The AI capex supercycle is the surge in capital spending on AI infrastructure — data centres, chips and power — led by the largest cloud companies, whose combined capital expenditure is on track to approach US$700 billion in 2026, roughly double 2025. For long-term asset owners it reshapes equity concentration, energy and infrastructure demand, and the risk that spending outruns returns.
A handful of companies are about to spend something close to the annual GDP of a mid-sized economy on machines, buildings and electricity — in a single year, and again the year after. The AI capital-expenditure supercycle is the largest concentrated build-out of productive infrastructure in modern corporate history. For universal owners who hold a slice of the entire market, it is not a sector story to watch from the sidelines. It is reshaping the portfolio they already own.
What the supercycle actually is
"Capex supercycle" describes a sustained, multi-year surge in capital spending well above normal replacement levels. The current one is driven by artificial intelligence: the data centres, specialised chips, networking gear and power infrastructure needed to train and run AI models. What makes it a supercycle rather than a spending blip is both its scale and its persistence — capex has been compounding at extraordinary rates for several years and shows little sign of slowing.
The combined capital expenditures of the largest US cloud builders have grown at an average of roughly 72% per year since mid-2023, around the time large language models entered mainstream use. That is not a typical investment cycle. It is an arms race for compute.
The numbers
The headline figures are difficult to overstate. In 2026, the five largest US hyperscalers — Amazon, Alphabet, Meta, Microsoft and Oracle — have collectively guided toward capital expenditure of roughly US$660 billion to US$725 billion, depending on the estimate, nearly double 2025 levels. Approximately three-quarters of that aggregate is expected to fund AI-specific infrastructure, putting AI capex alone around the US$450 billion mark.
The individual commitments are striking. Amazon has signalled roughly US$200 billion in 2026 capex, much of it for data centres. Alphabet has guided to US$175-185 billion, Meta to US$115-135 billion, Microsoft toward US$120 billion or more, and Oracle near US$50 billion. These are not speculative projections from analysts; they are spending plans the companies themselves have put on the record.
Beyond the hyperscalers, dedicated mega-projects have raised the ceiling further. The Stargate project — a venture of OpenAI, Oracle, SoftBank and Abu Dhabi's MGX — targets up to US$500 billion and roughly 10 gigawatts of AI capacity, and by late 2025 the partners said more than US$400 billion was already committed across announced sites. The binding constraint on all of this is increasingly power: the International Energy Agency projects global data-centre electricity demand will more than double by 2030 to around 945 terawatt-hours, roughly Japan's entire consumption today. We examine that constraint in data center power demand and the grid.
Why a diversified owner cannot ignore it
The supercycle reaches the long-term portfolio through three channels.
The first is equity concentration. The companies leading the build-out are also the largest constituents of global equity indices. A passive global-equity allocation — the backbone of most asset-owner portfolios — is now heavily exposed to a small group of AI-spending mega-caps. The returns of a "diversified" index increasingly hinge on whether that concentrated spending pays off, which quietly undermines the diversification owners think they have.
The second is infrastructure and energy demand. AI data centres consume electricity on an industrial scale, reviving demand for power generation, grid upgrades, natural gas, nuclear and renewables, and the physical real estate of data centres themselves. For owners of infrastructure, real assets and utilities, the supercycle is a demand tailwind that touches assets far from the technology sector.
The third is the financing chain. The spending is so large that it can no longer be funded from corporate cash flow alone. It is increasingly financed with debt, drawing in bond investors, and supplemented by sovereign vehicles such as Abu Dhabi's MGX, asset managers like BlackRock, and dedicated infrastructure partnerships. Long-horizon institutional capital is being pulled into the build-out as a direct financier, not merely a public-market shareholder.
The bull and bear cases, fairly stated
A sober asset owner holds both possibilities at once.
The bull case is that AI demand is real and accelerating, that compute is the scarce input of the next economic era, and that building capacity now is a strategic necessity. In this view, today's capex is the foundation of a productivity boom, and the firms that build aggressively will dominate. The infrastructure — power and data centres especially — retains value regardless of which specific models or applications win.
The bear case is that the spending is running ahead of the revenue. AI capex is increasingly funded by debt and is visibly depressing the free cash flow of companies once prized for generating it. The monetisation of AI — the actual revenue that justifies hundreds of billions in annual investment — has not yet caught up to the scale of the build-out. If demand disappoints, the result could be overcapacity, write-downs and a sharp repricing of the mega-caps that dominate the indices. History offers cautionary parallels in the telecom and fibre build-out of the late 1990s, where genuine technological demand still produced enormous capital destruction along the way.
The honest position is that both scenarios are plausible, and the supercycle will likely deliver elements of each.
How long-term portfolios should think about it
For most large asset owners, the first point is uncomfortable but important: you already own this trade. Holding global equities means holding the mega-cap AI builders at index weight, so the supercycle is embedded in the portfolio whether or not it was chosen deliberately. The starting task is therefore to measure that exposure honestly rather than assume diversification protects against it.
Beyond awareness, the most de