four reported quarters, its change from two years earlier, and the share of revenue kept as gross profit.
Two things don’t fit a top-down picture, and they are the reason this map exists. First, the money has to come back. Capital spending is a cost, not a return; it pays off only if the revenue on the right-hand side of the map keeps rising. Second, some of the money no longer flows one way. Suppliers have started investing in, lending to and guaranteeing their own customers. We trace both below, and each has its own page.
Layer 8
Software and applications
- Revenue
- $193B
- 2-yr change
- +37%
- Gross margin
- 78%
Layer 1
AI clouds
- Revenue
- $1.91T
- 2-yr change
- +35%
- Gross margin
- 61%
Layer 2
Servers and systems
- Revenue
- $297B
- 2-yr change
- +56%
- Gross margin
- 16%
Layer 5
Networking and optics
- Revenue
- $141B
- 2-yr change
- +45%
- Gross margin
- 53%
Layer 6
Power, cooling & data centers
- Revenue
- $229B
- 2-yr change
- +37%
- Gross margin
- 27%
Layer 3
Chips
- Revenue
- $544B
- 2-yr change
- +106%
- Gross margin
- 66%
Layer 4
Memory and storage
- Revenue
- $183B
- 2-yr change
- +289%
- Gross margin
- 75%
Layer 7
Foundry, equipment and design tools
- Revenue
- $98.9B
- 2-yr change
- +37%
- Gross margin
- 52%
Five things the filings say
1. The spending now uses almost all of the buyers’ cash
In Q3 2023 the five biggest buyers spent 35% of their operating cash flow on capital projects. In Q2 2026 they spent 97%: $182B of capex against $186B of cash from operations, in a single quarter. Four-quarter capex has risen from $184B to $566B in two years.
That changes who is funding the build-out. When a company spends a third of its cash flow, the spending is a choice made with spare money. When it spends nearly all of it, AI competes directly with buybacks and dividends, and at the margin it is paid for with debt. Oracle is already there: $75.7B of capex against $46.9B of operating cash flow over its latest four quarters, so its free cash flow is −$28.7B. CoreWeave spent $20.6B on equipment while booking $7.6B of revenue. Follow the money in detail.
The five biggest buyers now spend almost all their operating cash on capex
Microsoft, Amazon, Alphabet, Meta and Oracle combined, per calendar quarter
- Operating cash flow
- Capital spending
- Depreciation and amortization
Source: company 10-K and 10-Q filings (SEC XBRL). Capex is cash paid for property and equipment; finance leases excluded. Fiscal quarters aligned to the nearest calendar quarter.
Show the numbers
| Quarter | Capex | Operating cash flow | Capex ÷ cash flow | D&A |
|---|---|---|---|---|
| Q1 2021 | $27.8B | $61.6B | 45% | $9.9B |
| Q2 2021 | $31.6B | $75.4B | 42% | $10.4B |
| Q3 2021 | $33.8B | $76.9B | 44% | $11.4B |
| Q4 2021 | $37.5B | $75.9B | 49% | $12.4B |
| Q1 2022 | $36.6B | $65.6B | 56% | $11.8B |
| Q2 2022 | $38.4B | $69.2B | 56% | $12.3B |
| Q3 2022 | $41.0B | $74.0B | 55% | $13.0B |
| Q4 2022 | $41.7B | $79.3B | 53% | $15.7B |
| Q1 2023 | $36.6B | $71.0B | 51% | $17.0B |
| Q2 2023 | $35.3B | $96.9B | 36% | $17.8B |
| Q3 2023 | $38.3B | $109.8B | 35% | $18.9B |
| Q4 2023 | $44.0B | $99.8B | 44% | $21.1B |
| Q1 2024 | $46.0B | $104.5B | 44% | $19.3B |
| Q2 2024 | $55.6B | $114.6B | 49% | $20.2B |
| Q3 2024 | $61.2B | $123.0B | 50% | $27.0B |
| Q4 2024 | $76.3B | $136.3B | 56% | $30.4B |
| Q1 2025 | $77.8B | $120.2B | 65% | $29.5B |
| Q2 2025 | $97.3B | $134.6B | 72% | $32.0B |
| Q3 2025 | $105.8B | $167.1B | 63% | $35.8B |
| Q4 2025 | $130.7B | $180.9B | 72% | $40.5B |
| Q1 2026 | $148.4B | $157.9B | 94% | $42.6B |
| Q2 2026 | $181.5B | $186.4B | 97% | $46.2B |
2. The profit pools at two toll booths
Over the past two years the supplier layers we track added $627B of revenue. Of the extra gross profit that came with it, 83% went to just two layers: chips (mostly Nvidia) and memory (mostly Micron). Memory is the bigger surprise. Its gross margin, which swings with the commodity cycle, went from 25% to 75% as high-bandwidth memory became as scarce as the GPUs it sits beside.
