In the high desert east of Reno, Nevada, construction crews are flattening the golden foothills of the Virginia Range, laying the foundations of a data center city. Google, Tract, Switch, EdgeCore, Novva, Vantage, and PowerHouse are all operating, building, or expanding huge facilities within the Tahoe Reno Industrial Center, a business park bigger than the… When we set out to write a story on the best available estimates for AI’s energy and emissions burden, we knew there would be caveats and uncertainties to these numbers. But, we quickly discovered, the caveats are the story too.  This story is a part of MIT Technology Review’s series “Power Hungry: AI and our energy future,” on the energy demands and carbon costs of the artificial-intelligence revolution. Measuring the energy used by an AI model is not like evaluating a car’s fuel economy or an appliance’s energy rating. There’s no agreed-upon method or public database of values. There are no regulators who enforce standards, and consumers don’t get the chance to evaluate one model against another.  Despite the fact that billions of dollars are being poured into reshaping energy infrastructure around the needs of AI, no one has settled on a way to quantify AI’s energy usage. Worse, companies are generally unwilling to disclose their own piece of the puzzle. There are also limitations to estimating the emissions associated with that energy demand, because the grid hosts a complicated, ever-changing mix of energy sources.  It’s a big mess, basically. So, that said, here are the many variables, assumptions, and caveats that we used to calculate the consequences of an AI query. (You can see the full results of our investigation here.) Measuring the energy a model uses Companies like OpenAI, dealing in “closed-source” models, generally offer access to their  systems through an interface where you input a question and receive an answer. What happens in between—which data center in the world processes your request, the energy it takes to do so, and the carbon intensity of the energy sources used—remains a secret, knowable only to the companies. There are few incentives for them to release this information, and so far, most have not. That’s why, for our analysis, we looked at open-source models. They serve as a very imperfect proxy but the best one we have. (OpenAI, Microsoft, and Google declined to share specifics on how much energy their closed-source models use.)  The best resources for measuring the energy consumption of open-source AI models are AI Energy Score, ML.Energy, and MLPerf Power. The team behind ML.Energy assisted us with our text and image model calculations, and the team behind AI Energy Score helped with our video model calculations. Text models AI models use up energy in two phases: when they initially learn from vast amounts of data, called training, and when they respond to queries, called inference. When ChatGPT was launched a few years ago, training was the focus, as tech companies raced to keep up and build ever-bigger models. But now, inference is where the most energy is used. The most accurate way to understand how much energy an AI model uses in the inference stage is to directly measure the amount of electricity used by the server handling the request. Servers contain all sorts of components—powerful chips called GPUs that do the bulk of the computing, other chips called CPUs, fans to keep everything cool, and more. Researchers typically measure the amount of power the GPU draws and estimate the rest (more on this shortly).  To do this, we turned to PhD candidate Jae-Won Chung and associate professor Mosharaf Chowdhury at the University of Michigan, who lead the ML.Energy project. Once we collected figures for different models’ GPU energy use from their team, we had to estimate how much energy is used for other processes, like cooling. We examined research literature, including a 2024 paper from Microsoft, to understand how much of a server’s total energy demand GPUs are responsible for. It turns out to be about half. So we took the team’s GPU energy estimate and doubled it to get a sense of total energy demands.  The ML.Energy team uses a batch of 500 prompts from a larger dataset to test models. The hardware is kept the same throughout; the GPU is a popular Nvidia chip called the H100. We decided to focus on models of three sizes from the Meta Llama family: small (8 billion parameters), medium (70 billion), and large (405 billion). We also identified a selection of prompts to test. We compared these with the averages for the entire batch of 500 prompts.  Image models Stable Diffusion 3 from Stability AI is one of the most commonly used open-source image-generating models, so we made it our focus. Though we tested multiple sizes of the text-based Meta Llama model, we focused on one of the most popular sizes of Stable Diffusion 3, with 2 billion parameters.  