What is AI capex, and why does it matter for stocks?

AI capex is the money a company spends on the long-lived things that run AI: chips, servers, networking gear, data center buildings and the power to run them. It is capital expenditure, so the cost is spread over several years as depreciation. One company's capex is other companies' revenue, which is why the market watches it so closely.

The catch is that nobody reports a line called "AI capex". Every figure you see is either a company's total capital spending or someone's estimate of the AI share of it. That is why two articles can quote two different totals for the same year and both be right.

What counts as AI capex?

Anything bought to build or run AI capacity that lasts more than a year:

• AI chips: GPUs and custom accelerators
• Servers and the high-bandwidth memory inside them
• Networking and optical gear that connects the servers
• Data center buildings and land
• Cooling, including liquid cooling
• Electrical equipment and on-site power

What does not count: salaries, research spending, and computing rented from someone else. A company that rents GPUs by the hour from a neocloud books that as an operating expense. The neocloud that bought the GPUs books them as its own capex. That is why CoreWeave and Nebius appear in capex tables next to Microsoft and Amazon.

How much are companies spending on AI right now?

The seven buyers in Supercycle's Buildout Tracker (Amazon, Microsoft, Alphabet, Meta, Oracle, CoreWeave and Nebius) spent $657 billion over their last four reported quarters, as of October 4, 2026. In the July to September reporting round, four of them raised their guidance and none cut it. The September capex report has the company-by-company table with a source for every row.

Why do the AI capex totals disagree?

Three choices change the total, and articles rarely say which ones they made:

1. Which companies are included. Some totals cover only Amazon, Microsoft, Alphabet and Meta. Others add Oracle, the neoclouds, xAI or the Chinese platforms.
2. What counts as spending. Meta's 2026 guidance includes principal payments on finance leases, which other companies leave out. Amazon's headline capex figure for Q2 2026 was $53.1 billion, while its cash purchases of property and equipment in the same quarter were $54.2 billion.
3. Whose year. Microsoft's fiscal year ends in June and Oracle's in May, so "2026 capex" means different months for different companies.

Show the check: each company's cash flow statement has the line, usually called purchases of property and equipment, in its quarterly 10-Q. The tracker links the filing or results release behind every figure it shows.

Where does an AI capex dollar go?

Into the supplier layers below the buyers. Each one is a category on Supercycle, with its tracked public companies and their AI Exposure Scores.

What the money buys	Supercycle layer	Tracked stocks
AI chips	Semiconductors	67
The machines that make the chips	Chip Equipment & Foundries	37
Memory and storage	Memory & Storage	9
Servers and racks	Data Centers & Hardware, EMS & Manufacturing	65, 9
Networking and optics	Networking & Optics	24
Cooling	Cooling & Thermal	10
Power and grid equipment	Power & Energy, Utilities & Grid	62, 42

Where the data stops: the buyers do not say how their capex splits across these layers, so the table shows where the money goes, not how much of it each layer gets.

Why does capex matter for stocks?

It cuts two ways, depending on which side of the purchase a company is on.

For suppliers, capex is revenue. A buyer's guidance is a public forecast of orders, given months before those orders show up in any supplier's results. That is why a hyperscaler's earnings call can move chip, power and data center stocks that report weeks later. The hyperscaler capex entry covers this in more detail.

For the company spending, capex is a cost that arrives later. The cash leaves now, and the expense reaches the income statement over the asset's useful life as depreciation. That makes the assumed life of a server a real number for earnings. Alphabet depreciates servers and network equipment over six years, according to its 2025 annual report. Meta moved its servers to 5.5 years at the start of 2025, while Amazon shortened part of its server fleet from six years to five, citing the faster pace of AI hardware. A shorter life means a bigger yearly expense for the same spending.

How can you track AI capex yourself?

1. Open the company's latest quarterly report (10-Q) or results release. You should see a cash flow statement.
2. Find purchases of property and equipment, or additions to property and equipment. That figure, as a negative number, is the quarter's cash capex.
3. Read the guidance on the earnings call or in the release. You should find a range for the coming year, often with a word on next year.
4. Compare it with the company's row on the Buildout Tracker. The row shows the same figures with the definition used and a link to the source.

The next round of reports from the large buyers arrives in late October 2026, and it is the first update to every number in this post.