Capex is the new original sin

The 2026 capital budgets of the four American hyperscalers now resemble a wartime industrial programme. Amazon has guided to about $220 billion. Alphabet to $195–205 billion, with $80.6 billion already spent in the first half. Microsoft around $190 billion. Meta $130–145 billion. OpenAI's Stargate file, with partners, still talks in hundreds of billions and gigawatts. Against that, Apple's choice to pay Google for a model and not pour a city of data centres has gone from looking timid to looking like an operating-margin insurance policy. The question in every board packet this autumn is no longer “are we in AI?” It is “when does this earn?”

1. 01Alphabet raised 2026 capex guidance to $195–205 billion, driven by servers and data centres. First-half spend was $80.6 billion.Company / Times reporting
2. 02Amazon, Microsoft, and Meta have printed similarly historic numbers, with AI infrastructure named as the principal driver.
3. 03OpenAI and partners have described a Stargate footprint on the order of $400 billion-plus and nearly 7 GW under development toward 2027.
4. 04From late June, Apple rallied as money rotated out of AI-capex names. Year-to-date Apple led the Magnificent Seven; the semiconductor index did not.Bloomberg
5. 05Gemini as a consumer app has crossed a billion users. That is the first piece of distribution evidence that the factory has a customer. It is not yet evidence of return on capital at this scale.

We assess with high confidence that 2026 is the year the market stopped paying for ambition and started paying for the shape of the payoff. That does not mean the buildout is a bubble in the 1999 sense. Power, chips, and land are real constraints; someone will own the capacity. It means the multiple now depends on a story about utilisation, pricing power, and whether enterprises will pay for agents at a rate that services this debt of steel and silicon.

Apple is not a counter-proof. Apple is a different business, with a billion devices and a services engine, renting intelligence the way it has long rented search. Copying Apple without the devices is not a strategy. Copying Apple's refusal to pretent that training is the product — that, for most companies, is the briefing.

The original sin is not spending. It is spending without a unit of return you can name in a sentence.

• CFOs, not research leads, now own the AI narrative at public companies.
• Secondary suppliers — power, cooling, networking, construction — remain a cleaner way to sit near the buildout than most model labs.
• “We are investing in AI” is no longer an acceptable line in an earnings call. The follow-up is “show me the utilisation.”

If you are spending, name the unit of return this quarter: tokens served, seats paying, hours saved, or revenue per GPU. If you cannot name it, you are not in the AI business. You are in the construction business, and construction businesses are judged on the building.

• Next hyperscaler earnings: capex guidance versus commentary on utilisation.
• Any Apple language on the inference bill after 9 September.
• Power-purchase and grid-queue stories — they are the real governor.

• Company guidance — 2026 capex figures as reported
• Bloomberg — Rotation from capex names into Apple