Hyperscaler capex is still outrunning revenue: up 93% against 19%
Five companies' data-centre spending has become the largest single source of demand for chips and power; the question is how long it keeps rising faster than their own revenue.Is hyperscaler capex still the thing to watch, and when does its lift to earnings end?
Capex is still outrunning the companies' revenue
Is hyperscaler capex still the thing to watch, and when does its lift to earnings end?- The five companies spent $691bn on capex in the four quarters to Q3 2026, +93% on a year earlier.→ Exhibit 1
- Capex grew 74 points faster than revenue over the four quarters to Q3 2026.→ Exhibit
- Semiconductor output was +19.9% on a year earlier in September 2026.→ Exhibit
| What would change this view | Trigger | Today | Decided by |
|---|---|---|---|
| The spending stops growing | the five companies' capex over four quarters below a year earlier | 93.5% (Jul 2026)not crossed | – |
| Spending falls behind revenue | capex growth below revenue growth, four quarters each | 74.4pp (Jul 2026)not crossed | – |
| The chip cycle turns | semiconductor output below a year earlier | 19.9% (Sep 2026)not crossed | – |
The reasoning behind each point
- LevelThe five companies spent $691bn on capex in the four quarters to Q3 2026, +93% on a year earlier.
A year earlier the four-quarter total was $357bn. In Q3 2026 alone Amazon spent the most, $77.3bn (34% of the five).
- Against revenueCapex grew 74 points faster than revenue over the four quarters to Q3 2026.
Four-quarter revenue rose +19.1% and capex +93.5%. The test of the view is capex growth falling below revenue growth.
- ChipsSemiconductor output was +19.9% on a year earlier in September 2026.
It has been above a year earlier for 86 months. Since 1990 it has fallen below a year earlier 5 times, for 1 to 11 months.
Next: Industrial production — output across all industries
The claim as written
Hyperscaler capex is still the thing to watch: while it grows faster than the companies' own revenue it carries chip demand, and its lift ends when it stops growing.
The canvas's board S10 reads the five companies' capex as the driver of tech and semiconductor earnings and asks when the lift ends. This page tracks the spending in the companies' own filings — the level, its growth against revenue — and the chip cycle in the Federal Reserve's output data. It does not yet print the earnings effect of a pause: that needs the sector engine.
Argued in the : 2 dated documents, listed with their sections under Sources in the Appendix.
Capex rose 93% over four quarters, revenue rose 19%
How much are the five companies spending, and is it growing faster than what they earn?Exhibit 1 · The five companies spent $691bn on capex in the four quarters to Q3 2026, +93% on a year earlier.
About this exhibit
Capex grew 74 points faster than revenue over the four quarters to Q3 2026.
About this exhibit
Capex by company, Q3 2026
| Company | Quarter | A year earlier | y/y | Share |
|---|---|---|---|---|
| Microsoft | 39.4 | 19.4 | +103% | 17% |
| Amazon | 77.3 | 35.1 | +120% | 34% |
| Alphabet | 51.4 | 24.0 | +114% | 22% |
| Meta Platforms | 33.9 | 18.8 | +80% | 15% |
| Oracle | 28.5 | 8.5 | +235% | 12% |
Quarterly capex of each of the five hyperscalers
The chip cycle the spending feeds
Is semiconductor output still rising, and how long have its downturns lasted?Semiconductor output was +19.9% on a year earlier in September 2026
About this exhibit
Runs of semiconductor output growth below zero
| From | To | Months |
|---|---|---|
| Nov 2001 | Dec 2001 | 2 |
| Nov 2008 | Sep 2009 | 11 |
| May 2015 | Nov 2015 | 7 |
| Jan 2016 | Jan 2016 | 1 |
| Nov 2018 | Jul 2019 | 9 |
| Aug 2019 | Sep 2026 (current, rising) | 86 |
3 tests of the view, none crossed
What would change the view?What would change the view, test by test
| Status | Test | Latest reading | Threshold | If crossed | Threshold source |
|---|---|---|---|---|---|
