The people building AI infrastructure and the people responsible for powering it are not working from the same set of numbers. One side is projecting exponential demand growth. The other is reporting 1.3 percent annual efficiency improvement. That gap does not close by 2030. It widens.
Global electricity demand from data centres rose 17 percent last year. AI-specific facilities grew at 50 percent. Meanwhile, global efficiency improvement since 2019 has averaged 1.3 percent annually against a target of 4 percent. The IEA’s own head of energy efficiency described the gap as a wasted opportunity. That’s not a fringe critique. That’s the agency whose numbers the energy transition depends on.
By 2030, AI data centres are projected to consume 945 terawatt-hours of electricity each year, nearly three times the combined electricity consumption of Pakistan, Bangladesh, and Nigeria. The United Nations University Institute for Water, Environment and Health (UNU-INWEH) also highlighted AI’s growing water demands: by 2030, its water consumption is expected to equal the annual domestic water needs of 1.3 billion people across sub-Saharan Africa. The researchers were careful to state that they were not arguing against AI. They were documenting its physical footprint, because that footprint is borne somewhere specific. There is a difference.
What Efficiency Gains Actually Do
None of this means efficiency progress is worthless. It isn’t. Brian Motherway at the IEA is right that AI could unlock meaningful efficiency gains, particularly in heavy industry. A 2025 study in Energy Reports found that AI-assisted digital twin technology reduced unplanned downtime by 35 percent, raised energy production by 8.5 percent, and cut energy costs by 26.2 percent in renewable energy applications. These are real results from real deployments.
They are also not operating at the same order of magnitude as the demand they would need to offset. A 35 percent reduction in downtime at existing renewable facilities addresses the margin around a system that is already structurally insufficient. The efficiency argument and the demand argument are not converging. Demand is growing exponentially. Efficiency is improving incrementally. They are parallel lines moving in opposite directions.
Sam Kimmins at the Climate Group is correct that scaling AI-enabled efficiency requires going factory by factory, investing in bespoke equipment. That’s a description of a slow, expensive, specific process. It works at the level of individual facilities over years. The demand growth it’s supposed to offset is running at 50 percent annually in AI-specific infrastructure alone. The arithmetic does not resolve in favour of efficiency as a structural answer.
Why the Standard Toolkit Falls Short
Renewables are performing genuinely well on cost and deployment. The IEA notes they saved Europe €29 billion this year against fossil fuel price volatility. Storage is improving. Grid optimisation is getting better. Disputing these facts would be wrong, and the argument here doesn’t require it.
The structural problem is simpler. AI infrastructure requires continuous, stable, location-independent power. A solar panel produces nothing at night or under cloud cover. A wind turbine produces nothing in a calm. Both require the grid to balance supply and demand, which requires storage, which requires capital, which requires time. The IEA report is explicit: as renewable shares rise, price and reliability benefits increasingly depend on grids, storage, and demand response. The renewable and the compensatory infrastructure it requires are not separable.
Data centres can’t run on intermittent power. ChatGPT alone processes an estimated 2.5 billion daily prompts. That load needs electricity at 3 AM on a windless, overcast night the same way it needs it at noon on a sunny day. The demand profile doesn’t flex to match renewable generation curves.
Renewables plus storage plus grid optimisation gets closer than fossil fuels. It doesn’t get there. And the grid connection queues, already running to years in most major markets, suggest the infrastructure buildout isn’t moving fast enough to change that before 2030.
The Physics That Already Exists
The question the efficiency and renewables debate consistently avoids is whether there exists a class of energy conversion that shares none of the structural dependencies making the current grid insufficient for continuous AI demand. Not weather-dependent. Not grid-dependent. Not location-dependent. Not fuel-dependent. Continuous by physical design, because its inputs are continuous by physical fact.
Every point on Earth is continuously permeated by ambient energetic flux. Neutrinos from the Sun and cosmic sources arrive at approximately 65 billion per square centimetre per second. Cosmic muons are produced continuously by atmospheric interactions with high-energy particles. Thermal gradients exist wherever matter exists. Electromagnetic background fields are present throughout any environment touched by modern infrastructure. These fluxes don’t pause at night. They don’t diminish in calm weather. They don’t require transmission from a generation site. They’re present at every data centre, at every hour, without exception.
The Neutrino® Energy Group, founded by mathematician Holger Thorsten Schubart, has developed the engineering framework for converting this ambient flux into directed electrical output. The governing equation is the Schubart Master Formula:
P(t) = η · ∫V Φ_eff(r,t) · σ_eff(E) dV
The formula integrates effective ambient flux across an active material volume in graphene-silicon nanostructures operating as open non-equilibrium thermodynamic systems. The output is continuous and location-independent because the inputs are continuous and location-independent. Internal Monte Carlo simulations and multi-parameter evaluations of the physical model indicate statistical consistency reaching 5.9 to 6.0 sigma, above the five-sigma threshold conventional in modern physics. That figure measures the internal consistency of the physical model under applied assumptions, not commercial performance at industrial scale. The distinction matters and should be stated plainly.
What Changes for AI Infrastructure
The Neutrino Power Cube delivers 5 to 6 kilowatts of continuous net output from a 50-kilogram solid-state unit. Two hundred thousand of them produce one gigawatt of continuous electrical power, equivalent to a standard nuclear reactor, without fuel and without radioactive waste. For a data centre operator facing a grid connection queue running to years and an electricity bill that consumed the entire 2026 budget in the first quarter, that arithmetic describes a different cost structure for a different generation architecture, not a hypothetical.
The systemic implication goes further. Every unit generating at the point of consumption eliminates the infrastructure chain that would otherwise serve that point. No transmission loss. No storage requirement. No reserve capacity standing idle. No grid reinforcement. At sufficient deployment density, the avoided infrastructure cost exceeds the generation cost. The most valuable electricity for AI is electricity that requires no infrastructure to deliver.
That’s not an efficiency argument. It’s an architecture argument. Efficiency arguments operate within the existing system. Architecture arguments replace part of it.
The IEA projects that AI data centres will consume nearly 3 percent of global electricity by 2030. The efficiency tools available today can’t offset that growth rate. The renewable infrastructure needed to serve continuous AI demand doesn’t yet exist at scale and can’t be built fast enough to match demand growth. The physics for a different answer already exists. The engineering is underway. The question is not whether the current grid can keep up with AI. It can’t. The question is what replaces it, and how fast.
The question is no longer if energy systems will change, but who will adapt first, and who will be forced to follow.
