AI diffusion is imposing sweeping harmful impacts on critical resources and the planetary biosphere, on a much greater scale than has been previously recognized. That’s the conclusion of our latest piece on the global energy and emissions footprint of AI, which we just released this morning. This was a labor of love that tooks months of writing and research, and I want to thank my co-author @Jesse Damiani for the phenomenal collaboration. I think we’ve created something special here and I’m proud of the final outcome. Nonetheless, we emphasize in the piece that our results should be seen as a first-order approximation, not a definitive final answer, especially because we used a very macro analysis that relies on many assumptions and eschews detailed microlevel inputs, but also because Big Tech itself deliberately hides a lot of critical data that would allow researchers and the public to get a better sense of AI’s ecological footprint.

The specifics can be found in the post, but here are the main headlines and conclusions:

We estimate that the AI industry will be responsible for at least 1.1% of global energy consumption and for 1.2% of all global GHG emissions this year. These are higher figures than most conventional estimates because we use the most recent and updated data for 2026, because we define AI as a tech industry that encompasses way more than just LLMs like ChatGPT, because we adopt a lifecycle analysis that goes far beyond electricity consumption and tracks primary energy footprints and all major upstream and downstream impacts of AI diffusion, and because we explore the impact of AI through new sectors and methods that have been ignored or marginalized in prior studies (for example, we have a first-of-its-kind global assessment for this sector on the emissions associated with land use changes from data centers and other facilities supporting the AI infrastructure boom). In contrast to the message of sustainability that many leaders in Silicon Valley are promoting, we demonstrate that AI is a major global industry that’s already imposing historic consequences for the planetary biosphere. Too often, conversations about the harmful ecological impacts of AI are framed as something that’s going to happen in the future. In reality, AI is already reshaping the global energy and emissions landscape right now. It’s almost certainly the dominant driver of new global energy consumption and emissions, and if it continues at this rate and the corresponding financial bubble keeps inflating until 2030, the AI industry will rival the ecological footprint of steel and cement, the world’s dominant industrial polluters.

Our piece is comprehensive but also imperfect in many ways, and we acknowledge the shortcomings and limitations that come with this kind of analysis. In many places, we leaned on the side of caution and conservative interpretations. For example, we excluded the upstream energy footprint of the roughly 80 GW of natural gas power plants in the US that are being built to support data center expansion, thus in effect significantly understating our case. But in other places, our analysis includes macroeconomic rebound effects from AI diffusion that we know are going to be controversial, especially because they haven’t been quantified at a global level like this before. And yet, even in the case of rebounds and the Jevons Paradox, we used conservative assumptions that we believe are still understating our case (for example, we excluded by definition any potential rebounds or backfire effects outside the primary energy footprint of electricity consumption). Nevertheless, we are fully aware that there’s enough content in this piece to spark a million debates that will last for years over gross abatement factors, silicon wafer electrical intensities, and land use emission footprints, to name just a few. That’s the main reason we wrote it. Not because we thought the conclusions we reached are ironclad and irreversible (we ourselves acknowledge that our analysis has weaknesses), but because we hoped to spark a paradigm shift in how people talk about the energy footprint of AI. And our hope is that by sparking such a shift we may eventually arrive at better public policy that can establish much stronger controls and regulations over the AI industry, even though the current political moment makes anything like that fairly improbable in the short run.

I wanted to close on a few notes of context and perspective. The issues involved here are very deep, complex, and usually don’t have perfect solutions. How should we measure the Scope 3 emissions of an industry? There is no universal agreement. Walmart reports four times the carbon emissions footprint of Amazon. If you didn’t know any better, you would think Walmart is so much worse for the planet than a peer competitor which in every other way is roughly the same size! But the real reason for the gaping discrepancy is that Amazon has adopted a very restrictive (and many would argue useless) accounting standard for Scope 3 emissions: it simply excludes the entire upstream manufacturing footprint from its third-party vendors, even though over 60% of all e-commerce sales on the Amazon platform come from independent sellers. By contrast, Walmart includes the manufacturing footprint of major third-party sellers on its carbon ledger. As a result, Walmart looks way worse than Amazon purely because of radically different accounting standards, not because it’s actually that much worse for the planet. These same kinds of issues are relevant for this piece, especially the Scope 3 analysis of the downstream impacts in the final section. But the broader point is that one can adopt different standards, conventions, and assumptions and arrive at very different answers. We obviously think that our analysis is a substantial enhancement over the common assumptions in the dominant discourse on AI, but that doesn’t mean that our assumptions are perfect or can’t be questioned. They certainly can be, and we actually questioned them ourselves numerous times throughout the piece.

This analysis is partly aimed at experts and researchers, but since we’re obviously posting this on social media, it’s also written in a way that can be accessible for the general public, at least to some extent. There are deep yet balanced theoretical discussions in here that should be of interest to everyone. But we don’t want to overpromise in this regard either. Unit conversions and equations are generally omitted from the written text, but they are always there behind almost every conclusion. If you’re hoping to find the specific number of 167,000 TWh as the 2025 primary global energy consumption in The Energy Institute’s Statistical Review World of Energy, well you won’t. EI gives that figure in exajoules, so what you’ll find instead is 600 exajoules, and only by converting that to TWh will you get the numbers that are shown in our piece. For experts this is all trivial and obvious stuff, but if this world is new to you and you’re trying to follow along with our conclusions and sources, these kinds of nuances could prove to be somewhat annoying barriers, though hopefully only temporary ones.

Beyond this piece, I’ll be doing a whole series of media appearances on AI in the coming weeks and months, where I’ll be talking to people about many aspects of the AI story that weren’t in focus for this post, like the economic and political impacts related to jobs, the stock market, geostrategic rivalries, AGI and ASI, and so much more. AI is obviously a huge topic right now, and I want to play my small part in driving the public conversation forward and towards a more productive direction.