Three years ago, as the pandemic caused chaos for companies big and small, Colgate-Palmolive’s chief supply chain officer Luciano Sieber orchestrated a “logistics blitz.”  The result gave Sieber a better understanding of how Colgate-Palmolive moves its products around the world. But it stuck Sieber with another problem: too much data.  About a year ago, Sieber says he found a solution to that problem with Uber Freight. The ride-hailing service’s long-running logistics and analytics arm has been developing new ways to wrangle large amounts of data by using artificial intelligence. Colgate-Palmolive became one of the first companies to use one of its newest products, a a logistics-focused LLM Uber Freight calls Inisights AI.  Now, Uber Freight is more formally launching a suite of AI features to shippers around the world as part of its existing supply chain software. That includes an expansion of Insights AI, which Uber Freight quietly launched in 2023, as well as more than 30 AI agents built to “execute key logistics tasks throughout the freight lifecycle.”  Uber Freight is not alone in trying to tame unruly supply chains with modern artificial intelligence tools. Flexport announced its own suite of AI tools in February, and there are myriad startups trying to help companies wrangle data, reduce inventory stockpiles, and better predict supply and demand. But Uber Freight is betting its AI solutions can make an immediate impact on the bottom line of both its blue-chip customers and the nearly 10,000 other shippers it works with. That’s largely because of the knowledge base and relationships it has established in the eight years since it was created to match long-haul truckers with shippers. “Supply chain is inherently a data-rich problem. It’s complex, it’s nuanced, and AI can serve a fundamental role in shaping it and accelerating it,” Uber Freight founder Lior Ron said in an interview with TechCrunch.  Techcrunch event Join us at TechCrunch Sessions: AI Secure your spot for our leading AI industry event with speakers from OpenAI, Anthropic, and Cohere. For a limited time, tickets are just $292 for an entire day of expert talks, workshops, and potent networking. Exhibit at TechCrunch Sessions: AI Secure your spot at TC Sessions: AI and show 1,200+ decision-makers what you’ve built — without the big spend. Available through May 9 or while tables last. Berkeley, CA | June 5 REGISTER NOW ‘We’ve been building towards this moment’ Uber Freight began as a more straightforward brokerage business model when it launched in 2017. But the Uber subsidiary has steadily evolved over the years into more of a service provider to companies that ship goods around the world.  Many modern companies are trying to find ways to incorporate artificial intelligence (often to mixed results); it should come as no surprise that Uber Freight is putting the technology front and center. After all, both Ron’s undergraduate work and his master thesis were centered around AI – way back “in the dark ages when it was called ‘neural networks,’” he joked. Ron continued to work with machine learning technology when he was running Google Maps from 2007 until 2016. It was there, he said, that he saw “the potential of digitizing the physical universe.”  “That sort of led me to the foundational belief, nine years ago, that supply chain is fundamentally a data-first, technology-first challenge that could be accelerated with data connectivity, and over time, AI,” he said. “We’ve been building towards this moment, I think, since I started Uber Freight.” Ron said Uber Freight has used machine learning in its work since the beginning. But it was around two years ago that the team started trying to work with more advanced generative AI capabilities.  That “hasn’t been an easy road,” Ron said. Uber Freight’s initial attempts at building a sort of “co-pilot for logistics” were riddled with hallucinations and returned accurate answers only around 60% to 70% of the time. Now that technology has been “battle tested” and is “driving real business outcomes,” with an accuracy rate of 98%, according to Ron. The company says the Insights AI model has been trained on internal and external data related to the $20 billion worth of freight that it helps move every year. It also leverages multiple undisclosed AI models “providing optimal combinations of price, precision and performance,” according to Uber Freight. Ron said this AI push creates new ways for customers to work with the data related to their supply chain. They can ask Insights AI to quickly pull up, say, the worst-performing origin points for particular shipments. Or they can ask to be shown “all shipments to CVS in 2023.” Ron stressed that the queries can be far more complex than this, too, and the model always keeps up.  Insights AI is presented to customers much like other popular LLM interfaces; it will also show its work and make clear where all the data is coming from, just like other reasoning models. All of this lets a customer “gain insights on your network much faster, at close to 100% accuracy instantly, versus formulating what you want to know, sending it to some analysts, and waiting for two weeks for the PowerPoint presentation to come back to have a discussion,” Ron said. ‘What do you want to know?