AI Capex
What the hyperscalers are actually spending on AI infrastructure, and what it is producing
Thesis
Is AI investment turning into actual economic payoff?
- Hyperscaler capex runs at a record share of revenue.
- Data-centre power demand is the physical footprint of the spending.
- Margins, productivity and unit labour costs say whether it pays.
Index
Five indicators, standardized and equal-weighted, updated daily — 100 = January 2020, higher means more build-out and more evidence of payoff
AI Capital Spending Index
149.1
1 wk
↑ +0.0
1 mo
↓ -0.2
1 yr
↑ +16.1
Source: Liquiqui, from SEC EDGAR / FRED / EIA
Where each constituent stands today, in standard deviations from its own average since 2020
- Hyperscaler capexQuarterly · Jun 30, 2026+3.0
- Corporate profit marginQuarterly · Apr 1, 2026+0.9
- Data center power demandDaily · Oct 16, 2026+2.2
- Labor productivityQuarterly · Apr 1, 2026+1.7
- Unit labor costs (inv.)Quarterly · Apr 1, 2026-1.3
Notes
- What is happening
- The index stands at 149.1 on 10 October 2026, a point below its record of 150.2 set on 4 September, against a base of 100 in January 2020 and 133 a year earlier. Capex is at +3.0 standard deviations, the method's cap; power demand at +2.2, productivity at +1.7, margins at +0.9 after the annual revision of the national accounts on 30 September; unit labour costs, inverted, are the only constituent pulling down (−1.3).
- Why it matters
- An investment cycle is judged on two things: how much is spent, and whether it pays. This index combines five measures at equal weight — the four hyperscalers' combined capex, corporate profit margins, data-centre power demand, labour productivity and unit labour costs inverted — each standardised against its own history since 2020. It rises when the build-out accelerates and when evidence of payoff accumulates; neither direction is a verdict on its own.
- What's driving it
- $165 billion of quarterly capex at Microsoft, Alphabet, Amazon and Meta, nearly double a year ago, and Northern Virginia's electricity load up 8% on the year.
- What to watch
- The balance between the index's two halves: as long as margins and productivity rise with capex, the build-out pays. Margins falling while capex stays at +3σ would be the first sign that spending has run too far ahead of returns.
Signals
The three legs of the AI thesis — spending, payoff, and conviction — scored by how strong the evidence is for each
| Criterion | Current Reading | Trend | Comment |
|---|---|---|---|
Big Tech Capex How much are the hyperscalers actually investing, and how fast is it accelerating? | High | ↑ | $165bn last quarter across the four, against roughly $8bn a quarter in 2016. Meta commits close to 50% of its revenue, Microsoft 40%. These are cash outflows already booked, not announcements. |
Macro Productivity Is the promised productivity gain showing up in the aggregate data? | Low | → | Productivity has genuinely picked up: 2.1% a year since 2019 against 1.2% through the 2010s. But that is far short of the IT boom's 3.2%, unit labor costs are rising faster still (2.8%), and nothing ties the acceleration to AI specifically. |
Sentiment & Adoption How convinced is the market, and is real adoption following? | High | ↑ | The SOX has quintupled since early 2023 and the Nasdaq-100 has outrun the equal-weight S&P 2.4-fold since 2016. Read with care: this measures market conviction, not enterprise adoption, which no series on this page yet covers. |
Raw data
Combined Hyperscaler Capex
Combined capex
165 B$
1 yr
↑ +77 B$
Source: SEC EDGAR (XBRL)
Notes
- What is happening
- $165 billion in the second quarter of 2026, against $88 billion a year earlier and $7 billion at the end of 2014. The total has doubled in a year and grown tenfold since 2019; each of the four is at a quarterly record.
- Why it matters
- The combined capex of Microsoft, Alphabet, Amazon and Meta is the most direct measure of the AI build-out: they are the four buyers of chips, land and power plants, and their quarterly SEC filings record cash already spent, not announcements.
- What's driving it
- The race for compute: each one invests so as not to be the one short of capacity when demand arrives, and none can slow down alone.
- What to watch
- The first quarter of sequential decline at any of the four. The 2000 telecom cycle turned on one quarter of cancelled orders, not on a year.
Capex by Company
Source: SEC EDGAR (XBRL)
Notes
- What is happening
- Amazon leads at $54 billion in the second quarter of 2026, Alphabet at 45, Microsoft at 36, Meta at 30. All four have at least doubled in two years; Microsoft and Alphabet more than doubled over the year.
- Why it matters
- The composition says who carries the cycle and how fast each one accelerates: four trajectories that will not turn at the same time.
- What's driving it
- Amazon builds for AWS and its own chips, Alphabet for Gemini and its cloud, Microsoft for OpenAI and Azure, Meta for its open models with no cloud business to monetise them.
- What to watch
- Meta: the only one of the four without cloud revenue to amortise its data centres. A turn at Meta would precede the others'.
Capex / Revenue
Source: SEC EDGAR (XBRL)
Notes
- What is happening
- Close to 50% for Meta and 40% for Microsoft in the second quarter of 2026, against 15 to 25% for Meta and 10 to 13% for Microsoft before 2023. Alphabet sits at 38% and Amazon at 27%, triple their pre-2023 norm.
- Why it matters
- Capex against revenue measures the intensity of investment: the share of each dollar earned that goes straight back into infrastructure. It is the ratio that separates a software company from a utility.
- What's driving it
- Capex grows twice as fast as revenue: AI revenues are arriving, but far behind the spending they are meant to justify.
- What to watch
- A ratio that levels off would mean revenues are catching up; a ratio above 50% at several of the four would put them beside the telecoms of 1999, whose intensity reached 40%.
