The defining physical story of AI in 2026 isn’t a chatbot or a model — it’s buildings. Enormous, power-hungry data centers are going up across the United States and beyond, and the sums involved are staggering. By widely reported estimates, the largest cloud providers and their partners have committed on the order of $690 billion in capital spending toward AI-ready data centers, with 2026 capex budgets at Amazon, Microsoft, Alphabet, and Meta each running into the hundreds of billions collectively.
The takeaway: AI has become a heavy industry. Running large models at scale requires vast fleets of chips, and those chips need buildings, cooling, and — above all — electricity. In 2026 the binding constraint stopped being money or even chips, and became power. That single fact explains most of the drama in this story, from canceled projects to deals with power companies.
The scale of the build-out
A few reference points make the scale concrete.
The combined 2026 capital-expenditure plans of the major hyperscalers are enormous: Amazon’s projected capex has been reported around $200 billion (not all data centers, but a large share), with Microsoft, Alphabet, and Meta each in the roughly $115–190 billion range. Together the biggest four spent an estimated $413 billion in 2025 and are projected to spend somewhere between $600 billion and $700 billion in 2026 — an increase that would have seemed implausible just two years ago.
The single largest named effort is the Stargate Project, a consortium including OpenAI, SoftBank, and Oracle, which has been described as a roughly $500 billion, multi-year plan to build around 10 gigawatts of US AI data-center capacity. Its flagship Texas campus began coming online through 2025 and 2026 with over a gigawatt of capacity and hundreds of thousands of Nvidia GPUs planned. Those chips are the reason Nvidia’s numbers look the way they do — see Nvidia in 2026: The AI Chip Giant's Big Year.
Why power became the bottleneck
The most important shift in 2026 is that electricity, not construction or silicon, now gates how fast AI infrastructure can grow.
The numbers explain why. US data-center power demand is projected to climb from roughly 31 gigawatts in 2025 toward the low 40s in 2026 and higher still by 2027. Globally, data-center electricity consumption is forecast to rise more than a quarter in 2026. A single large AI training facility can draw 100 to 500 megawatts continuously — comparable to a small city — and several gigawatt-scale campuses are expected to come online in 2026, each run by a different hyperscaler.
This demand is running straight into the limits of the power grid. Reporting through 2026 described a striking situation: a large share of US data centers planned for the year — on the order of half of announced capacity by some counts — have been delayed or canceled, largely because the power to run them isn’t available on schedule. One major cloud provider disclosed a multi-billion-dollar backlog of orders it couldn’t fulfill for lack of electricity. Demand, in other words, is outrunning even an aggressive build-out.
Who benefits, who’s squeezed
The build-out is reshaping several industries at once.
- Chipmakers. Nvidia above all, plus networking and memory suppliers, capture much of the spend.
- Utilities and power producers. Data centers are becoming among the largest new electricity customers in a generation, a windfall for those who can supply reliable power.
- Specialized cloud providers. Companies renting out GPU capacity have become significant players, some going public on the strength of it — a theme in our AI IPOs & Valuations in 2026 coverage.
- Construction and cooling. The physical build — electrical work, advanced cooling, land — is a booming sub-industry.
On the squeezed side are ordinary electricity customers in some regions, where surging data-center demand raises questions about grid capacity and prices, and communities weighing the local trade-offs of hosting these facilities. It’s a genuine tension, not just a growth story.
The money question
The uncomfortable part is that this spending is largely ahead of the revenue AI generates today. Companies are building capacity on the expectation that demand for AI will keep climbing steeply. If that demand grows as hoped, the infrastructure pays off. If it disappoints, some of these buildings could end up underused — a risk that sits at the heart of the wider debate in Is the AI Bubble Real? A Level-Headed Look.
There’s a reassuring counterpoint: even in a downturn, data centers, chips, and power connections retain value and get repurposed. Physical infrastructure doesn’t vanish the way a hyped app can. But the timing mismatch between today’s spending and tomorrow’s revenue is real, and worth holding in mind. For how the economics are meant to work, see How Tech Companies Actually Make Money From AI.
What to watch next
- Power deals. New agreements between data-center operators and power sources, including nuclear and dedicated generation, are the clearest sign of who can actually grow.
- Capex revisions. Whether the hyperscalers raise or trim their 2026–2027 budgets signals confidence in AI demand.
- Utilization. Reports of how fully these expensive facilities are actually used — the real test of whether the build-out was justified.
- Grid strain. How communities and regulators respond as data centers compete with households for electricity.
What it means for you
Most of this is invisible to everyday users — until it isn’t. These data centers are what make the AI tools on your phone and laptop fast and available. When capacity is tight, you may notice slower responses, usage limits, or paused sign-ups on popular AI services. Longer term, the efficiency of these facilities influences how cheaply AI can be offered to you. And if you pay an electricity bill in a data-center-heavy region, this build-out is quietly part of your world. For the tools all this infrastructure powers, browse our The AI Directory.
FAQ
How much are companies spending on AI data centers in 2026?
The largest cloud providers and their partners have collectively committed on the order of $690 billion toward AI data-center infrastructure, with the biggest four hyperscalers projected to spend roughly $600–700 billion in capital expenditure in 2026 alone — a steep rise from about $413 billion in 2025.
What is the Stargate Project?
Stargate is a large AI-infrastructure consortium including OpenAI, SoftBank, and Oracle, described as a multi-year effort worth around $500 billion to build roughly 10 gigawatts of US data-center capacity. Its flagship campus in Texas began coming online through 2025 and 2026 with over a gigawatt of capacity.
Why is power a problem for AI data centers?
AI facilities draw enormous, continuous electricity — a single large training site can use 100 to 500 megawatts, like a small city. US data-center power demand is rising fast, and the grid can’t always supply it on schedule, so a significant share of planned 2026 capacity has been delayed or canceled for lack of power.
How much electricity do data centers use?
US data-center power demand is projected to rise from roughly 31 gigawatts in 2025 toward the low 40s in 2026, with global consumption forecast to grow more than a quarter that year. Some projections see data centers consuming a rising share of total US electricity by the late 2020s.
Will the AI data center boom raise my electricity bill?
Possibly, in regions with heavy data-center concentration, where surging demand can strain local grids and pressure prices. The picture varies widely by location and depends on how utilities add generation and how regulators allocate costs, so it’s an area to watch rather than a universal outcome.
Is the data center build-out a bubble?
Spending is running ahead of the revenue AI earns today, which is a genuine risk if demand disappoints. But unlike a hyped app, data centers, chips, and power connections keep their value and can be repurposed. It’s better seen as a high-stakes bet on future demand than a guaranteed return.
Zen Tech Hub may earn a commission from links on this page, at no extra cost to you.