‘ Uber Freight works with a lot of Fortune 500 companies, but it found a particularly willing partner in Colgate-Palmolive to trial Insights AI and its other new tools. The conglomerate already makes a suite of AI models available to all of its employees, according to Sieber. It also makes those workers take a mandatory training on AI ethics that was developed in-house. “I think it’s great, because it turns the conversation from fear into, ‘how that makes me more efficient, and how [do] I become a better professional and deliver more by having access and using those new technologies,’” Sieber said.  For instance, Sieber said his company has used Insights AI to easily identify carriers who are accepting fewer shipments than they’re contractually obligated to to move. From there, they can work out why those levels are low, and either come up with a solution to get the carrier back in compliance or drop them in favor of another. This was previously a challenge to solve in real time, Sieber said, because companies like Colgate-Palmolive work with thousands of carriers. Each of those might work with different systems and workflows, and all of that resulting information was never really centrally managed.  The next step with AI, both Sieber and Ron said, has been finding ways to create more proactive solutions. Ron said this is another place Uber Freight can flex its data strengths. “We know the facilities, we know the lanes, we know the prices,” he said. “What do you want to know?” These more proactive integrations come in the form of alerts that tell a customer like Colgate-Palmolive they’re overpaying on certain routes, or that there are faster options available for a particular shipment.  Any single suggestion like that may only save a few hundred, or perhaps a few thousand dollars. But aggregated over an entire network, it could make a big difference. That’s why, when asked, Sieber was quick to answer that Colgate-Palmolive’s chief financial officer is the executive who’s most pleased with what Uber Freight’s enabled. “He loves to see logistics costs coming down,” Sieber laughed.

Popular battle royale game Fortnite has finally returned to the U.S. App Store amid game maker Epic Games’ lengthy legal skirmish with Apple. As of Tuesday, Fortnite is also available on the Epic Games Store and AltStore in the EU. It’ll show up in App Store searches soon, Epic said in a post on X. Apple removed Fortnite from the App Store in 2020 after Epic added an option for people to make in-game purchases using Epic’s virtual currency, for which the company offered a 20% discount. Epic then went on the offensive and sued Apple. A judge ruled in Epic’s favor, finding that Apple was in “willful violation” of rules that bar anticompetitive pricing. Last month, that judge said Apple must allow developers to collect payments outside the App Store and not levy fees on them. Following the ruling, Epic resubmitted Fortnite to the App Store — but faced an approval delay. The company accused Apple of blocking the game from going live in the U.S. and EU. Earlier this week, U.S. District Judge Yvonne Gonzalez Rogers said in a filing that Apple should approve Epic’s submission or show it had a legal basis to block it. That possibly tilted things in Epic’s favor. Image Credits: EPic Games The new version of Fortnite available on the App Store also offers a 20% cashback for buying in-game items directly from Epic instead of using Apple’s in-app purchase method. Techcrunch event Join us at TechCrunch Sessions: AI Secure your spot for our leading AI industry event with speakers from OpenAI, Anthropic, and Cohere. For a limited time, tickets are just $292 for an entire day of expert talks, workshops, and potent networking. Exhibit at TechCrunch Sessions: AI Secure your spot at TC Sessions: AI and show 1,200+ decision-makers what you’ve built — without the big spend. Available through May 9 or while tables last. Berkeley, CA | June 5 REGISTER NOW Topics

With seemingly no limit to the demand for artificial intelligence, everyone in the energy, AI, and climate fields is justifiably worried. Will there be enough clean electricity to power AI and enough water to cool the data centers that support this technology? These are important questions with serious implications for communities, the economy, and the environment.  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. But the question about AI’s energy usage portends even bigger issues about what we need to do in addressing climate change for the next several decades. If we can’t work out how to handle this, we won’t be able to handle broader electrification of the economy, and the climate risks we face will increase. Innovation in IT got us to this point. Graphics processing units (GPUs) that power the computing behind AI have fallen in cost by 99% since 2006. There was similar concern about the energy use of data centers in the early 2010s, with wild projections of growth in electricity demand. But gains in computing power and energy efficiency not only proved these projections wrong but enabled a 550% increase in global computing capability from 2010 to 2018 with only minimal increases in energy use.  