Corporate Profit Margin
Corporate Profit Margin
16.6%
1 yr
↑ +2.7%
Source: BEA, via FRED
Notes
- What is happening
- 16.6% in the second quarter of 2026 (accounts revised on 30 September), against 13.9% a year earlier and a record 17.6% in 2021. Margins have climbed back towards their peak after two years of drift.
- Why it matters
- The after-tax margin of US nonfinancial corporations, as a share of their value added, says whether the capital being invested still turns into profit: it is the first place an overinvestment cycle shows, when depreciation catches up with revenue.
- What's driving it
- Prices that hold, wage costs contained by a loosening labour market, and AI spending still capitalised rather than expensed.
- What to watch
- Data-centre depreciation reaches the income statement two to five years after the capex. Margins falling below 15% while capex continues would be the signal that the bill is arriving.
Semiconductor Index (SOX)
Source: Yahoo Finance
Notes
- What is happening
- 12,324 on 12 October 2026, up 83% on a year ago and five times early 2023, after a peak of 14,635 on 22 June 2026 and a trough of 10,447 in late July: the retreat from the top, which reached 29%, is back to 16%.
- Why it matters
- The Philadelphia Semiconductor Index is the market's daily read on the AI hardware cycle: chips are the largest item in hyperscaler capex, and their makers get paid before anyone else.
- What's driving it
- AI chip orders exceeding production capacity; the summer's doubts about how long the order cycle lasts gave way in mid-September to a rebound on fresh hyperscaler orders.
- What to watch
- The SOX leads capex by two to three quarters: a fall of more than 30% from the peak, as in 2022, has always preceded a pause in orders.
Market Concentration
Source: Yahoo Finance
Notes
- What is happening
- About 325 on 12 October 2026, base 100 in 2003, against 298 a year earlier and a record 335 in June 2026. The Nasdaq-100 has outrun the average S&P stock 2.4-fold since 2016.
- Why it matters
- The Nasdaq-100 against the equal-weight S&P 500 measures how much of the market's performance rests on the largest technology names: the higher the ratio, the more the index depends on a handful of AI-linked stocks.
- What's driving it
- The earnings of the hyperscalers and Nvidia, a third of the S&P 500, against an average stock that has barely moved since 2021.
- What to watch
- A ratio falling without a market decline would mean the rally is broadening: the healthy path. A ratio falling with the market would say the leaders are taking everything down with them.
Data Center Electricity Demand
Source: US Energy Information Administration
Notes
- What is happening
- 391 gigawatt-hours a day on average over the twelve months to 10 October 2026, just under the record of 393 set in early September, against 362 a year earlier and 280 in 2020: +8% on the year, +40% in six years — far beyond what population or ordinary growth would produce.
- Why it matters
- Electricity demand in the Dominion zone of Northern Virginia, the world's largest data-centre cluster, is the physical footprint of the capex: servers bought have to be plugged in, and load cannot be announced, only measured. Read it as a proxy, not a measurement: it is the zone's total demand, which nobody breaks down by use; the proxy holds because Dominion attributes most of its load growth to data centres.
- What's driving it
- Data centres connected in 2024 and 2025 ramping up; interconnection queues delay the rest.
- What to watch
- Annual growth falling below 5% would say the grid, not capital, now limits the pace of construction in Virginia — or that demand is moving to other states.
Productivity & Unit Labor Costs
Source: BLS, via FRED
Notes
- What is happening
- Productivity rose 2.2% on the year in the second quarter of 2026 and 2.1% a year since 2019, against 1.2% through the 2010s; unit labour costs are up 1.4% on the year, after 2.8% a year on average since 2019.
- Why it matters
- Labour productivity and unit labour costs are where an AI gain would eventually show up: more output per hour, and wage costs per unit that stop rising. Without that, capex is an expense, not an investment.
- What's driving it
- Investment in software and equipment, a loosening labour market, and perhaps the first AI gains in services — which nothing yet lets us isolate.
- What to watch
- The 1990s IT boom delivered 3.2% productivity growth a year for eight years. Two years above 2.5% with unit labour costs below 2% would be the first macro evidence of AI's payoff.
Tech Employment & Semiconductor Output
Source: BLS and Federal Reserve, via FRED
Notes
- What is happening
- Semiconductor production stands at 211 in August 2026 on the index rebased (2022 = 100) at the annual revision, a record, +22% on the year; computer services employment is at 121 in September, down from its March 2023 peak of 127 and 1.3% below a year earlier.
- Why it matters
- Employment in computer systems design and semiconductor industrial production measure the cycle from the production side: the people hired and the chips made, both rebased to 100 in January 2016.
- What's driving it
- Chip plants shipping flat out for data centres, while computer services replace part of their hiring with the tools they sell.
- What to watch
- The gap between the two lines: chip production rising while software employment falls is the first, still ambiguous, sign that AI is substituting capital for labour.
Methodology
1Sources
SEC filings (hyperscaler capex), BEA via FRED (profit margin), EIA (Dominion power demand), BLS via FRED (productivity, unit labour costs).
2Normalisation
Capex as a four-quarter sum, margin as % of value added, power demand as a trailing annual mean; quarterly inputs carried forward on the daily calendar.
3Z-scores
Each constituent minus its mean since 2020, divided by its standard deviation; capped at ±3. Unit labour costs enter inverted.
4Weighting
Five constituents, equal weights.
5Index
Average of the five z-scores × 20 points, rebased to 100 in January 2020. Updated daily.
6Interpretation
Higher means more build-out and more evidence of payoff; neither direction is a verdict on its own. The breakdown says which side is moving.