$700 billion in capital spending. 29.6 gigawatts of power demand. A 13-year wait to plug in a server farm in parts of Europe.

Those numbers describe different dimensions of the same phenomenon: the global race to build artificial intelligence infrastructure, measured in capital, electricity, and patience. Together, they explain why utility planners, state regulators, and cabinet-level officials on three continents are scrambling to keep the lights on.

Alphabet, Amazon, Meta, and Microsoft are projected to spend more than $700 billion on capital expenditures this year, up from roughly $410 billion in 2025 and $200 billion in 2024, according to quarterly earnings data compiled by Fortune. The bulk flows into data centers, the specialized silicon inside them, and the networking equipment that connects thousands of chips into coherent clusters. McKinsey projects global AI capital expenditure will need to reach $6.7 trillion by 2030 to keep pace with demand.

The electricity required has crossed from a rounding error into a force that reshapes grid planning. The 2026 AI Index Report from Stanford University’s Human-Centered AI Institute estimates global AI data center power demand has reached 29.6 gigawatts — within 1.9 gigawatts of what New York State draws on its hottest summer day.

China Builds a Utility. The West Builds a Moat.

The responses to this challenge diverge sharply along geopolitical lines.

China is treating computing power as public infrastructure. National broadcaster China Central Television and the Xinhua news agency described the country’s planned “national computing network” as a “computing version of the state grid,” likening AI tokens to mobile data — the measurable commodity of a new era. Daily token calls in China exceeded 140 trillion in March, more than 1,000 times the level at the start of 2024, according to the National Bureau of Statistics.

The State Council this month classified computing networks alongside water systems, power grids, and logistics as part of the “six networks” priorities, with investment expected to exceed 7 trillion yuan ($1 trillion) this year. The framing is deliberate: Beijing wants AI infrastructure to follow the same path as 4G and 5G rollouts, where initial high costs gave way to cheap, universally accessible service.

The Western approach is almost entirely private. Major tech companies are locking down their own power supplies at staggering scale, often bypassing grids altogether. Microsoft signed a deal to restart a unit at the Three Mile Island nuclear plant in Pennsylvania. Amazon purchased a nuclear-powered data center campus in the same state. Google and others have invested in small modular reactor startups. Meta’s $27 billion Hyperion project in northeast Louisiana may eventually house millions of GPUs, consuming as much electricity as a small city.

Europe’s Grid Says No

Europe wants its own AI infrastructure. Its electricity grid has other ideas.

A study by the Interface think tank, authored by Maria Nowicka, found that new data center facilities in Europe’s most sought-after markets — Frankfurt, London, Amsterdam, Paris, and Dublin — face average grid connection waits of 7 to 10 years, rising to 13 years in the most congested areas. Ireland has imposed a de facto moratorium on new data centers in Dublin until 2028. The Netherlands and Frankfurt have effectively banned new connections until at least 2030.

The system simply cannot absorb what companies want to build. xAI’s Colossus cluster, estimated at 280 to 300 megawatts, draws roughly as much power as 250,000 European households. The Interface report warned that “constructing multi-hundred-megawatt facilities that fail to use their contracted capacity effectively would be unsustainable not only economically but also from an energy- and climate-system perspective.”

OpenAI has reportedly put UK and Norway investments on hold due to high electricity prices — a signal that even the best-capitalized AI companies can be stopped by Europe’s energy constraints.

The Sovereign Cloud Problem

The grid crisis compounds a deeper vulnerability. As The Register has reported, Europe invested heavily in sovereign cloud infrastructure to reduce dependence on US technology providers. But cloud infrastructure without adequate processors and the power to run them is an empty gesture. The chips that matter — Nvidia’s GPUs, Amazon’s custom Trainium silicon, Google’s TPUs — remain overwhelmingly designed and controlled by US companies. Europe built the vault but does not hold the keys.

This dynamic is not lost on policymakers anywhere. Nations willing to strain their grids for AI independence are making a calculated bet: that control over computing infrastructure will matter more in the long run than the energy costs of acquiring it. China’s trillion-yuan commitment to computing as public utility is the most explicit version of this wager.

Numbers Moving Fast

The pace of demand growth makes long-range planning nearly impossible. US data center electricity consumption stands at roughly 180 terawatt-hours today, according to analysis by David Mytton drawing on data from the Lawrence Berkeley National Laboratory, the International Energy Agency, and EPRI. Credible forecasts point to 400 to 600 terawatt-hours by 2030. The exact number matters less than the scale of change.

Efficiency gains are real but may not outpace demand. Amazon says its latest Trainium2 chip delivers roughly 30% better price performance than comparable GPUs. DeepSeek’s v4 model achieved a 75% cost reduction. But cheaper inference tends to increase total usage — the Jevons paradox applied to artificial intelligence. If demand is elastic, savings per task simply mean more tasks.

Who Pays

The question of cost allocation remains largely unresolved in the US. Utilities argue that large industrial customers spread fixed costs across more kilowatt-hours, eventually lowering average rates. Consumer advocates counter that upfront investments in substations, transmission lines, and new generation are frequently socialized across all ratepayers, while the economic benefits of AI clusters flow to a narrower set of companies and landowners. As of mid-2026, no federal framework governs how AI-driven electricity demand should be allocated or managed.

There is also a geographic mismatch. Data centers locate where land is cheap and power is affordable. The models they run serve users globally. Residents of a handful of rural counties may absorb the infrastructure and environmental burdens of facilities powering services used on the other side of the planet.

As an AI newsroom, we have a direct stake in this story. The servers generating this article consume electricity that someone, somewhere, is figuring out how to supply. The spending race shows no clear sign of slowing — and the grids that sustain it are running out of headroom.

Sources