In the late 2010s, however, the trends that had saved us began to break. As the accuracy of AI models dramatically improved, the electricity needed for data centers also started increasing faster; they now account for 4.4% of total demand, up  from 1.9% in 2018. Data centers consume more than 10% of the electricity supply in six US states. In Virginia, which has emerged as a hub of data center activity, that figure is 25%. Projections about the future demand for energy to power AI are uncertain and range widely, but in one study, Lawrence Berkeley National Laboratory estimated that data centers could represent 6% to 12% of total US electricity use by 2028. Communities and companies will notice this type of rapid growth in electricity demand. It will put pressure on energy prices and on ecosystems. The projections have resulted in calls to build lots of new fossil-fired power plants or bring older ones out of retirement. In many parts of the US, the demand will likely result in a surge of natural-gas-powered plants. It’s a daunting situation. Yet when we zoom out, the projected electricity use from AI is still pretty small. The US generated about 4,300 billion kilowatt-hours last year. We’ll likely need another 1,000 billion to 1,200 billion or more in the next decade—a 24% to 29% increase. Almost half the additional electricity demand will be from electrified vehicles. Another 30% is expected to be from electrified technologies in buildings and industry. Innovation in vehicle and building electrification also advanced in the last decade, and this shift will be good news for the climate, for communities, and for energy costs. The remaining 22% of new electricity demand is estimated to come from AI and data centers. While it represents a smaller piece of the pie, it’s the most urgent one. Because of their rapid growth and geographic concentration, data centers are the electrification challenge we face right now—the small stuff we have to figure out before we’re able to do the big stuff like vehicles and buildings. We also need to understand what the energy consumption and carbon emissions associated with AI are buying us. While the impacts from producing semiconductors and powering AI data centers are important, they are likely small compared with the positive or negative effects AI may have on applications such as the electricity grid, the transportation system, buildings and factories, or consumer behavior. Companies could use AI to develop new materials or batteries that would better integrate renewable energy into the grid. But they could also use AI to make it easier to find more fossil fuels. The claims about potential benefits for the climate are exciting, but they need to be continuously verified and will need support to be realized. This isn’t the first time we’ve faced challenges coping with growth in electricity demand. In the 1960s, US electricity demand was growing at more than 7% per year. In the 1970s that growth was nearly 5%, and in the 1980s and 1990s it was more than 2% per year. Then, starting in 2005, we basically had a decade and a half of flat electricity growth. Most projections for the next decade put our expected growth in electricity demand at around 2% again—but this time we’ll have to do things differently.  To manage these new energy demands, we need a “Grid New Deal” that leverages public and private capital to rebuild the electricity system for AI with enough capacity and intelligence for decarbonization. New clean energy supplies, investment in transmission and distribution, and strategies for virtual demand management can cut emissions, lower prices, and increase resilience. Data centers bringing clean electricity and distribution system upgrades could be given a fast lane to connect to the grid. Infrastructure banks could fund new transmission lines or pay to upgrade existing ones. Direct investment or tax incentives could encourage clean computing standards, workforce development in the clean energy sector, and open data transparency from data center operators about their energy use so that communities can understand and measure the impacts. In 2022, the White House released a Blueprint for an AI Bill of Rights that provided principles to protect the public’s rights, opportunities, and access to critical resources from being restricted by AI systems. To the AI Bill of Rights, we humbly offer a climate amendment, because ethical AI must be climate-safe AI. It’s a starting point to ensure that the growth of AI works for everyone—that it doesn’t raise people’s energy bills, adds more clean power to the grid than it uses, increases investment in the power system’s infrastructure, and benefits communities while driving innovation. By grounding the conversation about AI and energy in context about what is needed to tackle climate change, we can deliver better outcomes for communities, ecosystems, and the economy. The growth of electricity demand for AI and data centers is a test case for how society will respond to the demands and challenges of broader electrification. If we get this wrong, the likelihood of meeting our climate targets will be extremely low. This is what we mean when we say the energy and climate impacts from data centers are small, but they are also huge. Costa Samaras is the Trustee Professor of Civil and Environmental Engineering and director of the Scott Institute for Energy Innovation at Carnegie Mellon University. Emma Strubell is the Raj Reddy Assistant Professor in the Language Technologies Institute in the School of Computer Science at Carnegie Mellon University. Ramayya Krishnan is dean of the Heinz College of Information Systems and Public Policy and the William W. and Ruth F. Cooper Professor of Management Science and Information Systems at Carnegie Mellon University.

