New research finds AI and data centers cause 0.5% of global CO2 emissions today — but the bigger climate risk is AI-driven economic growth, not the servers.
The surging resource demands of artificial intelligence make headlines almost daily. The United Nations recently noted that by 2030, data centers could be using three times as much electricity as the combined annual demand of Pakistan, Bangladesh, and Nigeria — countries collectively home to more than 650 million people. That’s fueled a genuine debate: will AI make climate change dramatically worse, or help solve it?
Management professor John Sterman and Koru Labs president Jennifer Turliuk, SFMBA ’25, take on that question directly in a new paper, “The Net Climate Impact of Artificial Intelligence: Balancing Current Costs With Future Climate Benefits,” forthcoming in the book “Sustainable Climate-Resilience: Incremental Pathways for Cities and Communities” and available now as a preprint. Their framework is built into En-ROADS, an interactive climate simulator developed by Climate Interactive that lets users visualize how policy and technology choices shape global temperature, sea level rise, extreme weather, energy consumption, and GDP.
“AI is a remarkable technology, both promising and problematic,” Sterman said. “But, by itself, it does not solve the climate crisis. We need sound policies to do that.”
AI and Data Centers Cause 0.5% of Global CO2 Emissions Today
Sterman and Turliuk reviewed the most current data on AI’s climate footprint and found that carbon emissions from AI and data centers are responsible for roughly 0.5% of global CO2 emissions from energy today — a small but rapidly growing share.
“AI and data center emissions are making climate change worse today, and they are growing rapidly,” Sterman said.
According to En-ROADS projections, the direct impact of AI and data center emissions on the climate is expected to be on the order of a tenth of a degree Celsius by 2100. That may sound small, but every tenth of a degree of warming worsens climate harms and makes it harder to hold warming under the Paris Agreement’s 2°C (3.6°F) ceiling.
Three factors currently limit AI’s direct climate impact: emissions from AI and data centers remain small relative to the broader economy (transportation, buildings, industry, agriculture, deforestation, and fossil-fuel methane leakage); renewable energy is decarbonizing electricity production, including power for data centers; and AI systems themselves are becoming more efficient through better algorithms, chips, and data center design.
That last point cuts both ways, though. As AI and data centers get more efficient, the cost of training and running AI drops — which stimulates more AI use and offsets some of the efficiency gain. Researchers call this the direct rebound effect. Using En-ROADS, Sterman showed that even sharply faster reductions in AI’s energy intensity would produce only a negligible reduction in global warming by 2100, absent new climate policy.
The Indirect Rebound Effect Could Push Warming From 3.3°C to 3.6°C by 2100
The bigger risk, according to Sterman and Turliuk, isn’t AI’s direct energy use at all — it’s the indirect rebound effect. AI is widely expected to boost productivity and accelerate economic growth, but faster growth means more energy demand and more emissions economy-wide, not just from AI itself.
Using a scenario in which AI eventually boosts gross world product by 25% — a figure well below more optimistic industry projections — Sterman found that the added production of goods and services would raise energy demand and emissions enough to push expected warming from 3.3°C (about 5.9°F) by 2100 up to 3.6°C (6.5°F). That additional 0.3°C would meaningfully worsen sea level rise, wildfires, extreme weather, and crop yield declines.
Sterman illustrated the mechanism with electric vehicles: an EV can save its owner several thousand dollars a year in gas and maintenance costs. “When I ask executives what they would do with those extra thousands,” he said, “they often say ‘buy more stuff’ and ‘take my family to Disney World.’ But the emissions from that trip to Orlando can outweigh the reduction from driving your EV, increasing your carbon footprint and worsening climate change.”
“Indirect rebound is the big issue,” Sterman said. “Individuals, corporations, and governments all want faster growth in their income, sales, and economy, and AI may help achieve those goals. But unless we implement policies and actions that rapidly lower emissions across the economy as a whole, then any economic boost from AI will generate a lot more emissions, and the climate will get notably worse.”
Why AI-Driven Climate Innovation Won’t Arrive Fast Enough on Its Own
Proponents argue AI will accelerate innovations that cut emissions — from emissions-free energy and efficient buildings to low-carbon cement, steel, and fertilizer production and practical carbon capture. If so, the economy could theoretically grow faster while emissions fall.
Sterman and Turliuk point to two problems with that logic: focus and timing. On focus, the vast majority of AI use today isn’t directed at solving climate or other environmental and social problems. On timing, even genuinely useful AI-driven innovations take years to move from lab breakthrough to pilot project to raised capital, community acceptance, and the production capacity and supply chains needed to scale globally.
To test this, Sterman modeled a scenario in which AI-generated innovation cut the cost of new nuclear power plants in half. The climate benefit was negligible: nuclear power did grow faster, squeezing out coal and gas, but it also displaced renewable energy deployment — substituting one clean energy source for another rather than adding net new decarbonization.
Even if AI-enabled innovation eventually lowers emissions enough to offset the earlier increase, Sterman notes that the damage already done doesn’t reverse. “Even if emissions eventually fall back to what they would have been, sea levels won’t fall back,” he said. “The forests burned, the crops lost, and all the businesses and homes destroyed along the way by the extra warming from AI won’t reappear.”
1 thought on “AI Climate Impact: New Research Finds the Real Risk Is Indirect”