UN agency warns AI data centres are outpacing power grids

What happened

The UN Economic Commission for Europe, known as UNECE and based in Geneva, issued the warning about data centres, AI facilities and other very large electricity users. Its main point is a mismatch in time. A large facility can often be built and connected to the grid in two to five years. Expanding transmission lines and other grid equipment can take more than ten years, because of planning, approval and construction. Demand can therefore arrive long before the network is ready for it.

The numbers come from the International Energy Agency (IEA). Global data-centre electricity use was about 485 terawatt-hours in 2025 and is projected to reach about 950 terawatt-hours by 2030, roughly 3 percent of world electricity demand. Spending on data-centre infrastructure is forecast to rise from around 800 billion US dollars a year in 2026 to 1.8 trillion dollars a year by 2050. Both UN News and the legal news site JURIST reported these figures.

UNECE also named examples of countries that have already acted. Ireland, which has one of the highest concentrations of data centres, restricted new grid connections in Dublin. The Netherlands limited where data centres may be built. UNECE said there is no consistent way to divide the cost of grid upgrades between data-centre developers, power companies and ordinary consumers, and warned that this uncertainty may delay investment. It called for stronger regulation and coordination to keep energy systems reliable.

The engineering behind it

A data centre is a building full of servers, storage and network equipment that runs day and night. Almost all the electricity it uses becomes heat, so a large share of extra power goes to cooling. AI servers, which pack many graphics processors into each rack, draw much more power per rack than ordinary servers. That is why a single AI facility can need as much power as a small town, and why it needs a strong connection to the high-voltage network.

UNECE pointed to a second problem besides total energy: how quickly demand changes. According to UN News, AI workloads can rise and fall suddenly, and grids that depend heavily on renewable energy cannot always adjust in real time. Overloaded systems can suffer voltage oscillations, unintended disconnections and cascading failures, where one failure trips the next. UN News said this is already happening in some extreme cases. In general, a grid must keep supply and demand balanced every second, or voltage and frequency drift away from safe values.

Regulators also lack information. UNECE said they do not get enough real-time data about how large facilities use power. Facilities tend to cluster where connections and rules are favourable, so one local grid may face several large new loads at once. Rules on water use, emissions and local resources are spreading, but UNECE described them as fragmented. JURIST added that the IEA suggests managing project pipelines in advance, making power systems more flexible, and using AI itself to improve energy security.

Flexibility is a general engineering idea with several common forms. A data centre can store energy in batteries and draw on them during peaks. It can also shift work in time, because training an AI model can often wait a few hours while answering users cannot. Grid operators can pay large users to cut demand when the network is under stress, which is called demand response. Each of these needs measurement, control software and an agreement between the facility and the power company.

What it means in Nepal

Nepal's government wants to build exactly this kind of facility. The budget for fiscal year 2026/27, presented on 29 May 2026 by Finance Minister Swarnim Wagle, included a Sovereign AI Compute Center in Syuchatar, Kathmandu. According to Nepal News, the plan is to run thousands of AI processors on clean hydropower and offer computing to Nepali startups at subsidised rates. The paper reported that no exact processor count or government cost figure had been given.

The same Nepal News report, published on 1 July 2026, collected doubts that match the UNECE warning. AI developer Kshitiz Rimal said the grid would struggle to meet the demand. AI scientist Suresh Manandhar said power must be stable, because even a one-second outage disrupts services. Cybersecurity expert Rajiv Subba said a medium-sized data centre uses about as much power as 100,000 households and about 1.14 million litres of water a day for cooling. Others noted that current rules allow only smaller data centres.

Nepal News put existing and planned data-centre capacity in the country at about 15 to 20 megawatts, including a 3.5 megawatt WorldLink facility in Matatirtha. Any much larger AI centre would add a new, concentrated load to the Kathmandu grid. The UNECE paper shows the questions engineers would need to answer first: whether local substations and lines can carry the load, how to keep voltage stable when demand jumps, how to cool the servers, and who pays for the upgrades.