I’m writing this dripping in sweat caused by France’s fifth heatwave of the past four months. The fifth.
And yet clearly, big tech gives even less shits about that than ever.
It was revealed this week that Amazon will run a forthcoming AI data centre in Texas with its very own dedicated methane gas power station. Said facility, called GW Ranch, will emit 33 million tonnes of CO2 per year, making it the single biggest source of pollution in the USA.
If you’re wondering what 33 million tonnes of CO2 represents… it’s almost as much as Holland’s grid (visual via Ketan Joshi).
This means that, rather than disgusting others – including the neighbours whose air it poisons, nearly doubling nitrous oxide levels around it – Elon Musk’s monstruous methane-powered Colossus facility has inspired his fellow tech billionaire bros, Jeff Bezos first and foremost.
Indeed, just before Cleanview’s GW Ranch scoop, Amazon’s founder announced a new $200m fund, with Leonardo Di Caprio, to save rare animal species such as the Golden-Crowned Sifaka. This means he’s paying millions to save the planet, whilst investing billions to kill it. $12 billion in this case, according to Forbes, i.e. sixty times more.
If only this – and Colossus – were isolated incidents. They’re not. Data centre operators’ new love affair with behind-the-meter (BHM), or off-grid power sources, is booming, re. Cleanview (again, below):
82 gas-powered data centre projects are currently being built across the US right now, according to the New York Times; whence the above graph.
Should all of the gas capacity expected to have come online in 2025 actually come online, the USA’s emissions could increase by nearly 28%, says SAIG’s Boris Gamazaychikov…
And if the above graph isn’t worrying enough, spare a thought for Microsoft’s recent deal with Chevron, which has been tied to a $7bn gas power plant in… Texas (bless ’em…)
Why is this happening? Because new AI compute capacity ‘has to be built right now’ – to semi-quote big tech – and gas is the quickest energy source to deploy (Musk got Colossus running in just four months). Renewable sources like solar or wind only provide intermittent power (is their excuse, because they could always set up batteries to store this power, as indeed Google is already doing); and nuclear power stations take 8-10 years to build (yes, even Microsoft’s much-publicised acquisition of Three Mile Island). So much for the planet, then.
It could even be argued that big tech sees BHM as a solution to the tsunami of popular opposition to data centres sweeping the USA right now – $130bn in projects stopped in Q1 2026 alone – because taking power off-grid guarantees local electricity bills won’t go up when a new data centre comes online. Indeed, GW Ranch got the go-ahead despite (because of?) Texas’ recent moratorium on data centres, which was imposed in response to growing bipartisan data centre protests.
The backdrop: record big tech hypocrisy
Big tech’s “no shits given” stance was confirmed recently when the hyperscalers revealed that their emissions went up by 42% 2020-25 (and by even more if you count from 2019, except Microsoft has hidden its figures), in their latest annual Sustainability Reports. This despite their public pledges to be carbon neutral, zero or negative by 2030 or 2040.
Said Sustainability Reports were also brimming with world-class blame-shifting, given that these emissions increases are blatantly caused by GAFAM’s record-breakingly huge AI investments:
🗣️ “Many of the sustainability solutions are not scaling fast enough to keep pace with AI infrastructure growth” – Microsoft‘s chief sustainability officer (CSO) Melanie Nakagawa
🗣️ “Our AI infrastructure buildout is currently accelerating faster than the grid is decarbonizing” – Google 2026 Sustainability Report intro
🗣️ “The demand [for AI] may slow us down” in terms of the company’s climate ambitions – Kara H. Hurst, Amazon‘s CSO, admitted in her company’s report, adding that “we continue working to sustainably meet our customers’ demands for AI.”
Blaming consumer/client demand for a planet-damaging product is tactic taken straight from the big oil playbook (BP invented the notion of individual carbon footprints with a similar blame-shifting objective).
And yet big tech/AI’s hypocrisy continues.
Anthropic is still seen by many as the most responsible AI lab… AND YET it is investing $15bn a year in Musk’s aforementioned Colossus monstrosity, thereby turbocharging its rampant pollution. Based on the share of Anthropic’s total compute spend this represents, we estimate that one in three Claude prompts are currently going through Colossus.
Google, for its part, is still considered the least planet-killing of the three hyperscalers (in that some of its sustainability efforts are effective) AND YET it is currently paying Musk close to a billion a month for Colossus AI compute. Though we understand the latter example is temporary, both suggest big AI will have no future qualms whatsoever about pouring money into the 80+ new BHM-powered data centres planned in the USA.
“Yeah, but AI’s carbon footprint is tiny compared with planes’, so who cares?” I hear the AI boosters say.