Toll booths are where pricing power lives, and pricing power is what an investor is really buying. But toll booths also attract the most effort to get around them: every hyperscaler now designs its own accelerator, and every memory maker is adding capacity. The chips and memory pages look at how durable each booth is.
3. Volume is not value
The server makers and contract manufacturers added $93.9B of revenue over the same two years and kept $12.8B of it as gross profit: 14% of each new dollar. The rest passed straight through to Nvidia, AMD and the memory makers. Supermicro reported $39.1B of revenue in its latest four quarters and −$6.8B of operating cash flow, because the inventory it buys to fill orders ties up more cash than it earns.
If you buy “AI exposure” through the companies that assemble the hardware, you are mostly buying revenue, not margin. That can still work (Dell’s infrastructure segment has roughly doubled its operating income), but the stock is a bet on volume and execution, not on owning a scarce input. Servers and systems explains why.
4. Some of the money now moves in circles
Nvidia reports equity investments of $99 billion in its ecosystem, with $25 billion more committed, and it has agreed to buy back up to $36 billion of cloud capacity from AI clouds that buy its chips. Amazon has a credit line for Anthropic that opens as AWS delivers compute. Microsoft booked $24.1 billion of revenue from OpenAI in its latest fiscal year, after funding $11.9 billion of its investment in the company. AMD gave OpenAI and Meta warrants for up to 160 million shares each at a penny.
None of this is hidden; every item is in a filing, and we quote each one. Vendor financing is an old tool and often a sensible one. But it means some demand at the top of the stack is paid for by the suppliers below it, so reported growth overstates what end customers alone are paying. See every link, quoted.
5. The return is growing, but the bill is growing faster
Revenue at AWS, Microsoft Cloud, Google Cloud and Oracle’s cloud infrastructure reached $132B in Q2 2026, up +40% from a year earlier. That is real money, and the best evidence that AI is starting to pay: Google Cloud’s operating income over the latest four quarters was $31.8B, +198% on the year before. But over the same year the five buyers’ quarterly capex rose +87%.
Capex also arrives on the income statement slowly, as depreciation, over the five or six years the companies say their servers last. In Q2 2026 the five buyers spent 3.9 times what they recorded as depreciation and amortization. Today’s spending is tomorrow’s expense, and margins will have to absorb it. The software page asks whether the customers who are meant to pay for all of this are paying yet.
The eight layers
Each layer page has the companies’ reported numbers side by side, the charts that matter for that layer, what the filings say about its customers and suppliers, and what to watch in the next round of reports.
Layer 1
AI clouds: the buyers
Microsoft, Amazon, Alphabet, Meta and Oracle, plus the GPU-only clouds: the companies whose capital spending is everyone else’s revenue.
Can the biggest buyers keep paying for this out of their own cash?
$1.91T revenue · +35% in two years · 61% gross margin
Layer 2
Servers and systems
Dell, Supermicro, HPE and the contract manufacturers who turn chips into racks, carrying huge revenue on thin margins.
Why does the layer that ships the most hardware keep so little of the money?
