Capex to Revenue Ratio: 23% at Alphabet, 83% at Oracle

Key takeaways

• Microsoft's capex to revenue ratio went from 13.3% in fiscal 2023 to 34.9% in fiscal 2026, as additions to property and equipment rose from $28.1bn to $115.9bn.
• Alphabet spent 22.7% of 2025 revenue on property and equipment, Meta 34.7%, and Amazon 18.4%. Oracle spent 82.6% of revenue in its year to May 2026.
• Amazon's consolidated 18.4% hides AWS at 75.0%. The segment took $96.5bn of the group's $142.4bn of property additions on 18.0% of group revenue.
• Microsoft added $24.6bn of assets under finance leases in fiscal 2026, which never touch the capex line. Counting them lifts the ratio to 42.4%.
• Oracle's $55.7bn of fiscal 2026 capital spending against $32.0bn of operating cash flow produced free cash flow of minus $23.7bn, against a $638bn backlog.

The capex to revenue ratio has roughly doubled since 2023

How much of a cloud company's revenue now goes straight back into land, buildings, servers and power? That is what the capex to revenue ratio measures, and the annual filings answer it without much ambiguity.

Across five of the largest US cloud and data centre operators, the ratio has roughly doubled since 2023. At Oracle it has gone up almost five times. Here is capital expenditure divided by total revenue, taken from each company's most recent 10-K and the one before it. Each company reports on its own fiscal calendar, which matters and is dealt with further down.

The chart above plots the same ten figures. Two things stand out before any interpretation. Meta was already spending a fifth of revenue on property in 2023, well before anyone called this an AI capital cycle. And Amazon, which spent more cash on property and equipment in its latest year than any of the other four, has the lowest ratio of the five.

Both of those are artefacts of the denominator rather than findings about capital discipline. That is the whole problem with this ratio, and it is worth taking apart before leaning on it.

Where the two numbers come from on the filing

Capital expenditure is a cash flow line, not an income statement line. In the investing section of the cash flow statement, Microsoft calls it "Additions to property and equipment", Alphabet and Meta call it "Purchases of property and equipment", Amazon uses the same wording, and Oracle simply writes "Capital expenditures". It records cash that left the business during the year to buy long-lived assets.

Revenue is the top line of the income statement: $331,839m for Microsoft in the year to June 2026, $402,836m for Alphabet in 2025, $716,924m for Amazon, $200,966m for Meta and $67,357m for Oracle in the year to May 2026.

Dividing one by the other gives you a rough measure of capex intensity, meaning how asset-heavy the business has become relative to what it sells. A company whose ratio climbs is turning a larger share of every sale into fixed assets that have to be paid for now and earn their keep later.

What the ratio does not do is compare cleanly across companies, because the denominators are not the same kind of number.

Amazon looks the most restrained until you read the segment note

Amazon's $716.9bn of 2025 net sales includes merchandise it buys and resells. Online stores alone were $269.3bn, and cost of sales for the group came to $356,414m. Meta's top line is a different animal. Of its $200,966m of 2025 revenue, $196,175m was advertising, and its entire cost of revenue was $36,175m. Putting a data centre budget over each of those produces two numbers that look comparable and are not.

Amazon is the only one of the five whose segment note carries capital additions, and that note settles the question. In 2025 the group added $142,352m of property and equipment. AWS accounted for $96,496m of that, or 67.8%, on net sales of $128,725m, which is 18.0% of group revenue.

So AWS capex ran at 75.0% of AWS revenue in 2025. That is not a rounding difference from the consolidated 18.4%. It is the same company in the same year, and the gap is entirely the retail business diluting the denominator.

The AWS series on its own is the sharpest number in any of these filings. Segment additions were $24,843m on $90,757m of net sales in 2023, a ratio of 27.4%. In 2024 it was $53,267m on $107,556m, or 49.5%. Then 75.0%.

Nobody else gives you the same view. Microsoft's segment note reports revenue, cost of revenue, operating expenses and operating income for each of its three segments, and no capital spending at all. Alphabet's reports revenue and operating income. Meta's reports revenue and income from operations, plus long-lived assets split only between the United States and the rest of the world. Oracle's presents segment revenues and expense categories. The cross-company comparison isn't unfair so much as unavailable.

The numerator misses the leases and the unpaid bills

Cash capex is a payments measure. Two things routinely fall outside it.

The first is leased capacity. Microsoft's lease note reports $24,608m of right-of-use assets obtained under finance leases in the year to June 2026, on top of the $115,948m of cash additions. Those are data centres the company controls and depreciates, and they sit nowhere near the capex line. Add them and the ratio moves from 34.9% to 42.4%. Amazon's segment reconciliation makes the same adjustment visible from the other side: $142,352m of total property additions against $131,819m of cash purchases.

The second is timing. Alphabet disclosed $15,090m of purchases of property and equipment sitting in accrued liabilities and accounts payable at the end of 2025, up from $7,435m at the end of 2023. Oracle reported $5,279m of unpaid capital expenditures for fiscal 2026. Kit that has been ordered, delivered and installed can sit outside the cash line for a quarter or more, which means a year-on-year jump in the ratio is partly a story about payment terms.

None of this is hidden. All of it is in the notes, and none of it is in the headline number.