No.
Such arguments notably ignore AI’s indirect impacts, or rebound effects. Which Enabled Emissions’ Holly and Will Alpine have just confirmed are HUGE. AI is currently being deployed by the oil and gas sector, for example to help it find more fossil fuels more quickly. This industrial productivity boost leads to enabled emissions 3-13 times greater than the global emissions of data centres today, and up to eight times greater than emissions projected for 2035, affirms the duo’s new white paper. “AI [enabled fossil fuel] productivity gains… help sustain fossil supply economics and increase total fossil consumption,” states the paper, “leading to an increase in net global emissions.”
TL;DR: Both in terms of enabled emissions and BHM, AI is currently giving oil and gas a new lease of life, when on the contrary, fossil fuel usage must stop outright if we are to have any chance of slowing global warming (cf. multiple IPCC reports of past decades).
Finally, figures aside, AI is already forcing us into major environmental dilemmas. And not just in the USA.
- In Marseille, France‘s second city and home to a rapidly-growing number of data centres, a previous mayor had to ask operator Digital Realty if they could spare him a bit of electricity for the city’s buses (source). The same city’s plan to electrify its port is currently on hold for similar reasons (source)
- In West London, UK, plans to create more housing are on hold because planned data centres get first dibs on local electricity (source)
- And in Queretaro, Mexico, locals living near hyperscaler data centres get water for an hour or two a week so that these ever-bigger facilities can get it 24/7 (source: Ana Valdivia, University of Oxford).
The worst/best of all? None of the above impacts are necessary.
AI doesn’t need to be this big
Firstly, we can tweak AI so it consumes less energy at the inference stage (when users ask models questions). “Peer-reviewed research in Joule found optimized inference systems already use 4–20× less energy per query than widely cited estimates, and that further improvements across model design, serving systems, and hardware could reduce per-query energy use by 8–20×”, said sustainable AI stars GreenPT and Neuralwatt in a joint statement yesterday (above visual extracted from that; more info here).
Secondly, we can move to stop AI waste. Just a few examples:
- Two thirds of Dutch data centres are currently empty, re. Leitmotiv (source)
- One third of the 3 million Blackwell GPUs sold by NVIDIA are actually in use, re. Ed Zitron
- 18% of the 2GW of compute power that Microsoft said would be online in 2026 is actually online (idem)
- 14% of the GPU compute capacity purchased by one company was actually being used, re. Mark Butcher
- Half of the GPUs currently online worldwide – and therefore half the new data centres – would be needed if we used them 3x better and 3x longer, according to the seminal “From FLOPs to Footprints” white paper.
Thirdly, and critically, we can start using smaller models. ChatGPT is around 200 times bigger than most users need it to be. This means it consumes considerable resources (energy, water, hardware) for nothing. But smaller, less impactful models are harder to find, and use. This is why I created the Frugal AI Switchboard (below), which I strongly encourage you to check out, use and share. For AI’s six most common enterprise uses/needs, it suggests small, open source alternatives to the big usual suspects. A crucial first step in your quest to #QuitBigAI!
There are countless other ways to make AI less impactful, including by choosing data centres in countries with low-carbon energy, like France or Sweden; using more efficient chips, like Google’s TPUs; or running smaller models offline, on-device. Many of these tips come with huge savings in terms of money and confidentiality too. You’ll find out more about those elsewhere on this blog, and in my Frugal AI training course, of course.
All of the above points assume you do need (generative) AI, too; you don’t always. I often cite the example of startup consultancy Terra Cognita, whose client, a delivery company, asked them to create a chatbot to handle client and colleague requests for data extraction. Terra Cognita established that an old-fashioned dashboard could do just as good a job, without AI, saving the company a lot of money, and an estimated 150 kg of CO2e per month in emissions.
I could go on, but I think you get the point. Big AI’s impacts are currently exploding – when they should be doing the opposite – at a time they simply don’t need to. As the excellent Cory Doctorow put it recently:
If Donald Trump ordered Big Tech to turn off all of your country’s chatbots tomorrow, nothing would change. Every one of your country’s ministries and corporations would chug on with nary a hitch. Households, too, though perhaps a few of the younger members of those families would have to do their own homework again.
Contrast this with what would transpire if Trump directed his tech giants to switch off your country’s Office 365 access, or to brick your Android and iOS phones, or to killswitch your John Deere tractors. Your country would effectively cease to exist. If “digital sovereignty” means anything, it means doing something about this urgent fact.)
Thank you for reading. I look forward to your comments and shares as ever. Now I’m off to jump in a bucket of ice water… 🥵
Featured image: © Kristin Charleton via Unsplash
