CATS: A Carbon-Aware Task Simulator for Reducing AI Data Center Emissions

A trace-driven framework for quantifying spatial and temporal carbon savings in global AI inference fleets

Dayuan Chen; Ziliang Zong· 2026· DOI 10.48550/arXiv.2609.14775

The core problem

The rapid rise of generative AI is accelerating cloud data center expansion, with electricity demand projected to double by 2026. Because carbon intensity varies by more than 5.5x across grids and times of day, where and when inference tasks execute significantly affects operational emissions. Chen and Zong address this issue through three contributions: a global alignment dataset, the CATS simulator, and a quantification of achievable CO2 savings under realistic constraints. The work is motivated by a siting-carbon mismatch: 50% of current cloud sites lie in medium-to-high carbon-intensity grids, indicating unrealized carbon reduction potential. The central research question is how much operational emissions can be reduced by carbon-aware spatial and temporal scheduling on today's fleets, given realistic task mixes, diurnal curves, and service-level agreement (SLA) constraints.

In a 24-hour trace with 600,000 tasks at fleet utilization of 0.37, spatial shifting reduces CO2 by 38.4% versus the speed-first baseline. Temporal shifting yields 16% savings with bounded SLA violations at 3.27%. These results demonstrate that carbon-aware scheduling on today's fleets can achieve substantial operational emissions reduction. The spatial savings are larger because moving tasks to lower-carbon regions exploits the 5.5x variation in grid carbon intensity, while temporal shifting is limited by SLA deadlines and diurnal patterns. The reported SLA violation rate of 3.27% indicates a manageable trade-off between carbon savings and quality of service. The dataset analysis also reveals that 50% of current sites are in medium-to-high carbon-intensity grids, confirming a siting-carbon mismatch that could be addressed by locating future data centers in low carbon-intensity grids.

The rapid rise of generative AI is accelerating cloud data center expansion, with electricity demand projected to double by 2026. Because carbon intensity varies by more than 5.5x across grids and times of day, where and when inference tasks execute significantly affects operational emissions. Chen and Zong address this issue through three contributions: a global alignment dataset, the CATS simulator, and a quantification of achievable CO2 savings under realistic constraints. The work is motivated by a siting-carbon mismatch: 50% of current cloud sites lie in medium-to-high carbon-intensity grids, indicating unrealized carbon reduction potential. The central research question is how much operational emissions can be reduced by carbon-aware spatial and temporal scheduling on today's fleets, given realistic task mixes, diurnal curves, and service-level agreement (SLA) constraints.

The authors compile a global alignment dataset unifying 140 operational and planned cloud regions across 8 major providers with five-minute carbon-intensity traces for 145 grid regions from 2022 to 2024. They then develop CATS (Carbon-Aware Task Simulator), a flexible trace-driven framework that profiles six AI inference tasks across multiple GPU types, synthesizes realistic diurnal curves, geographical and task mixes, and SLA constraints, and evaluates spatial and temporal schedulers against two baselines. The simulator reports comprehensive metrics including carbon emissions, energy consumption, runtime, queue delay, and hardware utilization. The evaluation uses a 24-hour trace with 600,000 tasks at a fleet utilization of 0.37. Two baselines are considered: a speed-first baseline and presumably a carbon-agnostic alternative. The schedulers are assessed for their ability to shift tasks spatially (across regions) and temporally (across time), while respecting SLA bounds.

The findings advocate locating future data centers in low carbon-intensity grids and demonstrate that carbon-aware scheduling on today's fleets can achieve substantial operational emissions reduction. The 38.4% spatial reduction is particularly notable because it requires no new hardware, only intelligent task placement across existing regions. However, the 3.27% SLA violation rate for temporal shifting suggests that deferring tasks is less attractive when latency-sensitive inference is involved. The study's trace-driven approach allows for reproducible evaluation, but it also relies on synthetic diurnal curves and task mixes, which may not capture all real-world variability. The global alignment dataset covering 140 cloud regions and 145 grid regions from 2022 to 2024 provides a valuable resource for future research. The taxonomy candidates (Architecture, Cybersecurity, Network, Cryptography) are not directly addressed, but the work intersects with network-aware scheduling and architectural considerations for sustainable AI infrastructure. Overall, CATS offers a practical tool for cloud operators to quantify and reduce emissions, and its results support policy and investment decisions favoring low-carbon siting.

CS practitioners and researchers

Opening member content…