Oracle is the stress test at 82.6% of revenue

Oracle's fiscal 2026, which ended in May 2026, is where the ratio stops being an abstraction. Capital expenditures were $55,663m against total revenues of $67,357m. Operating cash flow was $31,977m. Oracle's own free cash flow table in the filing reports the result: minus $23,686m, against minus $394m the year before.

The company explains the change in one sentence: "Cash used for capital expenditures increased from $21.2 billion in fiscal 2025 to $55.7 billion in fiscal 2026 primarily due to the expansion of our data centers." It adds that it expects "this upward trend to continue during fiscal 2027 and in the following fiscal years".

The funding shows up in the financing section. Oracle issued $43.0bn of senior notes during fiscal 2026 and $5.0bn of mandatory convertible preferred stock, and disclosed $129.5bn of outstanding debt at the end of May 2026. A ratio of 82.6% has to be financed by someone, and the filing names the sources.

The strongest objection: a $638bn backlog is not in the denominator

Here is the serious case against reading any of this as overspending, and it comes from the same filing.

Oracle disclosed remaining performance obligations of $638 billion at the end of May 2026, against $138 billion a year earlier. Remaining performance obligations are contracted revenue not yet recognised, and the increase is attributed to "certain significant cloud contracts that were entered into during the period". On that reading, fiscal 2026 capex is not being spent to serve fiscal 2026 revenue at all. It is being spent to serve a book of business that has already been signed.

Which is exactly the flaw in a ratio that puts this year's spending over this year's sales. Capex leads revenue. Capacity has to exist before it can be sold, and the contract that pays for it is recognised across the years it runs. A company mid-build looks reckless on this measure. A company that has finished building looks disciplined, whatever either of them earns in the end.

The objection is real and it constrains what the ratio can be used for. It does not make the ratio useless, because backlog is a promise and depreciation isn't. A signed contract can be renegotiated, and a customer can fail. The asset still has to be paid for, still has to be depreciated, and that arithmetic is covered in more depth in our piece on AI capex depreciation.

Capex against operating cash flow asks a harder question

If the denominator is the problem, one fix is to stop using revenue. Operating cash flow is closer to comparable across these five, because it already nets out the cost of goods that makes Amazon's top line so large.

The ranking changes, and that is the point. Amazon, lowest of the five on capex to revenue at 18.4%, is second most stretched here at 94.5%. Nearly every dollar its operations generated in 2025 went back out of the door as capital spending. Microsoft, at 34.9% of revenue, still had roughly a third of its operating cash flow left over. Three years earlier its capital spending took just 32.1% of it.

Once capital spending exceeds operating cash flow outright, the build is being funded from the balance sheet rather than from the business. Oracle is the only one of the five past that line, and it isn't close.

Neither measure says anything about whether the spending earns its cost of capital. Alphabet's segment note gets one step closer by reporting Google Cloud's revenue and costs on their own, which is the subject of our piece on the Google Cloud operating margin.

What these filings can't tell you about the capex to revenue ratio

Start with the fiscal calendars. Microsoft's year ends in June, Oracle's in May, and the other three in December. Comparing Microsoft's year to June 2026 with Alphabet's 2025 puts six extra months of a steeply rising series into one of them. That flatters Microsoft's number relative to the rest, and there is no way to fix it from annual filings alone.

Then the restatements. Meta's 2023 10-K reported purchases of property and equipment of $27,266m for 2023. Its 2025 10-K reports $27,045m for the same year, a difference of $221m. Any ratio history assembled from filings inherits the vintage of whichever filing you read it from, and this one uses the most recent presentation of each figure.

The bigger limit is what the ratio measures. It is an input measure. It tells you how much cash went into fixed assets, and nothing about utilisation, pricing, the useful life assumed, or the return earned. Microsoft at 34.9% and Meta at 34.7% finished their latest years in almost the same place, and nothing in either ratio says whether the capacity behind it is full. The filings do not disaggregate that, and the ratio cannot be made to.

Finally, three years is a short sample and it covers one capital cycle in one industry. It contains no downturn. Every figure here comes from audited annual reports, so it is about as clean as company data gets, but a clean number over a short window is still a short window. Concentration in the buyers of this equipment is its own exposure, which we cover in our piece on NVIDIA index concentration.

What would change the conclusion

The ratio is a build-phase artefact, so the first thing that would change it is the build ending. If capital spending flattens while revenue keeps compounding, every one of these ratios falls without a single decision being made about discipline. Watch the denominator's growth rate, not the numerator's.

Segment disclosure would change it more. Amazon reports capital additions by segment and that single note reprices the entire comparison, from 18.4% to 75.0% for the business people actually mean. If Microsoft, Alphabet or Meta began reporting the same thing, the group ratios in this article would become close to irrelevant overnight.

And leases would change it again. The wider the gap between cash capex and total property additions, the less the headline ratio is worth. Microsoft's finance lease additions grew from $11,633m in the year to June 2024 to $24,608m two years later. If that line keeps growing faster than the capex line, a reader tracking only capex to revenue is tracking a shrinking share of the actual build.

The single number worth watching next is Oracle's operating cash flow against its capital budget in fiscal 2027. Its filing says it expects the upward trend to continue. Whether the $638 billion backlog starts converting into cash at the same speed is the test that a ratio calculated at 174.1% has already set for it.