Google co-founder Sergey Brin said he “made a lot of mistakes with Google Glass” during an onstage interview at Google I/O 2025 on Tuesday. Brin was a surprise addition to an interview with Google DeepMind CEO Demis Hassabis conducted by Big Technology Podcast’s Alex Kantrowitz. Brin went on to say he “didn’t know anything about consumer electronic supply chains,” or how difficult it would be to build smart glasses at a reasonable price point. The Google co-founder noted he’s a big believer in the form factor of smart glasses, and said he’s glad the company is pursuing them yet again, this time with “great partners who are helping us build this.” Earlier on Tuesday, Google unveiled its latest effort to develop Android XR smart glasses, roughly a decade after the Google Glass project shut down. Onstage, Google leaders showcased how their smart glasses, powered by DeepMind’s Project Astra, could assist with live translations, directions, and generally AI queries. To develop Android XR glasses with AI and AR capabilities, Google is working with a network of partners, such as Samsung and Xreal. Google is also investing up to $150 million in a partnership with Warby Parker, and taking an equity stake in the eyewear company, to support its smart glasses efforts. As Brin referenced, those experienced eyewear and electronics makers may be able to help with some of the supply chain problems associated with producing smart glasses. Brin noted how the advent of generative AI makes the capabilities of smart glasses much more tangible than when Google Glass was around. Earlier in the interview, Brin acknowledged how he’s effectively come out of retirement to work on Google’s Gemini efforts. The Google co-founder says he’s in the Mountain View, California office nearly every day, and says he’s helping the Gemini team with multimodal projects, such as Google’s video-generating model Veo 3. “Anybody who’s a computer scientist should not be retired right now,” said Brin. “They should be working on AI.” Techcrunch event Join us at TechCrunch Sessions: AI Secure your spot for our leading AI industry event with speakers from OpenAI, Anthropic, and Cohere. For a limited time, tickets are just $292 for an entire day of expert talks, workshops, and potent networking. Exhibit at TechCrunch Sessions: AI Secure your spot at TC Sessions: AI and show 1,200+ decision-makers what you’ve built — without the big spend. Available through May 9 or while tables last. Berkeley, CA | June 5 REGISTER NOW Previous reports have suggested that Brin has pushed Google’s Gemini teams quite aggressively to compete in the AI race. Reportedly, Brin told Google employees in a memo they should be in the office at least every weekday, and that working 60 hours a week is likely the sweet spot for productivity.