$297B revenue · +56% in two years · 16% gross margin
Layer 3
Chips: accelerators and custom silicon
Nvidia, Broadcom, AMD, Marvell and the CPU makers: the toll booth where most of the build-out’s profit has collected so far.
How long can one layer keep most of the profit?
$544B revenue · +106% in two years · 66% gross margin
Layer 4
Memory and storage
Micron, Sandisk, Western Digital and Seagate: high-bandwidth memory turned the industry’s most cyclical layer into the build-out’s second toll booth.
Has memory stopped being a commodity, or is this the top of a cycle?
$183B revenue · +289% in two years · 75% gross margin
Layer 5
Networking and optics
Arista, Cisco, Ciena, Coherent, Lumentum, Credo, Astera Labs, Amphenol and Corning: the switches, optics and cables that make thousands of GPUs act as one.
Which links in the network have pricing power, and which are just volume?
$141B revenue · +45% in two years · 53% gross margin
Layer 6
Power, cooling and data centers
Vertiv, Eaton, GE Vernova, Quanta, the nuclear generators and the data-center landlords: the slow, physical layer that sets how fast everything else can grow.
Is power the real bottleneck, and who gets paid for relieving it?
$229B revenue · +37% in two years · 27% gross margin
Layer 7
Foundry, equipment and design tools
TSMC, ASML, Applied Materials, Lam Research, KLA, Teradyne, Amkor, Synopsys and Cadence: the companies every chip has to pass through.
What does the layer under the chips tell us about how long the build-out lasts?
$98.9B revenue · +37% in two years · 52% gross margin
Layer 8
Software and applications: the return
Salesforce, ServiceNow, Palantir, Adobe and the rest: the layer whose customers have to pay for AI if the spending below is to earn a return.
Is the money coming back up the stack yet?
$193B revenue · +37% in two years · 78% gross margin
Deep dive
Is AI a bubble?
Eight warning signs computed from the filings, each compared with a year earlier: cash, leases, seller financing, backlog and more.
Deep dive
Following the AI capex money
Where the buyers’ cash goes, who keeps each new dollar, the depreciation bill, and whether cloud revenue is catching up.
Deep dive
The AI money loop
Investments, guarantees, warrants and buy-back deals between suppliers and their customers, each quoted from the filing.
Deep dive
Nvidia’s customers and suppliers
Who buys Nvidia’s chips, who makes them, and how much it has committed to both.
Deep dive
Who owns and funds OpenAI
Microsoft’s stake, Amazon’s $50 billion, Nvidia’s guarantees and AMD’s warrant, from the filings.
Deep dive
Who owns and funds Anthropic
Amazon’s notes, preferred stock and a credit line that opens as AWS delivers compute.
Tools
Tool
Company pages
Nvidia, Microsoft, TSMC and 50 more, each read from its own filings.
Tool
Glossary
Hyperscalers, neoclouds, HBM, RPO and the other terms in AI companies’ filings, in plain English.
Tool
Quarterly AI Stack Report
After each earnings season: who spent, who got paid, and what changed in the filings.
Tool
AI capex tracker
Quarterly capital spending by the six biggest buyers, with cash-flow coverage, leases and backlog.
Tool
Capex payback calculator
How much new revenue a year of AI spending needs, against what the clouds actually earn.
Tool
Useful-life calculator
What one year of difference in server life does to depreciation and operating income.
How to use the map
Every AI stock sits somewhere on this map, and three questions place it. Which layer is it in, and does that layer sit at a bottleneck or in the flow? Does it keep what it is paid: is its gross margin rising with its revenue, or are new dollars passing straight through? And who pays it, with whose money: does it depend on one or two customers, and does it finance any of them?
Then watch the signals that would change the picture. The buyers’ capex guidance and their capex-to-cash-flow ratio say whether the spending can go on. Cloud segment growth and operating margins say whether it is paying. Gross margins at the toll booths say whether scarcity is easing. New investments, guarantees and warrants in the filings say whether suppliers are carrying more of their customers’ costs. Each layer page ends with what to watch for that layer.
