Is the AI juice worth the carbon squeeze?

Data centers are anything but. The global cloud comprises 100 million or more individual computer servers, running continuously in over 10,000 factory-sized data centers. With the exponential growth of power-hungry AI services, Deloitte estimates that data center power consumption will almost double by 2030. OpenAI alone is planning a massive $100 billion data center called Stargate that will need five gigawatts of electricity, enough to power hundreds of thousands of homes.

AI systems, and the cloud technology that enables them, won’t drift away on a breeze. For all their faults, they’re here to stay. But can they evolve from today’s stream of hallucinated facts, rip-off images, and plagiarized text into something more useful—or at least less carbon-intensive?

• • •

No. We Need A Course Correction

1. A training wreck. Building AI models in the first place is inefficient. Deloitte unpacks the numbers in this engrossing, data-rich read. Some state of the art AI models in 2021 needed over a gigawatt of electricity each time they were trained on new data. By 2024, those models had scaled up by a factor of ten, and today’s are larger still. In 2018, OpenAI admitted the cost of training its software was doubling about every 15 weeks.

2. Water woes. About 40% of energy data centers use goes to cooling, and powerful AI-specific processors need special attention. Data centers use water in two main ways: about a third from evaporative cooling on site and two thirds indirect usage from the electricity generation process. Even if a data center is located well for power generation, that doesn’t mean it’s also good for local water supplies. A 2021 study out of Virginia Tech found that even back then, nearly half of US data centers were fully or partially powered by power plants located within water-stressed regions. Deloitte estimates that AI data centers’ freshwater demand could rise to as much as 1.7 trillion gallons by 2027.

3. Kicking bad AI habits. Having AI responses switched on by default isn’t helping. A generative AI request consumes ten times the electricity of a typical Google search, according to the International Energy Authority. And most AI answers today are riddled with errors, researchers at Columbia Journalism Review’s Tow Center for Digital Journalism found recently. Leading AI-driven search engines incorrectly answered more than 60% of test queries. What if in the name of saving carbon, we went back to relying on traditional (and even human) sources for many of our queries?

Yes. We Can Have More Data With Less Carbon

1. AI ❤ renewables. Amazon, Apple, Google, Meta and Microsoft account for over 45 gigawatts of wind and solar energy purchases worldwide—over half the global corporate renewables market, according to S&P Global. Data centers are also supporting the development of a range of new low-carbon energy sources, including advanced geothermal, small modular fission reactors and even nuclear fusion start-ups.

2. Power couples. Siting a data center properly can improve the affordability of electricity for domestic customers while reducing overall grid emissions, researchers at the Rocky Mountain Institute found. The trick is to locate a data center near a renewable energy source (no surprise there) but also close to an existing generator with an approved grid connection, such as a natural gas power plant. This counterintuitive pairing of fossil fuels with renewables means  the new data center avoids stressing the grid, thus protecting other customers from paying for infrastructure upgrades. RMI has identified dozens of suitable “power couple” locations across the US.

3. A digital drop in the ocean. AI looms large in our cultural life, but the carbon footprint of making and moving data pales in comparison with making and moving physical stuff. In 2023, the International Energy Authority calculated that energy demand from data centers and data transmission networks each account for at most 1.5% of global electricity use. And only about 10% of that is attributable to AI, according to an in-depth survey of AI’s carbon footprint at the Sierra Club.

• • •

What To Keep An Eye On

1. Hot AI is cool.  Chip makers should focus on building servers that run hotter not faster if they’re serious about saving power and water, according to research from Hong Kong. If computer chips could function at 41 Celsius, rather than the 25 Celsius they do today, every data center around the world could swap thirsty AC systems for cheap, low-power fans, slashing nearly 60% off their power bills.

2. More slop for less juice? The tech world was rocked earlier this year with the emergence of a Chinese AI model called DeepSeek that takes 10 to 40 times less power to train than some American AIs. Tech and power stocks slumped on the news but it now appears that DeepSeek is less efficient when it comes to answering queries, suggesting that AI’s power consumption might even out over the long-term.

3. AI in your hand. Moving AI systems from the cloud to your phone can mean a 100 to 1,000-fold reduction in energy consumption per task, according to AI startup DeepX founder Kim Lokwon, writing at the World Economic Forum. He also suggests an energy credit trading scheme (developed by his company) could drive lower-carbon AI models.