Google’s family of “open” AI models, Gemma, is growing. During Google I/O 2025 on Tuesday, Google took the wraps off Gemma 3n, a model designed to run “smoothly” on phones, laptops, and tablets. Available in preview starting Tuesday, Gemma 3n can handle audio, text, images, and videos, according to Google. Models efficient enough to run offline and without the need for computing in the cloud have gained steam in the AI community in recent years. Not only are they cheaper to use than large models, but they preserve privacy by eliminating the need to transfer data to a remote data center. During a keynote at I/O, Gemma Product Manager Gus Martins said that Gemma 3n can run on devices with less than 2GB of RAM. “Gemma 3n shares the same architecture as Gemini Nano, and is and is engineered for incredible performance,” he added. In addition to Gemma 3n, Google is releasing MedGemma through its Health AI Developer Foundations program. According to the company, MedGemma is its most capable open model for analyzing health-related text and images. “MedGemma [is] our […] collection of open models for multimodal [health] text and image understanding,” Martins said. “MedGemma works great across a range of image and text applications, so that developers […] can adapt the models for their own health apps.” Also on the horizon is SignGemma, an open model to translate sign language into spoken-language text. Google says that SignGemma will enable developers to create new apps and integrations for deaf and hard-of-hearing users. Techcrunch event Join us at TechCrunch Sessions: AI Secure your spot for our leading AI industry event with speakers from OpenAI, Anthropic, and Cohere. For a limited time, tickets are just $292 for an entire day of expert talks, workshops, and potent networking. Exhibit at TechCrunch Sessions: AI Secure your spot at TC Sessions: AI and show 1,200+ decision-makers what you’ve built — without the big spend. Available through May 9 or while tables last. Berkeley, CA | June 5 REGISTER NOW “SignGemma is a new family of models trained to translate sign language to spoken-language text, but it’s best at American Sign Language and English,” Martins said. “It’s the most capable sign language understanding model ever, and we can’t wait for you — developers and deaf and hard-of-hearing communities — to take this foundation and build with it.” Worth noting is that Gemma has been criticized for its custom, non-standard licensing terms, which some developers say have made using the models commercially a risky proposition. That hasn’t dissuaded developers from downloading Gemma models tens of millions of times collectively, however. Updated 2:40 p.m. Pacific: Added several quotes from Gemma Product Manager Gus Martins. Topics

Luminar, the lidar company founded by recently replaced CEO Austin Russell, is going through another restructuring, according to a recent regulatory filing. This new round of layoffs, which the company did not provide figures for, follows extensive cuts to the workforce in 2024. Luminar cut about 30% of its workforce in 2024, a reduction that was expected to cost $4 million to $6 million in additional cash charges. Some of those layoffs spilled into the first quarter of 2025. A total of 212 employees were laid off. In its latest regulatory filing, the company said it began additional layoffs May 15. These new layoffs are expected to cost $4 million to $5 million in cash charges. These costs are expected to be incurred in the second and third quarters of this year. The layoffs are the latest complication for Luminar. Earlier this month, the company’s board replaced Russell as CEO and board chair. The board issued a press release saying he resigned as the result of an ethics inquiry without providing any additional information. Luminar’s board replaced Russell and appointed Paul Ricci to the role. Ricci is the former chairman and CEO of Nuance. A day after the leadership change was announced, board member Jun Hong Heng also resigned, according to a regulatory filing, which stated his decision was not due to any disagreements with the company on any matter relating to the company’s operations, policies, or practices. The company has not responded to requests for comment. Russell became a billionaire after his lidar startup Luminar went public in 2021 after the company merged with special purpose acquisition company Gores Metropoulos Inc., with a post-deal market valuation of $3.4 billion. Luminar raised $250 million prior to the SPAC announcement. Techcrunch event Join us at TechCrunch Sessions: AI Secure your spot for our leading AI industry event with speakers from OpenAI, Anthropic, and Cohere. For a limited time, tickets are just $292 for an entire day of expert talks, workshops, and potent networking. Exhibit at TechCrunch Sessions: AI Secure your spot at TC Sessions: AI and show 1,200+ decision-makers what you’ve built — without the big spend. Available through May 9 or while tables last. Berkeley, CA | June 5 REGISTER NOW