The team uses a dataset of example prompts to test a model’s energy requirements. Though the energy used by large language models is determined partially by the prompt, this isn’t true for diffusion models. Diffusion models can be programmed to go through a prescribed number of “denoising steps” when they generate an image or video, with each step being an iteration of the algorithm that adds more detail to the image. For a given step count and model, all images generated have the same energy footprint. The more steps, the higher quality the end result—but the more energy used. Numbers of steps vary by model and application, but 25 is pretty common, and that’s what we used for our standard quality. For higher quality, we used 50 steps.  We mentioned that GPUs are usually responsible for about half of the energy demands of large language model requests. There is not sufficient research to know how this changes for diffusion models that generate images and videos. In the absence of a better estimate, and after consulting with researchers, we opted to stick with this 50% rule of thumb for images and videos too. Video models Chung and Chowdhury do test video models, but only ones that generate short, low-quality GIFs. We don’t think the videos these models produce mirror the fidelity of the AI-generated video that many people are used to seeing.  Instead, we turned to Sasha Luccioni, the AI and climate lead at Hugging Face, who directs the AI Energy Score project. She measures the energy used by the GPU during AI requests. We chose two versions of the CogVideoX model to test: an older, lower-quality version and a newer, higher-quality one.  We asked Luccioni to use her tool, called Code Carbon, to test both and measure the results of a batch of video prompts we selected, using the same hardware as our text and image tests to keep as many variables as possible the same. She reported the GPU energy demands, which we again doubled to estimate total energy demands.  Tracing where that energy comes from After we understand how much energy it takes to respond to a query, we can translate that into the total emissions impact. Doing so requires looking at the power grid from which data centers draw their electricity.  Nailing down the climate impact of the grid can be complicated, because it’s both interconnected and incredibly local. Imagine the grid as a system of connected canals and pools of water. Power plants add water to the canals, and electricity users, or loads, siphon it out. In the US, grid interconnections stretch all the way across the country. So, in a way, we’re all connected, but we can also break the grid up into its component pieces to get a sense for how energy sources vary across the country.  Understanding carbon intensity The key metric to understand here is called carbon intensity, which is basically a measure of how many grams of carbon dioxide pollution are released for every kilowatt-hour of electricity that’s produced.  To get carbon intensity figures, we reached out to Electricity Maps, a Danish startup company that gathers data on grids around the world. The team collects information from sources including governments and utilities and uses them to publish historical and real-time estimates of the carbon intensity of the grid. You can find more about their methodology here.  The company shared with us historical data from 2024, both for the entire US and for a few key balancing authorities (more on this in a moment). After discussions with Electricity Maps founder Olivier Corradi and other experts, we made a few decisions about which figures we would use in our calculations.  One way to measure carbon intensity is to simply look at all the power plants that are operating on the grid, add up the pollution they’re producing at the moment, and divide that total by the electricity they’re producing. But that doesn’t account for the emissions that are associated with building and tearing down power plants, which can be significant. So we chose to use carbon

The thousands of sprawling acres in rural northeast Louisiana had gone unwanted for nearly two decades. Louisiana authorities bought the land in Richland Parish in 2006 to promote economic development in one of the poorest regions in the state. For years, they marketed the former agricultural fields as the Franklin Farm mega site, first to… The thousands of sprawling acres in rural northeast Louisiana had gone unwanted for nearly two decades. Louisiana authorities bought the land in Richland Parish in 2006 to promote economic development in one of the poorest regions in the state. For years, they marketed the former agricultural fields as the Franklin Farm mega site, first to auto manufacturers (no takers) and after that to other industries that might want to occupy more than a thousand acres just off the interstate. This story is a part of MIT Technology Review’s series “Power Hungry: AI and our energy future,” on the energy demands and carbon costs of the artificial-intelligence revolution. So it’s no wonder that state and local politicians were exuberant when Meta showed up. In December, the company announced plans to build a massive $10 billion data center for training its artificial-intelligence models at the site, with operations to begin in 2028. “A game changer,” declared