The scale of Big Tech's energy commitments for AI has entered territory that strains normal comprehension. Meta alone has signed nuclear deals totalling more than 6 gigawatts — enough to power approximately 5 million homes — through a patchwork of agreements with Constellation Energy, TerraPower, and Oklo. That's one company's power portfolio for one set of AI data centres. Multiply it across Google, Microsoft, Amazon, and OpenAI, and you begin to understand why California is now reconsidering a nuclear moratorium it has held for 50 years.
Data centres are on a trajectory to consume around 1,600 terawatt-hours globally by 2035 — roughly 4.4% of all electricity used worldwide, which would make them the fourth-largest electricity consumer on Earth if counted as a country. The grid was not built for this. Natural gas turbines capable of meeting the demand are largely sold out through the end of the decade. Advanced nuclear power, the technology most often cited as the long-term solution, won't reach commercial scale until the 2030s at the earliest. There is a meaningful gap between the power AI infrastructure needs and the power available to provide it, and that gap is now driving energy policy in ways that have nothing to do with climate goals.
California's reconsideration of nuclear energy, reported by Bloomberg in early March, is a striking example. The state banned new nuclear plants in 1976 and has long positioned itself as a clean energy leader on renewable terms. But AI demand is now straining the grid in ways that solar and wind alone can't address on the required timescale, and policymakers are confronting a genuine tension between decades-old nuclear scepticism and the practical reality of powering a technology sector that generates enormous economic value for the state.
The question CNBC raised in March — who is really footing the AI energy bill — is perhaps the sharpest angle on all of this. The large tech companies signing multi-gigawatt nuclear deals are, in many cases, negotiating agreements that connect their data centres directly to dedicated generation capacity, bypassing the traditional utility model. But not all of the grid costs can be privatised that way. Transmission infrastructure, grid balancing, the capital cost of rapidly expanding capacity — some portion of these costs flows through to electricity ratepayers, including households that have nothing to do with AI. Ratepayer advocates in several states have begun pushing back, arguing that the public shouldn't subsidise the infrastructure buildout for a technology whose commercial benefits largely accrue to shareholders.
Scientific American adds a further wrinkle: the rapid expansion of nuclear capacity that tech companies are pursuing — including small modular reactors and advanced designs — will generate nuclear waste on a scale the US doesn't yet have infrastructure to manage. The existing waste storage system is already under strain, and the regulatory and political path to new permanent storage sites remains deeply uncertain. Big Tech's nuclear ambitions may be solving one problem while quietly accelerating another.
None of this means AI's energy demands are unmanageable. But the current trajectory involves costs — to grids, to ratepayers, to waste management — that aren't fully visible in the announcements about gigawatt deals and nuclear partnerships. The tech companies are moving fast, as they do; the infrastructure required to support them is moving at a different pace entirely.