AI Capex Explained: Where Half a Trillion Dollars a Year Actually Goes — and Who Gets Paid
"Published 2026 estimates for hyperscaler AI capital spending range widely — roughly $430 billion to $800 billion depending on who is counting and what they include. The money flows through a stack: chip designers and memory makers, then server assemblers, then power and grid, then land and construction, then the financing that pays for it. Each layer has different margins, different competition and different fragility. The layer with the most attention usually has the most expectation already in its price.
"AI capex" is one of those phrases that gets repeated far more often than it gets explained. It matters because it is the mechanism by which an idea becomes an economy — and because it is the largest single driver of what has happened to stock markets over the past three years.
Here is what the money is, where it goes, and who ends up with it.
What capex actually means
Capital expenditure is money spent on long-lived physical assets rather than on day-to-day operations. Buying servers is capex. Paying the electricity bill to run them is not.
The accounting distinction matters more than it sounds. Operating costs hit the income statement immediately. Capital spending does not — it goes onto the balance sheet as an asset and is then written down as depreciation over the asset's assumed useful life, typically several years for computing equipment.
This creates a specific pattern that is worth internalising: the cash leaves now; the earnings hit arrives later, spread out, and it arrives whether or not the investment worked.
That lag is why a company can look profitable during a heavy investment phase and then face years of depreciation drag afterwards.
How big is it
Honestly? Estimates vary enormously, and anyone quoting a single confident number is not being careful.
Goldman Sachs has projected that AI companies may invest more than $500 billion in 2026. Other published 2026 estimates for hyperscaler capital spending run from roughly $430 billion at the conservative end to $600 billion, $690 billion, and around $800 billion at the aggressive end.
The spread is not incompetence. It is definitional:
- Who counts? The four or five largest cloud providers only, or also AI-native companies, sovereign programmes, telecom operators and enterprise buyers?
- What counts? Only chips and servers, or also buildings, land, cooling, and grid connections?
- When counts? Cash spent this year, or commitments signed this year that will be spent over five?
Whenever you see a headline number, the useful reflex is to ask which of those three questions it answered.
The stack: who gets paid, in order
Layer 1 — Silicon
Chip designers, foundries, memory manufacturers and semiconductor equipment makers. This is where the money arrives first and most visibly.
High-bandwidth memory has been a particular bottleneck, which is why memory manufacturers have swung so violently: Korea's KOSPI rose 4.6% in a single session in September 2026 on a memory rally, driven largely by two companies.
Characteristics: enormous margins during shortages, brutal cyclicality, capacity added in multi-year lumps, and a small number of firms controlling critical steps.
Layer 2 — Systems and networking
Server assemblers, networking equipment, optical interconnects, cooling systems.
Characteristics: high revenue, generally thinner margins than silicon. Assembling servers is a competitive business; designing the chip inside them is not.
Layer 3 — Power
This is the constraint people underestimate. A large data centre campus needs electricity at industrial scale, and the grid connection can take longer to obtain than the building takes to construct.
Beneficiaries span utilities, independent power producers, grid equipment manufacturers, turbine makers and fuel suppliers.
Characteristics: regulated, capital-intensive, slow, and rate-sensitive. With the US 10-year Treasury yield near 4.83% in September 2026, utility valuations face a direct headwind regardless of demand growth. There is also political risk: when large industrial users push up residential bills, regulators respond, and the regulatory outcome determines whether the utility earns a return.
Layer 4 — Land and construction
Engineering and construction firms, industrial real estate, data centre operators.
Goldman Sachs research indicates data-centre-related construction employment has risen by about 216,000 since 2022, with roughly 500,000 net new infrastructure jobs projected as needed by 2030 for power and data centre development.
Characteristics: visible backlogs, real revenue, cyclical exposure to a build cycle that will eventually end.
Layer 5 — The financing
The layer that gets the least coverage and carries the most risk.
Spending at this scale has increasingly been funded with debt and structured financing rather than purely from operating cash flow. That changes the risk profile fundamentally: a build-out funded from profits slows down when profits slow down. A build-out funded with debt has fixed obligations that do not care about the revenue outlook.
The July 2026 FOMC minutes noted explicitly that elevated AI-related equity valuations pose a repricing risk that could tighten financial conditions abruptly. Central banks do not usually name a specific theme as a financial stability concern.
The question that decides all of it
Does the revenue arrive?
Everything else is detail. If AI services generate enough revenue to justify hundreds of billions in annual spending, the assets are productive, depreciation is covered, and the build continues.
If revenue undershoots, three things happen roughly at once: capex guidance gets cut, which hits the suppliers' revenue; depreciation on assets already built keeps flowing through income statements for years; and the debt used to fund it becomes considerably harder to service.
That combination is what makes infrastructure cycles unwind faster than they build.
What an investor can reasonably do
Recognise the timing asymmetry. Suppliers get paid during the build. Buyers get paid — if at all — afterwards. Owning the whole market means owning both sides of that, which is a feature rather than a compromise.
Watch capex guidance more than product announcements. Quarterly capital spending guidance from the largest buyers is the leading indicator for every supplier below them in the stack. A single deferral reprices an entire supply chain.
Understand the depreciation lag. A company that spent enormously in 2025 and 2026 carries that cost through its income statement into 2029 and 2030. Earnings comparisons in those years will be harder for reasons that have nothing to do with how the business is performing today.
Do not confuse the theme being real with the price being right. The build-out is genuinely happening; hundreds of billions of dollars are genuinely being spent. Neither fact tells you whether a share price already reflects it. Those are separate questions, and conflating them is how people lose money being correct about the future.
Frequently asked questions
Why do estimates of AI capital spending vary so much?
Because there is no standard definition. Some counts include only the four or five largest cloud providers; others add AI-native companies, sovereign programmes, telecom operators and enterprise buyers. Some count only servers and chips; others include the buildings, land, cooling systems and grid connections. Goldman Sachs has projected AI companies may invest more than $500 billion in 2026, while other published estimates for hyperscaler capex alone run from roughly $430 billion to $800 billion. Always check what a headline number includes before comparing it to another.
Is AI capital spending a good thing for the companies doing it?
It depends entirely on whether the revenue arrives. Capital spending converts cash into assets that are then depreciated over several years, which reduces reported earnings for years after the cash has already gone out the door. If AI revenue grows fast enough to justify it, the spending is an investment. If it does not, the depreciation lands on income statements regardless, which is why analysts watch the gap between capex growth and revenue growth so closely.
Who actually makes money from AI infrastructure today?
In the near term, the suppliers do — companies selling chips, memory, networking, power equipment and construction services have visible revenue from the build-out now. The companies doing the spending are converting cash into assets whose returns are still ahead of them. This is the standard pattern in any infrastructure boom, and it is why the phrase picks and shovels persists: the sellers get paid during the build regardless of how the build turns out.
What would tell me the AI capex cycle is turning?
Watch for capex guidance being cut or deferred rather than raised, lease and purchase commitments slowing in disclosures, memory and chip pricing rolling over, and utilisation or backlog commentary softening at the equipment suppliers. The financing channel matters too: a build-out increasingly funded by debt is more sensitive to interest rates and credit conditions than one funded by operating cash flow.
Primary sources
Data & disclaimer: This article is for educational purposes only and is not financial or investment advice. Figures reflect data available as of Sep 10, 2026, and conditions change — always confirm current pricing, rates and rules with the provider before you act. Written by Elizabeta Dimoska. See our editorial standards and disclosure.
Comments