Quarterly · from SEC filings

The AI Stack Report

Every quarter, after the biggest AI buyers have filed their results, we update the AI stack map from the new filings and publish what it shows: how much was spent, which suppliers got paid and kept the money, whether cloud revenue is catching up, and what changed in the financing behind it all.

Latest report

Q2 2026 · published October 28, 2026

AI Stack Report: Q2 2026

  • The five largest buyers spent $182B on capital projects in Q2 2026, +22% on Q1 2026 and +87% on Q2 2025. That was 97% of their operating cash flow, against 72% a year earlier.
  • Revenue at the four big cloud businesses reached $132B, +40% on Q2 2025; Oracle cloud infrastructure grew fastest, +93%.
  • Among suppliers, memory grew fastest: +253% on Q2 2025, and the biggest margin move was in memory (+43.3 pts of gross margin).
  • Amazon, Alphabet, Oracle and CoreWeave spent more on capex than they generated from operations in the quarter.

Read the report →

All reports

Every AI Stack Report
QuarterPublishedBig Five capexvs a year earlierCapex ÷ cash flow
Q2 2026October 28, 2026$182B+87%97%

The cycle so far

The five largest buyers’ combined capital spending and operating cash flow by calendar quarter, as of the latest report. The ratio between them is the simplest measure of how hard the build-out is pressing on the companies paying for it: low, and the spending comes from spare cash; near or above 100%, and further growth has to be financed.

Quarterly capex and operating cash flow of the five largest buyers
QuarterCapexOperating cash flowCapex ÷ cash flowDepreciation
Q2 2026$182B$186B97%$46.2B
Q1 2026$148B$158B94%$42.6B
Q4 2025$131B$181B72%$40.5B
Q3 2025$106B$167B63%$35.8B
Q2 2025$97.3B$135B72%$32.0B
Q1 2025$77.8B$120B65%$29.5B
Q4 2024$76.3B$136B56%$30.4B
Q3 2024$61.2B$123B50%$27.0B
Q2 2024$55.6B$115B49%$20.2B
Q1 2024$46.0B$104B44%$19.3B
Q4 2023$44.0B$99.8B44%$21.1B
Q3 2023$38.3B$110B35%$18.9B

What each report measures

The spending. Capital spending by Microsoft, Amazon, Alphabet, Meta and Oracle, plus CoreWeave, for the quarter: cash paid for property and equipment, compared with the quarter before and a year earlier, and set against each company’s operating cash flow. This is the money that becomes revenue for every supplier layer, and the ratio to cash flow shows how much room the buyers have left.

The return. Revenue at AWS, Microsoft Cloud, Google Cloud and Oracle’s cloud infrastructure, and the operating income the segments that report it earn. Cloud revenue is the most direct evidence in the filings that the capacity being built is being sold, and operating income shows whether it is absorbing the depreciation of the new equipment.

Who got paid. Revenue at the supplier segments that matter most (Nvidia’s data-center business, Broadcom’s semiconductors, Micron’s DRAM, AMD’s data-center segment and Dell’s AI servers) and revenue and gross margin for each layer of the stack. Revenue shows where the money went; gross margin shows who kept it, which is where pricing power and scarcity appear first.

The loop and the commitments. The financing links between suppliers and customers that filings disclose (investments, guarantees, capacity buy-backs, customer warrants and long power contracts), and the data-center leases the buyers have signed that have not yet started. These are the parts of the build-out that don’t appear in capex, and the parts most likely to surprise investors if demand changes.

What changed. From the second edition on, each report lists what moved since the one before: new or changed financing links, changes in lease commitments, changes in the server lives companies use to depreciate their equipment, and the shift in the buyers’ capex-to-cash-flow ratio.

The phases so far

Measured by the five largest buyers, the build-out has moved in steps. In 2024 they spent $239B, 50% of their operating cash flow; In 2025 they spent $412B, 68% of their operating cash flow. Quarterly spending went from $38.3B in Q3 2023 to $182B in Q2 2026.

The share of cash flow tells the story better than the dollars. It was lowest in Q3 2023, at 35%, when the buyers were still digesting earlier capacity and spending from spare cash, and highest in Q2 2026, at 97%. Between the two, capex stopped being a line item that buybacks and dividends could easily sit beside and became the main use of the buyers’ cash.

The next phase will be decided by whether that ratio can keep rising. Above 100% for the group, the build-out depends on borrowing and on the leases and partnerships that keep spending off the balance sheet; a falling ratio would mean either cash flow catching up, as cloud revenue grows, or the buyers slowing down. Each report records which way it went.

