Six Years of AI Energy Data: What the Numbers Actually Say

The International Energy Agency released its 2028 AI Energy Consumption Report in March, and for the first time in the six years since they started tracking this sector seriously, there’s enough longitudinal data to say something definitive. Not optimistic. Not pessimistic. Definitive.

Here’s the number that matters: AI-related electricity consumption reached 1,100 TWh globally in 2028. That’s roughly 4% of total global electricity production. In 2022, before the generative AI explosion, the same figure was around 200 TWh — so we’re looking at roughly a 5.5x increase in six years. For context, the entire country of Germany uses about 570 TWh per year. The AI industry now consumes twice that.

These numbers come from both the IEA’s World Energy Outlook and the DOE’s 2028 Data Center Energy Report, which cross-referenced utility company data with data center construction permits and hyperscaler disclosure requirements that the EU mandated starting in 2026. Before those mandates, we were largely guessing. After them, we started getting real numbers — and the real numbers are high.

What’s less discussed is the shape of that consumption. Training large frontier models accounts for roughly 15% of AI energy use. The other 85% is inference — the billions of daily queries, image generations, code completions, and automated pipeline processes that run continuously. This matters enormously for how we think about reducing AI’s footprint, because the interventions look completely different depending on which part of the problem you’re trying to solve.

The narrative around AI energy has been polluted by motivated reasoning from both ends. Technology optimists point to efficiency gains — and they’re real. A query processed by a modern inference-optimized chip uses roughly 40% less energy than the same query would have used on 2023-era hardware. Model distillation techniques have brought capable smaller models into deployment at a fraction of the compute cost of their predecessors. The energy per useful output has genuinely improved.

But the total consumption went up by a factor of five.

This is the efficiency paradox that anyone who has studied energy economics recognizes immediately. The Jevons paradox, named after the 19th-century British economist William Stanley Jevons, describes how improvements in resource efficiency tend to increase rather than decrease total resource consumption — because lower costs enable higher adoption. The steam engine got more efficient; Britain burned more coal. Cars got better fuel economy; people drove more miles. AI got more efficient per query; the world ran far more queries.

The AI industry has spent considerable effort emphasizing per-unit efficiency improvements while being somewhat quieter about the total trajectory. The two statistics live in the same universe but tell different stories. A company can truthfully say their 2028 data center is 60% more energy efficient than their 2023 one while also consuming three times more total power.

The geographic distribution of that consumption is where things get politically complicated. The United States hosts approximately 38% of global AI compute infrastructure, followed by China at 29%, with the EU, Japan, Singapore, and a cluster of emerging data center markets splitting the remainder. This distribution is shifting — Southeast Asia and the Middle East have seen enormous data center investment since 2026, partly because land is cheaper, partly because power purchase agreements are more flexible, and partly because regulatory environments have been more permissive.

What this means for carbon accounting is significant. A data center in Texas running on a grid mix of 40% renewables has a very different carbon footprint than nominally equivalent compute in a coal-heavy grid. The IEA estimates that the same AI workload produces anywhere from 50 to 400 grams of CO₂ equivalent per kWh of electricity, depending purely on where the servers are located. Companies that route workloads to renewable-heavy regions genuinely do reduce their climate impact. Companies that buy renewable energy certificates from wind farms in Idaho to offset coal-powered servers in Ohio are doing something that looks the same in annual sustainability reports but is not the same in atmospheric terms.

The DOE’s 2028 report is admirably blunt about this: “Carbon accounting methodologies that allow renewable energy certificate matching without temporal and locational correspondence systematically understate the grid emissions attributable to data center operations.”

What about the benefits side of the ledger? This is where honest accounting gets genuinely difficult, and where I find myself more sympathetic to the complexity than most critics allow.

AI is being used to optimize electricity grid operations in ways that measurably reduce waste. The grid balancing systems deployed by several European transmission operators use ML models that have demonstrably cut curtailment of renewable generation — essentially, the problem of throwing away clean power because supply and demand don’t match in real time. The estimated emissions savings from these applications globally are somewhere between 80 and 200 TWh equivalent per year, depending on your assumptions about counterfactual grid operation. That’s real.

Industrial process optimization is another area where the claimed benefits are at least partially verifiable. Cement production, steel manufacturing, and chemical processing are collectively responsible for about 22% of global industrial emissions. AI-driven process optimization in these sectors has reduced energy intensity meaningfully in facilities where it’s been deployed at scale. The numbers are harder to aggregate globally because implementation varies so widely, but the directional effect is not in dispute.

What is in dispute is whether these benefits are additional — would they have happened anyway through other means? And whether the counterfactual is fair — comparing AI-enabled optimization against the baseline of no optimization, rather than against the baseline of what would have happened with conventional optimization techniques. These are legitimate methodological questions that the more credulous AI sustainability reports sidestep.

The honest answer, drawing on six years of actual data rather than projections, looks something like this: AI consumed roughly 4.8 billion metric tons of CO₂ equivalent over the 2023–2028 period, accounting for both direct electricity consumption and the embodied carbon of hardware manufacturing. Against this, the verifiable emissions reductions attributable to AI applications are somewhere in the range of 0.8 to 2.4 billion metric tons, depending on attribution methodology. The range is wide because attribution is genuinely hard.

This means the AI sector is, on net, currently a contributor to global emissions rather than a mitigant. The most optimistic credible reading has the benefits closing the gap to something like a 2:1 ratio of cost to benefit. The pessimistic credible reading has it at 6:1. Neither reading supports the industry framing of AI as an inherently green technology, and neither supports the most alarmist claims that AI is a straightforward climate catastrophe.

What it supports is the conclusion that AI’s environmental impact is a serious policy problem that requires actual solutions rather than accounting tricks — and that six years of data are finally enough to have that conversation without both sides hiding behind uncertainty.

The uncertainty was always partially convenient. Now we know enough to stop hiding behind it.