Governor Jeff Landry, citing 5,000 construction jobs and 500 jobs at the data center that are expected to be created and calling it the largest private capital investment in the state’s history. From a rural backwater to the heart of the booming AI revolution! The AI data center also promises to transform the state’s energy future. Stretching in length for more than a mile, it will be Meta’s largest in the world, and it will have an enormous appetite for electricity, requiring two gigawatts for computation alone (the electricity for cooling and other building needs will add to that). When it’s up and running, it will be the equivalent of suddenly adding a decent-size city to the region’s grid—one that never sleeps and needs a steady, uninterrupted flow of electricity. To power the data center, Entergy aims to spend $3.2 billion to build three large natural-gas power plants with a total capacity of 2.3 gigawatts and upgrade the grid to accommodate the huge jump in anticipated demand. In its filing to the state’s power regulatory agency, Entergy acknowledged that natural-gas plants “emit significant amounts of CO2” but said the energy source was the only affordable choice given the need to quickly meet the 24-7 electricity demand from the huge data center. Meta said it will work with Entergy to eventually bring online at least 1.5 gigawatts of new renewables, including solar, but that it had not yet decided which specific projects to fund or when those investments will be made. Meanwhile, the new natural-gas plants, which are scheduled to be up and running starting in 2028 and will have a typical lifetime of around 30 years, will further lock in the state’s commitment to the fossil fuel. The development has sparked interest from the US Congress; last week, Sheldon Whitehouse, the ranking member of the Senate Committee on Environment and Public Works issued a letter to Meta that called out the company’s plan to power its data center with “new and unabated natural gas generation” and said its promises to offset the resulting emissions “by funding carbon capture and a solar project are vague and offer little reassurance.” The choice of natural gas as the go-to solution to meet the growing demand for power from AI is not unique to Louisiana. The fossil fuel is already the country’s chief source of electricity generation, and large natural-gas plants are being built around the country to feed electricity to new and planned AI data centers. While some climate advocates have hoped that cleaner renewable power would soon overtake it, the booming power demand from data centers is all but wiping out any prospect that the US will wean itself off natural gas anytime soon. The reality on the ground is that natural gas is “the default” to meet the exploding power demand from AI data centers, says David Victor, a political scientist at the University of California, San Diego, and co-director of its Deep Decarbonization Project. “The natural-gas plant is the thing that you know how to build, you know what it’s going to cost (more or less), and you know how to scale it and get it approved,” says Victor. “Even for [AI] companies that want to have low emissions profiles and who are big pushers of low or zero carbon, they won’t have a choice but to use gas.” The preference for natural gas is particularly pronounced in the American South, where plans for multiple large gas-fired plants are in the works in states such as Virginia, North Carolina, South Carolina, and Georgia. Utilities in those states alone are planning some 20 gigawatts of new natural-gas power plants over the next 15 years, according to a recent report. And much of the new demand—particularly in Virginia, South Carolina and Georgia—is coming from data centers; in those 3 states data centers account for around 65 to 85% of projected load growth. “It’s a long-term commitment in absolutely the wrong direction,” says Greg Buppert, a senior attorney at the Southern Environmental Law Center in Charlottesville, Virginia. If all the proposed gas plants get built in the South over the next 15 years, he says, “we’ll just have to accept that we won’t meet emissions reduction goals.” But even as it looks more and more likely that natural gas will remain a sizable part of our energy future, questions abound over just what its continued dominance will look like. For one thing, no one is sure exactly how much electricity AI data centers will need in the future and how large an appetite companies will have for natural gas. Demand for AI could fizzle. Or AI companies could make a concerted effort to shift to renewable energy or nuclear power. Such possibilities mean that the US could be on a path to overbuild natural-gas capacity, which would leave regions saddled with unneeded and polluting fossil-fuel dinosaurs—and residents footing soaring electricity bills to pay off today’s investments. The good news is that such risks could likely be managed over the next few years, if—and it’s a big if—AI companies are more transparent about how flexible they can be in their seemingly insatiable energy demands. The