Google co-founder Sergey Brin said he “made a lot of mistakes with Google Glass” during an onstage interview at Google I/O 2025 on Tuesday. Brin was a surprise addition to an interview with Google DeepMind CEO Demis Hassabis conducted by Big Technology Podcast’s Alex Kantrowitz. Brin went on to say he “didn’t know anything about consumer electronic supply chains,” or how difficult it would be to build smart glasses at a reasonable price point. The Google co-founder noted he’s a big believer in the form factor of smart glasses, and said he’s glad the company is pursuing them yet again, this time with “great partners who are helping us build this.” Earlier on Tuesday, Google unveiled its latest effort to develop Android XR smart glasses, roughly a decade after the Google Glass project shut down. Onstage, Google leaders showcased how their smart glasses, powered by DeepMind’s Project Astra, could assist with live translations, directions, and generally AI queries. To develop Android XR glasses with AI and AR capabilities, Google is working with a network of partners, such as Samsung and Xreal. Google is also investing up to $150 million in a partnership with Warby Parker, and taking an equity stake in the eyewear company, to support its smart glasses efforts. As Brin referenced, those experienced eyewear and electronics makers may be able to help with some of the supply chain problems associated with producing smart glasses. Brin noted how the advent of generative AI makes the capabilities of smart glasses much more tangible than when Google Glass was around. Earlier in the interview, Brin acknowledged how he’s effectively come out of retirement to work on Google’s Gemini efforts. The Google co-founder says he’s in the Mountain View, California office nearly every day, and says he’s helping the Gemini team with multimodal projects, such as Google’s video-generating model Veo 3. “Anybody who’s a computer scientist should not be retired right now,” said Brin. “They should be working on AI.” Techcrunch event Join us at TechCrunch Sessions: AI Secure your spot for our leading AI industry event with speakers from OpenAI, Anthropic, and Cohere. For a limited time, tickets are just $292 for an entire day of expert talks, workshops, and potent networking. Exhibit at TechCrunch Sessions: AI Secure your spot at TC Sessions: AI and show 1,200+ decision-makers what you’ve built — without the big spend. Available through May 9 or while tables last. Berkeley, CA | June 5 REGISTER NOW Previous reports have suggested that Brin has pushed Google’s Gemini teams quite aggressively to compete in the AI race. Reportedly, Brin told Google employees in a memo they should be in the office at least every weekday, and that working 60 hours a week is likely the sweet spot for productivity.

You can now visit the Viking world of How to Train Your Dragon and walk among the dragon trainers from the hit DreamWorks Animation franchise. io9 was on hand for opening week’s media preview in advance of this Friday’s first day for the newest theme park to join Universal Orlando’s offerings: Epic Universe. Of the portal realms we have visited so far, the Isle of Berk has probably the most impressive tech when it comes to bringing the fantasy franchise’s main component to life. As Hiccup says, they have dragons, and boy, will they blow you away. The first encounter we had when arriving was meeting one of the NightLights, the offspring of the franchise’s hero dragon Toothless and the Lightfury, Pouncer. Gizmodo has written about those scary-looking robot dogs in the past—and what struck me when the baby dragon walked out on its own (without a person, on all fours, in the scorching humid heat) is that its legs looked to move very similarly to the Boston Dynamics models. The flame-throwing version of the robot style makes a whole lot of sense now—kidding, these don’t breathe fire, but they could, though theme parks don’t have a good track record with fire breathing dragons, just ask the other guys. But isn’t it so darn cute? We got to see it do a little wiggle and strut with the face of a precious pure little guy who totally wouldn’t hurt us, unlike its military cousins. There’s more dragons throughout the land, which highlight various areas. Near the outdoor dining stations serving Viking fare like grilled meats, veggies, and mac and cheese in bread horns, you get to witness the snow wraith emit its icy blast at unsuspecting passers-by. It’s delightful when it catches people off guard. Then of course dragons are placed in various areas that correspond with their abilities, so you really get a sense of the co-op way they live with the Vikings of Berk. Character meet and greets with the dragon riders from the films such as Hiccup and Astrid also include their winged friends. Meeting Toothless is totally magical because of the seamless tech; various functions of articulation really work within the role they play. When you have your one-on-one moment with a dragon, they react and respond in such a life-like manner, it’s kind of alarming how quickly you let your guard down with the blink of those two big adorable eyes. They make the noises that give you the same cuteness aggression you’d get from a dog or cat. I really thought I needed a dragon’s nod of approval when Toothless sent me off on a high-flying ride on what’s probably my favorite family-friendly thrill coaster, Hiccup’s Winged Gliders. And it being set to John Powell’s movie score, is chef’s kiss. I really felt like I was soaring. So far on this visit, we haven’t gotten a chance to watch The Untrainable Dragon, the land’s stage show, which is set to feature a flying dragon as part of the production. But if all the other dragons make you feel like you’re right at home in Berk, we’ve got a good feeling about it. I don’t think any other land has gone to the lengths Berk does to incorporate the film’s fantasy creatures in this way. Sure, at lands such as Jurassic World-themed ones, there are some creatures beyond the boat raft, and it’s crazy cool to witness their simulated escape or get to meet in scheduled moments with the gentler giants. Berk just takes it to a new level. The dragons don’t just come out like your average characters at any park, they feel like an active part of the land that you can encounter within the story-play of Berk’s immersion. And that’s a high-flying feat. Epic Universe opens this Friday at Universal Studios Orlando. Travel and accommodations were provided for the purposes of this review. Want more io9 news? Check out when to expect the latest Marvel, Star Wars, and Star Trek releases, what’s next for the DC Universe on film and TV, and everything you need to know about the future of Doctor Who.

