CauseWriter.ai generates 750k to 1M words for good every day and some of our clients asked about the energy involved with the use of AI. This is also something that our parent company, Whole Whale a Marketing and Analytics Agency focused on nonprofits cares a lot about.
Why “quietly”?
There are still a lot of unknowns about energy costs related tokens, querries, and model training. Rather than just waiting until we can get perfect numbers we are trying to figure it out with the current information we have. Additionally, we work with a mix of the best models which all have different ranges just to make it more difficult to calculate. Here is our current approach, it will evolve as new data emerges:
So how much energy does 1M AI processed words consume in kWh?
Processing 1 million words with modern AI models (like Gemini 2.5 Pro, GPT-4o, or Claude 3.5) uses between 0.03 and 2 kilowatt-hours (kWh)—about the same energy as running one LED lightbulb for a week. However this cost doesn’t include the cost of training the foundation model which is where the true energy costs live. This is clearly a big asterisk on the following calculation but it is still important consideration that the cost of these current models have already been realized and this now like driving on a road that is already built. Though to be intellectually honest, that road might have cut a path through a forest.
Estimated Monthly Energy Consumption for AI Inference (1M Words / ~1.33M Tokens)
| Model | Input Words | Output Words | Total Tokens | Energy per Token (Wh) | Monthly Energy (kWh) | Source |
|---|---|---|---|---|---|---|
| OpenAI GPT-4o | 1,000,000 | – | 1,330,000 | 0.00002 (lower bound) | 0.03 | Epoch AI |
| 0.0003 (upper bound) | 0.4 | Epoch AI | ||||
| 0.0015 (conservative max) | 2.0 | arXiv 2024 | ||||
| Anthropic Claude 3.5 | 1,000,000 | – | 1,330,000 | 0.00002 (lower bound) | 0.03 | Epoch AI |
| 0.0003 (upper bound) | 0.4 | Epoch AI | ||||
| 0.0015 (conservative max) | 2.0 | arXiv 2024 | ||||
| Google Gemini 2.5 Pro | 1,000,000 | – | 1,330,000 | 0.00002 (lower bound) | 0.03 | Epoch AI |
| 0.0003 (upper bound) | 0.4 | Epoch AI | ||||
| 0.0015 (conservative max) | 2.0 | arXiv 2024 |
Assumptions Based on Research
- Word to Token Conversion: 1 word ≈ 1.33 tokens, so 1 million words = 1.33 million tokens.
- Energy per Token:
- Lower bound: 0.00002 Wh/token (efficient/average case)
- Upper bound: 0.0003 Wh/token (pessimistic “worst case”)
- Max: 0.0015 Wh/token (conservative, rarely exceeded)
- Monthly Energy Consumption: = Total Tokens × Energy per Token ÷ 1,000 (to get kWh)
Research Sources on Energy Cost of AI
- Epoch AI: How much energy does ChatGPT use?
- arXiv 2024: Energy Considerations of LLM Inference
- EcoLogits: Environmental Impacts of LLM Inference
- Gemini 2.5 Pro (Google Cloud Vertex AI)
Carbon Neutral Chat Tool | CauseWriter.ai
CauseWriter.ai has chosen terrapass to cover the cost of the growing tokens processed per month at the highest estimates.
CauseWriter.ai — Monthly AI Inference Energy Use
| Metric | Estimate |
|---|---|
| Words processed per day | 1,000,000 |
| Words processed per month | 30,000,000 |
| Approx. tokens per month | 40,000,000 |
| Energy per 1M words (highest est) | 2 kWh |
| Monthly energy (kWh) | 60 kWh |
Calculation Details
1 word ≈ 1.33 tokens → 30M words ≈ 40M tokens
We also hope that by increasing the number of words for good created, the work our users do helps offset even more in a larger context.
1 million words/day × 2 kWh = 2 kWh/day
30 days × 2 kWh/day = 60 kWh/month