| not crossed | The five companies' capex over four quarters falls below a year earlier | 93.5% (Jul 2026)an earlier page of this site | < 0.0% | The spending has stopped growing; the canvas's pause path is the one under way. | Proposal: integrating the earnings-stretch valuation work into the site (rev 42) · Page designs — AI capex: would change the view · 17 Nov 2026 |
| not crossed | Capex growth falls below revenue growth, four quarters each | 74.4pp (Jul 2026)an earlier page of this site | < 0.0 pp | Spending no longer runs ahead of the companies' own revenue: the spending is being paced by what the companies earn. | Proposal: integrating the earnings-stretch valuation work into the site (rev 42) · Page designs — AI capex: open charts (capex growth less revenue growth) · 17 Nov 2026 |
| not crossed | Semiconductor output falls below a year earlier | 19.9% (Sep 2026)an earlier page of this site | < 0.0% | The chip cycle has turned with or ahead of the spending. | Proposal: integrating the earnings-stretch valuation work into the site (rev 42) · Page designs — AI capex: folded (the pause runs) · 17 Nov 2026 |
The paths the canvas named, without odds
What does the spending do next?What the spending does next
- The spending keeps risingCapex keeps growing faster than revenue, chip output keeps rising, and the lift to tech and semiconductor earnings runs on.
- A capex pauseCapex growth falls below zero over four quarters, chip output follows it down, and the earnings lift reverses — the canvas's pause run.
The canvas simulates the paths but its odds have no record yet (D-2026-11-17-OR13; the proposal's change 6), so none are shown.
What would each path move?
None of this theme's branches is on the Scenarios set; no model on the site prices a capex path yetWhat each path of the spending would move
Nothing is projected here. The spending keeps rising and A capex pause: the site's macro model carries no capex variable yet (D-2026-11-17-OR13).
The earnings effect of a capex pause on tech and semis needs the sector engine, which waits on a consensus source (D-2026-11-17-OR8).
What each path would do to core PCE, unemployment and real GDP is read on Thematic impact, under Outlook.
Consensus capex as quoted, with date and source
| What | As quoted | Source |
|---|---|---|
| Capex against cash flowsubscription data | Reuters analysis of LSEG consensus estimates (22 Jul 2026): Microsoft, Alphabet, Amazon, Meta Platforms and Oracle are expected to raise annual capex by about $534 billion between 2025 and 2027, against about $340 billion more operating cash flow. | Reuters, P. Murugaboopathy and G. Dogra, 22 Jul 2026 · analysis of LSEG consensus estimates · 22 Jul 2026 |
| Consensus revisionssubscription data | The same analysis: current-year consensus capex for the five rose from about $485 billion in January to about $730 billion in July 2026. | Reuters, P. Murugaboopathy and G. Dogra, 22 Jul 2026 · analysis of LSEG consensus estimates · 22 Jul 2026 |
Timeline
Dated events in the record
The dated events behind the view
- 22 Jul 2026Reuters' analysis of LSEG consensus: the five companies' capex is expected to outrun their free cash flow by 2027.
Reuters, P. Murugaboopathy and G. Dogra, 22 Jul 2026 · analysis of LSEG consensus estimates · 22 Jul 2026
- 17 Nov 2026The theme opens on the site, reading the companies' filings through the third quarter of 2026.
the site's specification for this page · 17 Nov 2026
Method
Filings, sums, output data and licences27 checks, all pass
Five companies — Microsoft, Amazon, Alphabet, Meta Platforms and Oracle — report their capital spending and revenue to the SEC every quarter. This page adds them up, sets spending growth against revenue growth, and watches the chip cycle the spending feeds.
Definitions
- ThemeWhat the page reads.
The theme reads the spending where it is recorded: each company's capital expenditure and revenue as filed with the SEC, read into calendar quarters the way the Fundamentals pages read every S&P 500 company, and semiconductor output from the Federal Reserve. Four-quarter sums remove the quarterly pattern and the companies' different fiscal years.
- CapexThe companies' own line.