With seemingly no limit to the demand for artificial intelligence, everyone in the energy, AI, and climate fields is justifiably worried. Will there be enough clean electricity to power AI and enough water to cool the data centers that support this technology? These are important questions with serious implications for communities, the economy, and the environment.  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. But the question about AI’s energy usage portends even bigger issues about what we need to do in addressing climate change for the next several decades. If we can’t work out how to handle this, we won’t be able to handle broader electrification of the economy, and the climate risks we face will increase. Innovation in IT got us to this point. Graphics processing units (GPUs) that power the computing behind AI have fallen in cost by 99% since 2006. There was similar concern about the energy use of data centers in the early 2010s, with wild projections of growth in electricity demand. But gains in computing power and energy efficiency not only proved these projections wrong but enabled a 550% increase in global computing capability from 2010 to 2018 with only minimal increases in energy use.  In the late 2010s, however, the trends that had saved us began to break. As the accuracy of AI models dramatically improved, the electricity needed for data centers also started increasing faster; they now account for 4.4% of total demand, up  from 1.9% in 2018. Data centers consume more than 10% of the electricity supply in six US states. In Virginia, which has emerged as a hub of data center activity, that figure is 25%. Projections about the future demand for energy to power AI are uncertain and range widely, but in one study, Lawrence Berkeley National Laboratory estimated that data centers could represent 6% to 12% of total US electricity use by 2028. Communities and companies will notice this type of rapid growth in electricity demand. It will put pressure on energy prices and on ecosystems. The projections have resulted in calls to build lots of new fossil-fired power plants or bring older ones out of retirement. In many parts of the US, the demand will likely result in a surge of natural-gas-powered plants. It’s a daunting situation. Yet when we zoom out, the projected electricity use from AI is still pretty small. The US generated about 4,300 billion kilowatt-hours last year. We’ll likely need another 1,000 billion to 1,200 billion or more in the next decade—a 24% to 29% increase. Almost half the additional electricity demand will be from electrified vehicles. Another 30% is expected to be from electrified technologies in buildings and industry. Innovation in vehicle and building electrification also advanced in the last decade, and this shift will be good news for the climate, for communities, and for energy costs. The remaining 22% of new electricity demand is estimated to come from AI and data centers. While it represents a smaller piece of the pie, it’s the most urgent one. Because of their rapid growth and geographic concentration, data centers are the electrification challenge we face right now—the small stuff we have to figure out before we’re able to do the big stuff like vehicles and buildings. We also need to understand what the energy consumption and carbon emissions associated with AI are buying us. While the impacts from producing semiconductors and powering AI data centers are important, they are likely small compared with the positive or negative effects AI may have on applications such as the electricity grid, the transportation system, buildings and factories, or consumer behavior. Companies could use AI to develop new materials or batteries that would better integrate renewable energy into the grid. But they could also use AI to make it easier to find more fossil fuels. The claims about potential benefits for the climate are exciting, but they need to be continuously verified and will need support to be realized. This isn’t the first time we’ve faced challenges coping with growth in electricity demand. In the 1960s, US electricity demand was growing at more than 7% per year. In the 1970s that growth was nearly 5%, and in the 1980s and 1990s it was more than 2% per year. Then, starting in 2005, we basically had a decade and a half of flat electricity growth. Most projections for the next decade put our expected growth in electricity demand at around 2% again—but this time we’ll have to do things differently.  