How each report is built

We wait until all five of the largest buyers have filed their 10-Q or 10-K for the quarter; Oracle’s fiscal quarters end a month or two after the others’, so that sets the timing. Then we re-read the filings: quarterly financial data and segment figures from each company’s XBRL, the notes on leases, purchase commitments and useful lives, and the sentences that disclose investments, guarantees and warrants between suppliers and customers. Each financing link is kept only if the filing’s own sentence contains its figures.

The numbers are then frozen into the report’s own file and the page is built from that file alone. That is a deliberate constraint: when a company later restates a quarter or our methods improve, the live pages change and the past reports don’t, so anyone citing a report can rely on it still saying what it said. Corrections to a report, if one is ever needed, are noted on the report itself rather than made silently.

Why a quarterly report

The AI build-out moves in quarterly steps. Each earnings season the largest buyers reveal how much more they are spending, the suppliers reveal how much of it they captured, and the notes to the financial statements reveal new commitments: leases, guarantees, investments in customers. Read one company at a time, those disclosures are easy to miss. Read together, in the same frame each quarter, they show the direction of the whole cycle.

Each report uses the same sources and methods as the AI stack map: XBRL financial data and segment disclosures from 10-K and 10-Q filings, calendar-aligned so companies with different fiscal years can be added up, and financing links quoted from the filings that disclose them. Its numbers are frozen when it is published, so the archive is a record of what the filings showed at each point in the cycle.

Investors can use each report as a checklist for the quarter: is the spending still accelerating, is cloud revenue keeping pace, are the toll booths still keeping the profit, and is more of the demand being financed by the suppliers themselves? Companies selling into the build-out can use it to see their customers’ economics, and boards can use it to frame the AI questions they put to management. Because every report uses the same structure, comparing two quarters side by side shows what actually changed rather than what the headlines emphasized, which is often not the same thing at all.

For a talk or board briefing built on the latest report, see work with us.

Free · quarterly · from SEC filings

Get each report when it’s published

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.

About the AI Stack Report

When is the AI Stack Report published?

Once a quarter, after Microsoft, Amazon, Alphabet, Meta and Oracle have all filed their quarterly reports with the SEC. That is usually five to eight weeks after a calendar quarter ends, depending on Oracle’s fiscal calendar.

What does each report cover?

The largest buyers’ capital spending and how much of their cash it used, revenue at the big cloud businesses, revenue and margins across the supplier layers of the AI stack, the financing links between suppliers and customers disclosed in filings, and what changed since the previous report.

Why do past reports never change?

Each report saves its numbers when it is published. Companies restate figures and our data is refreshed, but a dated report should keep saying what it said. The live AI stack pages always carry the latest numbers.

Can I cite the reports?

Yes. Please cite “Investing in AI, AI Stack Report” with the quarter and link to the report page. Every figure traces back to a company’s SEC filing.

Which companies does the report cover?

The five largest AI buyers (Microsoft, Amazon, Alphabet, Meta and Oracle) and CoreWeave for spending, the four cloud businesses whose results are disclosed for the return, and the companies on the AI stack map, from Nvidia and Broadcom to Micron, Arista, Dell, Vertiv and TSMC, for the supplier layers.

How is the report different from the AI stack map?

The AI stack map always shows the latest numbers and the analysis behind them. Each report is a dated snapshot of one quarter with the changes since the previous quarter, so together the reports form a record of the cycle that the live map can’t provide.

Why focus on five buyers?

Because their spending dominates. Microsoft, Amazon, Alphabet, Meta and Oracle account for most of the capital spending on AI data centers that public filings disclose, and their capex is the revenue of nearly every supplier on the map. CoreWeave is shown alongside them as the largest listed GPU-only cloud.

Does the report cover companies outside the US?

Where they file with the SEC. TSMC, ASML, ASE, Arm and Nebius file annual reports on Form 20-F, so they appear in the AI stack map with annual figures. Quarterly comparisons in the report use companies that file 10-Qs, because those are the ones with consistent quarterly data.

What happens if a company restates its numbers?

The live AI stack pages pick up the restated figures at the next data refresh. Past reports keep the numbers they were published with, and if a restatement would change a report’s conclusions, we add a dated note to that report rather than altering it.

Can I get the data behind a report?

The figures in each report are shown in its tables and sourced to the filings. For the underlying datasets, including quarterly capex, segment revenue and the financing links, in a form analysts can use, get in touch through the work-with-us page about licensing.

Is the report investment advice?

No. It reports what companies disclosed and what it shows about the AI build-out. It does not recommend buying or selling any security.

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.