AI spending is a capital-expenditure chain, not a return

themePublished 19 October 20263 min readequities, multi asset

Key points

• The IMF's October 2026 outlook says investment in artificial intelligence helped activity hold up, and lists a reassessment of AI expectations as a risk.
• The OECD's September 2026 interim outlook says AI-related investment supported production and trade, and that returns may fall short of the earnings implied by current spending plans.
• Chip design, fabrication, equipment, memory, data centres, power and software sit at different points in the same spending chain.
• Neither institution prints a company revenue, a margin or a shareholder return for that chain.

What is going on

The IMF's October 2026 World Economic Outlook, using information through 1 October 2026, says activity has held up through the war in the Middle East, helped by buffers, private-sector adaptation and investment in artificial intelligence. Its upside case includes stronger AI investment and productivity. Its risk list includes a reassessment of expectations around artificial intelligence. The OECD's interim outlook of 23 September 2026 says AI-related investment supported production and trade in the first half of 2026, and names, as a downside risk, AI investment returns that fall short of the earnings implied by current spending plans. Those are the institutional sentences. They do not identify a supplier.

Why it matters

Spending becomes someone else's revenue immediately, and someone's profit only after cost, depreciation and competition. A cloud operator that buys chips records an asset and then depreciates it. The chip seller may record revenue in the delivery quarter. The equipment maker may have booked the tool a year earlier. The power utility may not connect the site for longer than either of those cycles. A reader who treats 'AI investment is supporting GDP' as 'every listed supplier is earning its cost of capital' has changed the subject.

What to watch

The evidence that would weaken the spending story is not a single share price. It is a cut in disclosed capital-expenditure plans, a lengthening of unused capacity, or the OECD's own test: earnings that do not follow the spending. The evidence that would weaken a particular supplier's claim is a margin that disappears once the customer has a second source. This note does not have those company disclosures.

Risks

Customer concentration means one buyer's pause is the supplier's recession. Depreciation of short-lived servers can absorb the cash the project was supposed to return. Financing costs rise if the build is debt-funded. A productivity gain, if it arrives, can accrue to the customer rather than to the firm that sold the tool. The IMF's growth projections of 3.1 percent in 2026 and 3.4 percent in 2027 are conditional paths, not a valuation of this chain.

Where the spending goes

A usable map, without market shares, runs from design, to fabrication and packaging, to manufacturing equipment, to memory and networking, to the building and the cooling plant, to electricity, and only then to software that a customer will pay for more than once. Each step has its own capital intensity. Equipment and buildings tie up cash for years. Software can have a high gross margin and still fail if customers will not renew. This map is a structure. It is not a ranking.

What this note will not invent

No chip shipment, data-centre megawatt or supplier margin was on the IMF or OECD pages used here. None is added. The electricity note on this desk explains why an announced campus is not operating demand. The minerals note explains why the same build also pulls on copper and minor metals.