To manage these new energy demands, we need a “Grid New Deal” that leverages public and private capital to rebuild the electricity system for AI with enough capacity and intelligence for decarbonization. New clean energy supplies, investment in transmission and distribution, and strategies for virtual demand management can cut emissions, lower prices, and increase resilience. Data centers bringing clean electricity and distribution system upgrades could be given a fast lane to connect to the grid. Infrastructure banks could fund new transmission lines or pay to upgrade existing ones. Direct investment or tax incentives could encourage clean computing standards, workforce development in the clean energy sector, and open data transparency from data center operators about their energy use so that communities can understand and measure the impacts. In 2022, the White House released a Blueprint for an AI Bill of Rights that provided principles to protect the public’s rights, opportunities, and access to critical resources from being restricted by AI systems. To the AI Bill of Rights, we humbly offer a climate amendment, because ethical AI must be climate-safe AI. It’s a starting point to ensure that the growth of AI works for everyone—that it doesn’t raise people’s energy bills, adds more clean power to the grid than it uses, increases investment in the power system’s infrastructure, and benefits communities while driving innovation. By grounding the conversation about AI and energy in context about what is needed to tackle climate change, we can deliver better outcomes for communities, ecosystems, and the economy. The growth of electricity demand for AI and data centers is a test case for how society will respond to the demands and challenges of broader electrification. If we get this wrong, the likelihood of meeting our climate targets will be extremely low. This is what we mean when we say the energy and climate impacts from data centers are small, but they are also huge. Costa Samaras is the Trustee Professor of Civil and Environmental Engineering and director of the Scott Institute for Energy Innovation at Carnegie Mellon University. Emma Strubell is the Raj Reddy Assistant Professor in the Language Technologies Institute in the School of Computer Science at Carnegie Mellon University. Ramayya Krishnan is dean of the Heinz College of Information Systems and Public Policy and the William W. and Ruth F. Cooper Professor of Management Science and Information Systems at Carnegie Mellon University.

Google’s family of “open” AI models, Gemma, is growing. During Google I/O 2025 on Tuesday, Google took the wraps off Gemma 3n, a model designed to run “smoothly” on phones, laptops, and tablets. Available in preview starting Tuesday, Gemma 3n can handle audio, text, images, and videos, according to Google. Models efficient enough to run offline and without the need for computing in the cloud have gained steam in the AI community in recent years. Not only are they cheaper to use than large models, but they preserve privacy by eliminating the need to transfer data to a remote data center. During a keynote at I/O, Gemma Product Manager Gus Martins said that Gemma 3n can run on devices with less than 2GB of RAM. “Gemma 3n shares the same architecture as Gemini Nano, and is and is engineered for incredible performance,” he added. In addition to Gemma 3n, Google is releasing MedGemma through its Health AI Developer Foundations program. According to the company, MedGemma is its most capable open model for analyzing health-related text and images. “MedGemma [is] our […] collection of open models for multimodal [health] text and image understanding,” Martins said. “MedGemma works great across a range of image and text applications, so that developers […] can adapt the models for their own health apps.” Also on the horizon is SignGemma, an open model to translate sign language into spoken-language text. Google says that SignGemma will enable developers to create new apps and integrations for deaf and hard-of-hearing users. Techcrunch event Join us at TechCrunch Sessions: AI Secure your spot for our leading AI industry event with speakers from OpenAI, Anthropic, and Cohere. For a limited time, tickets are just $292 for an entire day of expert talks, workshops, and potent networking. Exhibit at TechCrunch Sessions: AI Secure your spot at TC Sessions: AI and show 1,200+ decision-makers what you’ve built — without the big spend. Available through May 9 or while tables last. Berkeley, CA | June 5 REGISTER NOW “SignGemma is a new family of models trained to translate sign language to spoken-language text, but it’s best at American Sign Language and English,” Martins said. “It’s the most capable sign language understanding model ever, and we can’t wait for you — developers and deaf and hard-of-hearing communities — to take this foundation and build with it.” Worth noting is that Gemma has been criticized for its custom, non-standard licensing terms, which some developers say have made using the models commercially a risky proposition. That hasn’t dissuaded developers from downloading Gemma models tens of millions of times collectively, however. Updated 2:40 p.m. Pacific: Added several quotes from Gemma Product Manager Gus Martins. Topics