Google said it has struck a deal with Commonwealth Fusion Systems (CFS), a private spin-off of the Massachusetts Institute of Technology, to buy power from a project in Virginia that would be the search giant’s first commercial fusion purchase.

The deal, whose financial terms were not disclosed, comes at a time when tech companies are seeking innovative sources for power to fuel data centres for energy-hungry generative AI services.

Google said it plans to buy 200 megawatts of clean fusion power from a CFS project known as ARC that it describes as the first grid-scale power plant using fusion, which to date has not reached commercial viability.

Fusion plant

The plant is located in Chesterfield County, Virginia, in proximity to the world’s biggest cluster of data centres, and is planned to have a total capacity of 400 MW.

ARC is planned to go online in the early 2030s, but there are still scientific challenges to be overcome for the start-up’s fusion approach, which involves the use of large magnets to instigate the nuclear fusion of light atoms.

Michael Terrell, Google’s head of advanced energy, acknowledged to reporters on a call that the project faced “serious physics and engineering challenges” but called the investment a “longer-term bet”.

Google has invested in CFS since 2021, following the start-up’s spin-off in 2018, and said it was expanding its investment, without disclosing an amount.

CFS chief financial officer and co-founder Bob Mumgaard said the current funding undertaken by the start-up was comparable to the 2021 round of $1.8 billion (£1.3bn) by Google and others.

“We’re excited to make this longer-term bet on a technology with transformative potential to meet the world’s energy demand, and support CFS in their effort to reach their scientific and engineering milestones needed to get there,” Google’s Terrell said.

Power supply

Researchers have sought sustainable fusion energy for decades as a source of carbon-free energy using the reaction that powers the sun, but reactions have yet to deliver more power than the energy required to start the reaction.

Other large tech firms have turned to conventional nuclear energy for AI data centre power, with Meta Platforms last month signing a 20-year deal to buy power from a nuclear power plant in Illinois.

Since late last year Chinese start-up DeepSeek has been offering high-performance AI models that use far less power than those of US competitors, but its approach has so far not been emulated by American tech firms.