Ireland’s data center backlash offers a warning for the U.S.

Ireland was an early warning

Marx began following the issue after data center moratoriums in the Netherlands and Singapore, both introduced in 2019, and growing concern in Ireland about electricity use in 2022.

A case in Dalles, Oregon, also caught his attention. Google tried to keep its water-use figures secret for nearly a year. When residents finally obtained the data, they found the company accounted for 29% of the city’s water consumption.

Ireland offered a clearer preview of how constraints can redirect development. In 2021, Dublin imposed a moratorium after the grid operator said there were too many data centers to connect more. Operators responded by looking beyond Dublin, exploring connections to the gas network and using generators. Some began building dedicated gas power plants rather than relying on the shared grid.

The country’s electricity network is publicly owned, and only the state utility can build transmission lines. Data center companies pushed for a law that would effectively let private firms build their own cables, raising concerns about equal access to the grid.

Irish operators also secured emissions permits to burn fossil fuels. The environmental agency warned that doing so would make climate targets much harder to meet. Marx says the pattern reminds him of what is happening in the U.S., where companies are racing to build large private gas plants for data centers. What surprises him is the scale and speed.

What survives an AI bust

Marx thinks the AI boom is a bubble. The open question is how much of the infrastructure under construction will be finished before companies reassess demand. Turbines, chips and other components are already in short supply, he says.

The dot-com bust offers a partial precedent: fiber networks were overbuilt, and data centers were built too, though at a much smaller scale than today. Many companies went bankrupt. Equinix and other survivors bought data centers and were positioned when demand returned.

But today’s facilities are being designed around AI workloads and the assumption that chips will need replacing every few years. Even if the buildings remain useful, the equipment that gives them value could be obsolete by the time demand comes back.

The gas plants pose a separate risk. If data centers are never built, or demand collapses, owners will still need customers to justify those plants. Marx does not think anyone is building them with the intention of simply switching them off.

My read is that the infrastructure bet is riskier than the buildings alone suggest: a data center can be repurposed, but its power supply and specialized hardware may not age as well.

A different model of AI

Marx points to a community in New Zealand that developed a more accurate speech-recognition model for the Māori language and kept the data with local cloud providers. He presents it as an alternative to extracting resources from local communities to build AI elsewhere.

He argues that governments should set strict rules for data center construction and regulate AI products themselves. His broader criticism is that the promised benefits of generative AI have not materialized, while its costs were not adequately disclosed.

What I’d want to know is whether policymakers will assess the infrastructure against what AI delivers now, rather than the promise made in early 2023. If demand is revised down, the power plants could remain long after the rationale for building them has gone.