Research & Impact

AI Carbon Impact

How AI model usage relates to CO₂ emissions — with cited research, Virginia grid context, and an interactive usage meter.

HF models trained 0 metric tons CO₂e (5,227 models)
LLM-assisted code 0 × higher footprint than manual
Reasoning AI prompts 0 × more CO₂ than concise answers
PJM CO₂ drop 0 % since 2005 (2023)

What the research shows

Cited findings on AI & code emissions

Peer-reviewed studies and open tools that quantify the carbon footprint of AI training, inference, and developer workflows.

Training emissions at scale

An analysis of 5,227 Hugging Face models estimates aggregate training emissions of approximately [1]60,000 metric tons CO₂e — comparable to the annual emissions of ~42,000 cars. Annual training emissions grew from ~450 tCO₂e (2020–2021) to over 40,000 tCO₂e (2024–2025), nearly two orders of magnitude in five years [1].

Training vs. inference

Hugging Face's BLOOM model emitted ~25 metric tons of CO₂ during training; doubling to ~50 metric tons when including hardware manufacturing, infrastructure, and post-training operation [2]. GPT-3 training was estimated at over 500 metric tons of CO₂ [2].

Prompt-level emissions

A study of 14 LLMs found that reasoning-enabled models produce [3]up to 50 times more CO₂ emissions than concise response models. The most accurate reasoning model (70B parameters) produced three times more emissions than similarly sized concise-answer models [3]. Users can reduce emissions by prompting for concise answers and reserving high-capacity models for tasks that genuinely require them [3].

LLM-assisted coding

LLM-assisted code generation leads to an average [4]32.72× higher carbon footprint than manual software development (range: 19.92×–43.81×). For LLM-assisted workflows, 98.68% of the carbon footprint comes from querying the LLM itself [4].

Virginia grid context

CarbonSense measures against the PJM Dominion (VA-DOM) zone. PJM's average CO₂ emission rates fell [5]43% from 2005 to 2023 as cleaner resources replaced older units [5]. In Virginia, natural gas represents 52% of installed capacity, with hydro at 15% and nuclear at 14% [6]. Solar represents 41% of new interconnection requests in the state [6].

Measurement tools

Open-source tools like [7]CodeCarbon track CO₂ from local compute (CPU, GPU, RAM) using regional grid carbon intensity, while EcoLogits estimates GenAI API impacts across providers like OpenAI, Anthropic, and Mistral [7]. Frameworks such as LLMCarbon model end-to-end LLM carbon footprints including training, inference, experimentation, and storage [8].

Interactive

AI Carbon Meter — try it yourself

Select AI tasks to see how quickly CO₂ adds up. Based on per-prompt estimates for different model types.

Your AI Carbon Footprint

Each task adds emissions based on real per-prompt estimates. Watch the meter fill as your character uses more AI.

Emissions 0 g CO₂

Total emissions

0 g CO₂

Start adding tasks to see your impact.

👨‍💻
Dev
Idle

References

Sources & citations

  1. Hugging Carbon — Quantifying Training Carbon Emissions of AI Models at Scale [1]
  2. The Carbon Footprint of Artificial Intelligence — Communications of the ACM [2]
  3. Some AI prompts cause 50× more CO₂ emissions than others — Frontiers [3]
  4. Comparative Analysis of Carbon Footprint in Manual vs. LLM-Assisted Code Development [4]
  5. Emission Rates in PJM Reach All-Time Low [5]
  6. 2024 Virginia State Infrastructure Report — PJM [6]
  7. CodeCarbon — Measure the environmental footprint of compute & GenAI [7]
  8. LLMCARBON — Modeling the End-to-End Carbon Footprint of LLMs — ICLR 2024 [8]