US capex overtook buybacks in 2025. Five names did it

For most of the past decade the complaint about US companies has been that they would rather buy their own shares than build anything. In 2025, measured across 361 listed non-financial companies, that stopped being true.

Capital expenditure was $764.7 billion. Share repurchases were $727.9 billion. It is the first year in this period that building outweighed buying back.

Then you remove five companies, and it is true again.

The panel

Filers reporting both buybacks and capex, every year 2021–2025    1,316
less: combined 2025 spend under $500M                            −905
less: unlisted subsidiary filers                                   −6
less: banks and other financials (SIC 6000–6799)                  −44
───────
Panel                                                                361

Financials are excluded because they run large repurchase programmes against almost no property and equipment. Leaving them in inflates the ratio without saying anything about the trade-off being measured. The panel is fixed — only companies present in all five years — so this measures change at the same companies rather than change in the population.

What the aggregate says

Over five years capital expenditure rose 89.6% while buybacks rose 21.2%. Read as a statement about corporate America, that is a decisive shift in how cash is being used.

Who moved it

Capital expenditure across the panel rose $361.4 billion between 2021 and 2025. Here is where that increase came from.

Those five account for 62.7% of the entire increase across 361 companies. The four technology names alone account for 58.9%. Alphabet by itself is 18.5% of it.

Remove them and the trend is gone

The same panel, minus those five:

For the remaining 356 companies, the ratio moved from 1.41 to 1.38 in five years. Capex rose 41.9%, which is real growth — but the balance between building and repurchasing is where it was.

“Corporate America is investing again” and “five companies are building data centres” produce the same aggregate. They are not the same statement, and only one of them survives disaggregation.

Why this matters for reading any market-wide number

The mechanism here is not specific to capex. When a handful of companies become large enough relative to an index, aggregate statistics start describing those companies rather than the population. The aggregate is not wrong — $764.7 billion was genuinely spent — it is just answering a different question than the one most readers think they are asking.

The test is cheap: recompute without the top few and see whether the finding survives. Here it does not, and that is the finding.

It also reframes a related observation. Capex exceeding operating cash flow has become common enough that 46 companies with over $1 billion of operating cash flow failed that test in 2025. Several names appear on both lists. The build is real; it is just not broad.

What would make this wrong

• A fixed panel has survivorship in it. Companies that delisted, were acquired or went bankrupt during the period are excluded, and so is anything that listed after 2021. Nothing here captures private capital spending, which is where a large share of datacentre construction now sits.
• The $500M floor is arbitrary. It removes 905 smaller filers to keep the aggregate from being driven by companies where a single project distorts the ratio. A different floor gives a different panel.
• Buybacks are gross, not net. These figures do not subtract shares issued for compensation, so they overstate the reduction in share count. A net measure would be lower for every company here.
• Capex is as-filed and unadjusted — no netting of asset sales, partner contributions or government incentives, and no adjustment for finance leases, which sit outside capex entirely.
• Fiscal years are not aligned. Oracle’s “2025” is the year ended 31 May 2026; Walmart’s ends in January. The SEC assigns each annual period to the nearest calendar frame, and off-cycle filers were not excluded.
• Two rows were spot-checked, not 361. Alphabet and Oracle were verified against their own filings; the rest come from the frames data as filed.
• This describes what happened. It is not a forecast and implies nothing about whether either use of cash was the right one.

Check it yourself

The three frame endpoints are linked below and return raw JSON. Keep companies that reported both buybacks and capex in all five years, take the larger of the two capex tags per company-year, drop SIC codes 6000–6799, apply a size floor, and sum. Then do it again without the five largest capex increases and compare.

If a figure here does not match a filing, tell us and it will be corrected on the article and on the corrections log, with the date. The reason two capex tags are needed rather than one is explained in our note on how reported capex understates spending.