AI Data Centers and the Growing Battle Over Electricity

The computing behind AI needs power on the scale of cities. That physical fact is shaping where data centres are built, what they cost and how communities respond.

Public opposition to data centres is growing in the United States, and the technology industry often explains it as general unease about big companies. Recent reporting from northern Virginia, published in MIT Technology Review's AI newsletter The Algorithm, argues for a simpler cause: electricity. This article summarises that reporting and draws out what it means for businesses that depend on AI services.

Why are data centres facing a backlash?

The reporting centres on Loudoun County, Virginia, often called Data Center Alley and described as the largest data centre hub in the world. Local campaigners quoted there reject the idea that the dispute is cultural. In their account it has always been about power: who pays for new transmission lines, what happens to the grid at peak demand, and whether extra demand is met with fossil fuels.

According to the report, protests have already produced state rules requiring data centres to pay the full cost of new grid connections, and developers expect further rules that could cut their grid supply at times of peak demand.

How much electricity does an AI data centre use?

The figures quoted in the reporting give a sense of scale:

• A typical data centre operating in that part of Virginia today has a capacity of about 54 megawatts.
• The smallest sites considered attractive for training AI models are around 300 megawatts.
• The facilities AI companies would ideally build are closer to 6,000 megawatts, which the report compares to the electricity demand of the city of Chicago.

Demand of that size is hard for an existing grid to supply, wherever the site is.

Why are new AI data centres generating their own power?

There are two ways to power a large data centre: connect to the grid or generate electricity on site. The reporting says most new AI facilities are choosing the second, in rural areas with little grid infrastructure, using natural gas turbines with diesel generators as backup.

On-site generation avoids straining the public grid, but it brings a power station to the neighbourhood, with the air pollution that gas and diesel produce. When a public utility builds on that scale, communities usually expect a say in where it goes. Private developers moving quickly have not always offered one.

One developer quoted in the report said the industry sees its plans as limited less by public opposition than by a shortage of electricity.

What does this mean for businesses that use AI?

Few businesses will ever build a data centre. Most will feel these pressures indirectly, through the services they buy.

1. Treat AI running costs as variable. If power is the limiting factor for providers, prices and capacity for AI services may not fall as steadily as many plans assume. Budget with a margin.
2. Use the smallest model that does the job. Routing simple tasks to smaller, cheaper models and keeping the largest for work that needs them reduces both cost and exposure to capacity limits.
3. Ask providers where and how workloads run. Region, energy source and capacity commitments are reasonable questions to put to a cloud or AI supplier, particularly if your business reports on sustainability.
4. Avoid dependence on one provider or region. Local rules on grid access are changing. Being able to move a workload is a practical form of resilience.

What happens next?

The reporting expects new sites to move towards places with fewer residents and easier access to gas, and notes that companies are also investing abroad. Wherever they are built, the constraint is the same: AI at scale is an energy business as much as a software one. Decisions about which AI tools to adopt are better made with that in mind.

This article summarises and comments on reporting published in MIT Technology Review's newsletter The Algorithm. Figures and quotations are as given in that reporting.