Capex is cash paid for property, plant and equipment as each company tags it (the tag read is named in the Data tab). Finance leases, which some companies report separately, are not added, so the totals are the companies' own capex line, not their full spending on data centres. A four-quarter sum is drawn only where all five companies report all four quarters.
- ChipsOutput and its runs.
Semiconductor output is the G.17 index for semiconductors and related electronic components (NAICS 3344), seasonally adjusted, as a percent change on twelve months earlier; a run is the stretch of months on one side of zero.
- ConsensusQuoted, never used.
Consensus capex is quoted from the record with its date and source and is never an input: the site has no cleared source for consensus estimates (D-2026-11-17-OR8).
- StatusHow triggers are scored.
crossed — the threshold condition holds on the latest observation (for a run, on each of the last n); closing — not crossed, the series has printed since the September report's data date (11 September 2026), and the gap to the threshold is half or less of the gap on that date; not crossed — otherwise; pending — the observable is dated in the future or no page of this site carries it yet; not tracked here — an event, or subscription data this site does not display. Colours flag attention, not good or bad news for the book.
- OddsNone given.
No odds are given for this theme. The two paths are named without a number.
- DocumentsWhere the view was argued.
- the site's specification for this page · 17 Nov 2026
- Proposal: integrating the earnings-stretch valuation work into the site (rev 42) · Page designs — AI capex (theme) · 17 Nov 2026
- LicencesWhat is quoted only.
The filings and the output data are published. The Reuters figures are quoted only.
How the numbers are made
U.S. Securities and Exchange Commission, EDGAR XBRL company facts (Microsoft, Amazon, Alphabet, Meta Platforms, Oracle) ──▶ each company's capex and revenue read into calendar quarters through the concept dictionary (core.aggregate, the Fundamentals pages' reader) ──▶ quarterly capex and revenue quarterly capex and revenue ──▶ four-quarter sums, drawn only where all five companies report all four quarters; growth on a year earlier and capex growth less revenue growth ──▶ capex against revenue Board of Governors of the Federal Reserve System, G.17 Industrial Production and Capacity Utilization ──▶ semiconductor output (IP.G3344.S) as a change on twelve months earlier, and its runs above and below zero ──▶ the chip cycle capex against revenue + the chip cycle ──▶ trackers, tests, tiles, takeaways and section titles generated from the numbers; the log's quoted consensus printed beside them, never used ──▶ the page
Sources
- publicU.S. Securities and Exchange Commission, EDGAR XBRL company facts (Microsoft, Amazon, Alphabet, Meta Platforms, Oracle)
Source: U.S. Securities and Exchange Commission (EDGAR XBRL financial data) · https://data.sec.gov/api/xbrl/companyfacts/
- publicBoard of Governors of the Federal Reserve System, G.17 Industrial Production and Capacity Utilization
Source: Board of Governors of the Federal Reserve System (G.17) · https://www.federalreserve.gov/releases/g17/
- publicU.S. Securities and Exchange Commission, EDGAR XBRL company facts (Microsoft, Amazon, Alphabet, Meta Platforms, Oracle)
Source: U.S. Securities and Exchange Commission (EDGAR XBRL financial data) · https://data.sec.gov/api/xbrl/companyfacts/
- publicBoard of Governors of the Federal Reserve System, G.17 Industrial Production and Capacity Utilization
Source: Board of Governors of the Federal Reserve System (G.17) · https://www.federalreserve.gov/releases/g17/
Pages this one copies from
| Page | Title | Updated | Data to | Arrays | Values | Source checks |
|---|---|---|---|---|---|---|
| Range and scenarios | Projections | 2026-11-17 | 16 Nov 2026 | 0 | 2 | 80 pass |
| Transmission | Transmission | 2026-11-17 | Oct 2026 | 0 | 0 | 67 pass |
Reproduction checks
| Check | Published | Computed | Tolerance | Result |
|---|---|---|---|---|