reign of natural gas Natural gas in the US is cheap and abundant these days. Two decades ago, huge reserves were found in shale deposits scattered across the country. In 2008, as fracking started to make it possible to extract large quantities of the gas from shale, natural gas was selling for $13 per million Btu (a measure of thermal energy); last year, it averaged just $2.21, the lowest annual price (adjusting for inflation) ever reported, according to the US Energy Information Administration (EIA). Around 2016, natural gas overtook coal as the main fuel for electricity generation in the US. And today—despite the rapid rise of solar and wind power, and well-deserved enthusiasm for the falling price of such renewables—natural gas is still king, accounting for around 40% of electricity generated in the US. In Louisiana, which is also a big producer, that share is some 72%, according to a recent audit. Natural gas burns much cleaner than coal, producing roughly half as much carbon dioxide. In the early days of the gas revolution, many environmental activists and progressive politicians touted it as a valuable “bridge” to renewables and other sources of clean energy. And by some calculations, natural gas has fulfilled that promise. The power sector has been one of the few success stories in lowering US emissions, thanks to its use of natural gas as a replacement for coal.   But natural gas still produces a lot of carbon dioxide when it is burned in conventionally equipped power plants. And fracking causes local air and water pollution. Perhaps most worrisome, drilling and pipelines are releasing substantial amounts of methane, the main ingredient in natural gas, both accidentally and by intentional venting. Methane is a far more potent greenhouse gas than carbon dioxide, and the emissions are a growing concern to climate scientists, albeit one that’s difficult to quantify. Still, carbon emissions from the power sector will likely continue to drop as coal is further squeezed out and more renewables get built, according to the Rhodium Group, a research consultancy. But Rhodium also projects that if electricity demand from data centers remains high and natural-gas prices low, the fossil fuel will remain the dominant source of power generation at least through 2035 and the transition to cleaner electricity will be much delayed. Rhodium estimates that the continued reign of natural gas will lead to an additional 278 million metric tons of annual US carbon emissions by 2035 (roughly equivalent to the emissions from a large US state such as Florida), relative to a future in which the use of fossil fuel gradually winds

This is today’s edition of The Download, our weekday newsletter that provides a daily dose of what’s going on in the world of technology. We did the math on AI’s energy footprint. Here’s the story you haven’t heard. It’s well documented that AI is a power-hungry technology. But there has been far less reporting on the extent… This is today’s edition of The Download, our weekday newsletter that provides a daily dose of what’s going on in the world of technology. We did the math on AI’s energy footprint. Here’s the story you haven’t heard. It’s well documented that AI is a power-hungry technology. But there has been far less reporting on the extent of that hunger, how much its appetite is set to grow in the coming years, where that power will come from, and who will pay for it.  For the past six months, MIT Technology Review’s team of reporters and editors have worked to answer those questions. The result is an unprecedented look at the state of AI’s energy and resource usage, where it is now, where it is headed in the years to come, and why we have to get it right.  At the centerpiece of this package is an entirely novel line of reporting into the demands of inference—the way human beings interact with AI when we make text queries or ask AI to come up with new images or create videos. Experts say inference is set to eclipse the already massive amount of energy required to train new AI models. Here’s everything we found out. Here’s what you can expect from the rest of the package, including: + We were so startled by what we learned reporting this story that we also put together a brief on everything you need to know about estimating AI’s energy and emissions burden.  + We went out into the world to see the effects of this energy hunger—from the deserts of Nevada, where data centers in an industrial park the size of Detroit demand ever more water to keep their processors cool and running.  + In Louisiana, where Meta plans its largest-ever data center, we expose the dirty secret that will fuel its AI ambitions—along with those of many others.  + Why the clean energy promise of powering AI data centers with nuclear energy will long remain elusive.  + But it’s not all doom and gloom. Check out the reasons to be optimistic, and examine why future AI systems could be far less energy intensive than today’s. AI can do a better job of persuading people than we do The news: Millions of people argue with each other online every day, but remarkably few of them change someone’s mind. New research suggests that large language models (LLMs) might do a better job, especially when they’re given the ability to adapt their arguments using personal information about individuals. The finding suggests that AI could become a powerful tool for persuading people, for better or worse. The big picture: The findings are the latest in a growing body of research demonstrating LLMs’ powers of persuasion. The authors warn they show how AI tools can craft sophisticated, persuasive arguments if they have even minimal information about the humans they’re interacting with. Read the full story. —Rhiannon Williams How AI is introducing errors into courtrooms It’s been quite a couple weeks for stories about AI in the courtroom. You might have heard about the deceased victim of a road rage incident whose family created an AI avatar of him to show as an impact statement (possibly the first time this has been done in the US).But there’s a bigger, far more consequential controversy brewing, legal experts say. AI hallucinations are cropping up more and more in legal filings. And it’s starting to infuriate judges. Just consider these three cases, each of which gives a glimpse into what we can expect to see more of as lawyers embrace AI. Read the full story. —James O’Donnell This story originally appeared in The Algorithm, our weekly newsletter on AI. To get stories like this in your inbox first, sign up here. The must-reads I’ve combed the internet to find you today’s most fun/important/scary/fascinating stories about technology. 1 Donald Trump has signed the Take It Down Act into US lawIt criminalizes the distribution of non-consensual intimate images, including deepfakes. (The Verge)+ Tech platforms will be forced to remove such material within 48 hours of being notified. (CNN)+ It’s only the sixth bill he’s signed into law during his second term. (NBC News) 2 There’s now a buyer for 23andMe Pharma firm Regeneron has swooped in and offered to help it keep operating. (WSJ $)+ The worth of your genetic data? $17. (404 Media)+ Regeneron promised to prioritize security and ethical use of that data. (TechCrunch) 3 Microsoft is adding Elon Musk’s AI models to its cloud platformErr, is that a good idea? (Bloomberg $)+ Musk wants to sell Grok to other businesses. (The Information $) 4 Autonomous cars trained to react like humans cause fewer road injuriesA study found they were more cautious around cyclists, pedestrians and motorcyclists. (FT $)+ Waymo is expanding its robotaxi operations out of San Francisco. (Reuters)+ How Wayve’s driverless cars will meet one of their biggest challenges yet. (MIT Technology Review) 5 Hurricane season is on its wayDOGE cuts means we’re less prepared. (The Atlantic $)+ COP30 may be in crisis before it’s even begun. (New Scientist $) 6 Telegram handed over data from more than 20,000 users In the first three months of 2025 alone. (404 Media) 7 GM has stopped exporting cars to ChinaTrump’s tariffs have put an end to its export plans. (NYT $) 8 Blended meats are on the risePlants account for up to 70% of these new meats—and consumers love them. (WP $)+ Alternative meat could help the climate. Will anyone eat it? (MIT Technology Review) 9 SAG-AFTRA isn’t happy about Fornite’s AI-voiced Darth VaderIt’s slapped Fortnite’s creators with an unfair labor practice charge. (Ars Technica)+ How Meta and AI companies recruited striking actors to train AI. (MIT Technology Review) 10 This AI model can swiftly build Lego structuresThanks to nothing more than a prompt. (Fast Company $) Quote of the day “Platforms have no incentive or requirement to make sure what comes through the system is non-consensual intimate imagery.” —Becca Branum, deputy director of the Center for Democracy and Technology, says the new Take It Down Act could fuel censorship, Wired reports. One more thing Are friends electric?Thankfully, the difference between humans and machines in the real world is easy to discern, at least for now. While machines tend to excel at things adults find difficult—playing world-champion-level chess, say, or multiplying really big numbers—they find it hard to accomplish stuff a five-year-old can do with ease, such as catching a ball or walking around a room without bumping into things.This fundamental tension—what is hard for humans is easy for machines, and what’s hard for machines is easy for humans—is at the heart of three new books delving into our complex and often fraught relationship with robots, AI, and automation. They force us to reimagine the nature of everything from friendship and love to work, health care, and home life. Read the full story. —Bryan Gardiner We can still have nice things A place for comfort, fun and distraction to brighten up your day. (Got any ideas? Drop me a line or skeet ’em at me.) + Congratulations to William Goodge, who ran across Australia in just 35 days!+ A British horticulturist has created a garden at this year’s Chelsea Flower Show just for dogs.+ The Netherlands just loves a sidewalk garden.+ Did you know the T Rex is a north American hero? Me neither