Data centers' electricity use could more than double to about 945 terawatt-hours by 2030. That surge would raise both costs and emissions tied to generative AI, and Goldman Sachs estimates about 60 percent of the extra power could come from fossil fuels, adding roughly 220 million tons of CO2. MIT researchers and industry partners are racing to cut model and hardware energy use so the tradeoffs between pricey chips, power and software design don't determine who wins the AI boom.

Energy math is changing the game

AI workloads are power hungry. Training and running large models needs lots of GPUs. An April 2025 International Energy Agency projection cited by researchers puts data-center demand at about 945 terawatt-hours by 2030. That's more electricity than some countries use today. An August 2025 Goldman Sachs analysis referenced by MIT researchers found that about 60 percent of the rising demand could be met by fossil fuels, adding roughly 220 million tons of carbon dioxide.

That scale matters for costs. More power means bigger bills for cloud providers and enterprises. It also means higher operational emissions. Those two factors change the economics of building and running AI systems.

The picture includes more than just running the servers. Vijay Gadepally, senior scientist at MIT Lincoln Laboratory, notes that the construction and retrofitting of data centers carries its own emissions. He calls the embodied carbon in the steel, concrete and chillers a major part of the total footprint. The largest data centers can be enormous. MIT's reporting points to the China Telecomm-Inner Mongolia Information Park as roughly 10 million square feet, a scale that brings both energy density and construction emissions into focus.

Hardware and software are diverging

Those costs push different firms down different paths. Firms that design or own chips, or that can place workloads where power is cheap or low-carbon, can control a big piece of the bill. Pure-software companies that rent GPU time face a different set of pressures.

The energy and capital intensity of running cutting-edge models tends to favor companies with access to specialized chips and custom data-center arrangements.

At the same time, software work can reduce hardware needs. MIT and industry labs report progress on shrinking models while keeping performance. The MIT-IBM Watson AI Lab has highlighted research that slashes computational overhead and makes models smaller without losing capability. Those kinds of efficiency gains change the calculus. They lower operating costs and curb emissions. But the gains often come from joint hardware-software innovation. That puts a premium on firms that pair both skills.

Mastheads

What changes isn't just speed or accuracy. It's who captures value. When running a model becomes a major line item on operating statements, ownership of the stack matters. Companies that can co-design chips, cooling and model architectures can reduce both costs and carbon. Companies that rely entirely on third-party cloud GPUs may find margins squeezed or face higher tradeoffs between performance and sustainability.

Academic and industry pushback

Researchers aren't standing still. MIT has launched initiatives aimed at guiding AI's next phase. The new MIT Generative AI Impact Consortium brings together industry partners and university researchers to study how to make generative AI more effective and safer. Anantha Chandrakasan, dean of the School of Engineering and MIT's chief innovation and strategy officer, who leads the consortium, said, "Generative AI and large language models [LLMs] are reshaping everything, with applications stretching across diverse sectors."

The consortium frames its work around questions about AI-human collaboration, how AI affects behavior, and how interdisciplinary research can produce safer designs. Daniel Huttenlocher, dean of the MIT Schwarzman College of Computing and co-chair of the GenAI Dean's oversight group, helped lay out those priorities. Tim Kraska, associate professor of electrical engineering and computer science in MIT CSAIL and co-faculty director of the consortium, argued the field lacks solid design principles. "Everybody recognizes that large language models will transform entire industries, but there's no strong foundation yet around design principles," Kraska said. "Now is a perfect time to look at the fundamentals."

The MIT-IBM Watson AI Lab offers a complementary example of how research and industry link up. The lab reports 54 patent disclosures, more than 128,000 citations and an h-index of 162, along with over 50 industry-driven use cases. Its work spans improving imaging for medical devices to modeling chemistry and cutting computational overhead. Aude Oliva, MIT director of the lab, says the partnership helps identify practical problems where AI can be deployed effectively.

Operational fixes and construction choices

Researchers are targeting both operational and embodied emissions. On the operational side, work focuses on algorithmic efficiency, better scheduling of workloads and rethinking cooling and power management. On the embodied side, the building materials and retrofitting choices matter. MIT reporting notes that companies such as Meta and Google are exploring more sustainable materials for data-center construction.

Gadepally emphasized the gap between the operational carbon conversation and the embodied-carbon reality. "The operational side is only part of the story," he said. Some interventions that cut energy use during operation may also reduce embodied carbon. But he added that more work is needed on the construction side.

The combination of rising energy demand, potential reliance on fossil fuels, and the race for efficiency creates uneven incentives. Firms that invest in custom silicon and colocated infrastructure can lower per-unit compute costs. They also can claim lower emissions if they secure low-carbon power. Firms that depend on public clouds still benefit from economies of scale at large cloud providers, but they face exposure to wholesale power costs and the carbon mix of the regions where their workloads run.

At the same time, progress in software can shift market power. The MIT-IBM lab's work on shrinking models and keeping performance points to a route where software ingenuity reduces reliance on the most power-hungry chips.

Related Articles

Tim Kraska said, "Now is a perfect time to look at the fundamentals, the building blocks that will make generative AI more effective and safer to use."

This article was created with AI assistance.