| Capex, four quarters to Q3 2026: the total equals the sum of the five companies' four-quarter sums ($bn) | 691.015 | 691.015 | 0 | pass |
| Capex in Q3 2026: the five companies' quarterly capex sums to the shares' denominator ($bn) | 230.422 | 230.422 | 0 | pass |
| Capex growth to Q3 2026 = four-quarter total over the total a year earlier, less one (%) | 93.468 | 93.468 | 0 | pass |
| Capex growth less revenue growth to Q3 2026 = the two growth rates' difference (pp) | 74.366 | 74.366 | 0 | pass |
| Company shares of the latest quarter's capex sum to 100 (%) | 100 | 100 | 0 | pass |
| Semiconductor output growth, 2026-09 = the G.17 index over its level twelve months earlier, less one (%) | 19.899 | 19.899 | 0 | pass |
| The runs of semiconductor output growth cover every month since 1990 once (months unaccounted) | 441 | 441 | 0 | pass |
| Every quarter drawn since 2018-07 carries all five companies' capex (quarters missing one) | 0 | 0 | 0 | pass |
| Trigger statuses are computed for all 3 triggers (statuses outside the rule) | 0 | 0 | 0 | pass |
| What each path would move: a branch Scenarios does not project, or a theme branch not on its set, carries no number here (numbers on such rows) | 0 | 0 | 0 | pass |
| What each path would move: every projected row copies the ten-year, the S&P 500 and the mortgage move of the one Scenarios branch it names (rows missing one) | 0 | 0 | 0 | pass |
| Headline: one conclusion of at most 14 words, not a question, with one or two figures ('Hyperscaler capex is still outrunning revenue: up 93% against 19%'; faults: none) | 0 | 0 | 0 | pass |
| Thesis.status is one of active, closing, resolved (now 'active'; violations) | 0 | 0 | 0 | pass |
| Active theme: no resolved block is carried before the status is resolved (fields present) | 0 | 0 | 0 | pass |
| Active theme: no closing_tracker is named before the status is closing (violations) | 0 | 0 | 0 | pass |
| Scenario set 'buildout': the issue gives no probabilities, none are shown, and the set cites its source (violations) | 0 | 0 | 0 | pass |
| All 3 triggers cite the issue document, section and date their threshold comes from (triggers without) | 0 | 0 | 0 | pass |
| Every data trigger has a numeric threshold (triggers without) | 0 | 0 | 0 | pass |
| All 2 timeline events carry a date, a statement and a source (events without) | 0 | 0 | 0 | pass |
| All 0 exhibit references map to a published panel, a panel declared pending, or none (unmapped) | 0 | 0 | 0 | pass |
| The Range and scenarios page: the source page is published with all its checks passing (dated 17 Nov 2026, 80 checks) | 1 | 1 | 0 | pass |
| The Range and scenarios page: the site still serves the page this one copied (dated 17 Nov 2026) | 1 | 1 | 0 | pass |
| The Transmission page: the source page is published with all its checks passing (dated 17 Nov 2026, 67 checks) | 1 | 1 | 0 | pass |
| The Transmission page: the site still serves the page this one copied (dated 17 Nov 2026) | 1 | 1 | 0 | pass |
| The Range and scenarios page: every copied single value equals the source page (2 values; mismatches) | 0 | 0 | 0 | pass |
| Every read point names the exhibit it is read from (`ex`; 3 points; without one: none) | 0 | 0 | 0 | pass |
| Every read point's exhibit code names a card of this page (themes.ai_capex-…; naming another page: none) | 0 | 0 | 0 | pass |
- U.S. Securities and Exchange Commission, EDGAR XBRL company facts (Microsoft, Amazon, Alphabet, Meta Platforms, Oracle) — Source: U.S. Securities and Exchange Commission (EDGAR XBRL financial data)
- Board of Governors of the Federal Reserve System, G.17 Industrial Production and Capacity Utilization — Source: Board of Governors of the Federal Reserve System (G.17)
- U.S. Securities and Exchange Commission, EDGAR XBRL company facts (Microsoft, Amazon, Alphabet, Meta Platforms, Oracle) — Source: U.S. Securities and Exchange Commission (EDGAR XBRL financial data)
- Board of Governors of the Federal Reserve System, G.17 Industrial Production and Capacity Utilization — Source: Board of Governors of the Federal Reserve System (G.17)