The 2030 AI Power Crunch

Japan just approved the restart of Kashiwazaki-Kariwa, the world's largest nuclear plant, after 15 years of dormancy. The reason? AI data centres. Meanwhile in Texas, the Stargate project chose a different path: a £394m ($500m, €465m) on-site natural gas plant with 360 MW of turbines.

This is not a technology debate. It is a race. And gas is winning.

Reuters reports 114 GW of new gas-fired capacity is now in the US development pipeline, more than double the level from a year earlier. Data centres account for almost half of the country's forecasted rise in power demand. Every gas plant built today operates for 20 to 30 years. That is fossil fuel infrastructure locked in until 2055.

Our analysis reveals why the next three years determine whether nuclear or gas powers the AI revolution.

Goldman Sachs projects US data centre power demand will reach 84 GW by 2027, with AI growing to 27% of the overall market. The IEA projects global data centre electricity consumption will double to 945 TWh by 2030. That represents more than Japan's entire electricity consumption today.

GPU clusters now demand 80 kW per rack or higher, more than double conventional data centre density. Stargate's planned 1.2 GW Abilene campus alone requires the equivalent of a large nuclear reactor.

Whilst hyperscalers announce nuclear partnerships, Business Insider reports utilities in states where Amazon, Microsoft, and Meta are building giant data centre hubs want to power them with new natural gas plants.

The examples are multiplying:

MIT Technology Review summarised the uncomfortable truth: "Even as tech companies tout plans for nuclear power, they'll actually be relying largely on fossil fuels, keeping coal plants open, and even building new natural gas plants that could stay open for decades."

Nuclear restarts face a brutal maths problem. Palisades, backed by £1.2bn ($1.52bn, €1.4bn) in DOE loan guarantees plus over £1bn ($1.3bn, €1.2bn) from USDA, just pushed its restart to early 2026. Three Mile Island's £1.3bn ($1.6bn, €1.5bn) restart won't deliver power until 2028.

The US interconnection queue tells an even starker story. Lawrence Berkeley National Laboratory reports 2,300 GW of generation and storage capacity seeking grid connection as of end 2024. Gas turbines bypass this entirely with on-site generation.

China already operates 1.2 GW of nuclear-data centre co-location across CNNC's Qinshan, Yangjiang, and Tianwan sites. State planning eliminates the queue delays, permitting challenges, and financing uncertainty that constrain Western projects.

Japan expects energy demand to rise over the coming decade due to AI data centres despite its shrinking population. TEPCO plans to restart Kashiwazaki-Kariwa in January 2026, and the region is already planning data centre developments nearby.

France's EDF launched Project Giga to power 1 GW data centres using nuclear generation. With 56 operational reactors, France offers what the US cannot: existing clean baseload without new construction.

The winning strategy emerging from our analysis: behind-the-meter nuclear co-location at existing operational plants. This bypasses interconnection queues, eliminates permitting delays for new generation, and delivers power in 18 to 24 months rather than 5 to 10 years.

But the window is closing. Every Stargate-style gas plant that breaks ground is a 30-year commitment to fossil fuel infrastructure. Every month of nuclear delay hands market share to gas.

UAE's Barakah, with 4 GW of operational nuclear capacity, represents the largest co-location opportunity globally. No gas plants needed. No carbon capture retrofits required. Clean baseload ready today.

The 2030 power crunch is not a forecast. It is a race. Nuclear must beat gas to the grid, or we lock in fossil fuel infrastructure until 2055. The £79bn ($100bn, €93bn) arbitrage opportunity belongs to investors who identify which operational nuclear facilities can displace gas fastest.

Next week: We dive into the IAEA data: What IAEA PRIS Data Doesn't Tell You…