Method and sources
Figures come from SEC filings: quarterly XBRL data for companies that file 10-Ks and 10-Qs, and the annual 20-F for TSMC, ASML, ASE, Arm and Nebius. Revenue, margins and cash flow are totals over each company’s latest four reported quarters, and growth compares them with the four before. Companies report on different fiscal calendars, so anything added up across companies is aligned to the nearest calendar quarter: Nvidia’s quarter to late July counts as the June quarter. Data last pulled October 30, 2026.
Capital spending is cash paid for property and equipment, as reported in the cash-flow statement. It leaves out equipment acquired under finance leases, which several buyers use heavily, so it understates their total commitment. Segment figures (AWS, Google Cloud, Microsoft Cloud, Nvidia’s data-center revenue) come from the XBRL facts in each 10-K and 10-Q. Links between companies are quoted from the filing that states them, with a link to the document. Layer totals add up only the companies listed on each layer page with at least $100 million of revenue, so they show direction and proportion, not the whole market.
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 the AI stack
What are AI infrastructure stocks?
Publicly traded companies whose revenue comes from building and running the data centers AI depends on: the cloud buyers that fund them, and the makers of chips, memory, servers, networking, power equipment and chipmaking tools. This map covers 78 of them across eight layers, using only their SEC filings. It describes what each reports; it doesn’t recommend buying or selling any of them.
What is the AI stack?
It is the chain of companies that turn spending on artificial intelligence into physical infrastructure and back into revenue. At the top are the software companies and cloud customers who pay for AI. Below them are the clouds that buy the equipment, then the makers of servers, networking gear and power systems, then the chip and memory makers, and at the bottom the foundries and toolmakers that every chip passes through. We group it into eight layers and follow the money between them.
How much are the big tech companies spending on AI data centers?
Microsoft, Amazon, Alphabet, Meta and Oracle reported $566B of capital spending in the four quarters to Q2 2026, up from $184B two years earlier. Not all of it is for AI, and the companies don’t split it out, but their filings attribute most of the increase to AI and cloud infrastructure. The figure is cash capex; equipment acquired under finance leases comes on top.
Which companies have made the most money from AI spending so far?
By reported gross profit, the chip and memory makers. Over the past two years, 83% of the extra gross profit earned by the supplier layers we track went to those two layers, led by Nvidia and Micron. Companies that assemble servers grew revenue quickly but kept only a small share of it.
Why do AI server makers have such low margins?
Most of the cost of an AI server is the GPUs, memory and networking chips inside it, which the server maker buys from suppliers with far more pricing power. The assembler adds integration, cooling and service, which customers value, but competitors can match. So revenue rises with every rack shipped while gross margin stays in the teens.
What is circular financing in AI?
It is when a supplier helps pay for its own customer’s purchases: by investing in the customer, lending to it, guaranteeing its leases, agreeing to buy back capacity it can’t sell, or giving it shares through warrants. Nvidia, Amazon, Microsoft, AMD and HPE all disclose arrangements like this. They are legal and disclosed, but they mean some reported demand is funded by the sellers themselves.
Is AI capital spending sustainable?
The filings show it is getting harder to fund from cash flow alone. The five biggest buyers spent 97% of operating cash flow on capex in Q2 2026, against 35% in Q3 2023. Cloud revenue is growing fast, but not yet as fast as the spending. Whether that is sustainable depends on whether cloud and AI software revenue keeps compounding while the new equipment is depreciated.
Where does the data on this page come from?
From the companies’ own SEC filings: 10-K and 10-Q financial data in XBRL form, segment figures from the same filings, annual 20-F reports for foreign issuers, and quoted sentences for every link between companies. No estimates, consensus figures or paid data are used.
Does this tell me which AI stocks to buy?
No. This is research on how the industry’s money moves, not a recommendation. It can help you see where a company sits, how much of what it is paid it keeps, and who it depends on, which are the questions to answer before you think about price. Nothing here considers valuation or your circumstances.
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.