A new arXiv preprint 2603.20897v2 estimates AI data centers could contribute up to 1.4°C of additional global warming by 2060. The model assumes 30% annual growth in AI compute demand, extrapolating from recent trends in training and inference workloads.
Key facts
- 1.4°C: upper-bound warming contribution from AI data centers by 2060.
- 30%: assumed annual growth rate for AI compute demand.
- 2030: year AI data center emissions could exceed aviation.
- 0.3°C: lower-bound warming under moderate growth and efficiency gains.
- 1.4: current average PUE for data centers in the model.
The paper, titled "Quantifying the impact of AI data centers in a warming world," uses a coupled energy-climate model to simulate the warming impact of direct and indirect emissions from AI data centers. Direct emissions from AI data centers could exceed current aviation emissions by 2030, the authors estimate. The 1.4°C figure represents the upper bound under a scenario with no efficiency improvements or carbon capture. Under a more moderate scenario with 15% annual compute growth and aggressive efficiency gains, the contribution drops to 0.3°C by 2060.
Key Takeaways
- AI data centers could add 1.4°C to global warming by 2060, per a new arXiv preprint, assuming 30% annual compute growth.
- The paper highlights the need for policy intervention.
How the model works
:max_bytes(150000):strip_icc()/GettyImages-2196329960-d5d4407cd222466c9211ccb1dc1f9ea0.jpg)
The researchers model AI compute demand as a function of training and inference compute, using published figures from major labs. They then map compute to energy consumption using current data center power usage effectiveness (PUE) averages of 1.4, and project future PUE improvements down to 1.1. Emissions are calculated using grid carbon intensity projections from the IPCC's Shared Socioeconomic Pathways. The climate response is simulated using a simple climate model (MAGICC6) that translates emissions into temperature change.
Key assumptions and limitations
The 30% growth rate is derived from recent trends in AI compute demand, which has grown roughly 10x per year for frontier models since 2018. The authors acknowledge this may slow as scaling laws hit diminishing returns. The model does not account for potential breakthroughs in low-power AI hardware, such as analog or photonic chips. It also assumes no major policy intervention, such as carbon taxes or emissions caps on data centers. The authors call for policy interventions to decouple AI growth from emissions, including mandatory efficiency standards and investment in carbon-free energy.
The paper is a preprint and has not yet been peer-reviewed. The authors did not disclose funding sources or potential conflicts of interest. The model code is not publicly available.
What to watch
Watch for follow-up papers incorporating hardware efficiency breakthroughs, such as analog AI chips or optical interconnects, which could significantly alter the energy trajectory. Also watch for policy moves in the EU or US to mandate data center emissions reporting, which would provide real-world data to validate or challenge the model's assumptions.
Source: dcmag.fr
[Updated 04 Jul via dck_news]
NERC's 2026 report warns that AI data center campuses approaching gigawatt scale and clustering regionally pose a new threat to grid stability, requiring updated modeling standards and regulatory frameworks [per Data Center Knowledge]. The report highlights that abrupt disconnections at such scales could destabilize regional grids, adding a grid-reliability dimension to the warming concerns raised by the arXiv paper.
[Updated 04 Jul via dcd_news]
Canada Pension Plan Investment Board (CPP Investments) commits $1.75 billion to EQT and EdgeConneX's AI data center build-out, signaling major institutional capital flowing into the sector [per Data Center Dynamics]. EdgeConneX claims a 10GW development pipeline, underscoring the scale of expansion that could drive the warming projections in the arXiv paper.
