DeepSeek-R1 Environmental Impact
Open-source reasoning model — highest energy use
- Architecture
- Transformer Mixture-of-Experts (decoder-only)
- Parameters
- 671B
- Context
- 128,000 tokens
- Provider
- DeepSeek
Energy per query
29.0 Wh
97x more than a Google search (0.3 Wh)
CO2 per query
14.0 g
China grid (550 gCO₂/kWh)
Water per query
150 mL
~7 queries to fill 1 litre
Processing location
China (Hangzhou / Hainan)
Provider
DeepSeek
Category
Text / Chat
Grid carbon intensity
550 g CO2/kWh (30% renewable)
How does DeepSeek-R1 compare?
Detailed Breakdown
Energy Consumption
DeepSeek-R1 is the most energy-intensive model benchmarked, at 29 Wh per query — nearly 100x a Google search. As a reasoning model, it performs extended chain-of-thought computation. Despite using a Mixture-of-Experts architecture designed to be efficient, the sheer scale of reasoning steps and its deployment on older hardware contribute to its high energy draw.
Power Source & Carbon
DeepSeek processes queries on infrastructure in mainland China, primarily in Hangzhou (home of parent company High-Flyer) and an underwater data center off the coast of Hainan Island. China's electricity grid has a carbon intensity of approximately 550 g CO2/kWh — roughly 45% higher than the US average — with about 60% of electricity generated from coal. This makes DeepSeek's carbon emissions per query among the highest of any model.
Water Usage
At over 150 mL per query, DeepSeek-R1 has the highest water consumption of any text model. This is driven by both its extreme energy use (requiring more cooling) and the hot, humid climate at the Hainan underwater data center location. A session of 10 queries would consume over 1.5 liters of water.
About DeepSeek-R1
DeepSeek-R1 made waves as the first open-source model to rival OpenAI's o1 in reasoning benchmarks — but it shares the same environmental weakness. Extended chain-of-thought inference means each query can consume 100x more energy than a standard chat exchange. What makes R1 uniquely interesting from a sustainability perspective is that it is open-source: anyone can run it on their own hardware, choosing their own electricity source. A developer running R1 on a home server powered by rooftop solar has a fundamentally different carbon footprint than the same model running in a coal-heavy data centre region.
DeepSeek-R1 in Context
The efficiency alternative
Gemini Nano performs the same type of task using just 0.01 Wh per query — 100% less energy than DeepSeek-R1. For a user sending 25 queries per day, switching would save 264.5 kWh per year.
At global scale
With an estimated 10M+ daily users averaging 10 queries each, DeepSeek-R1 consumes roughly 2900 MWh of electricity per day — enough to power 96667 homes.
Your yearly DeepSeek-R1 footprint
At 25 queries per day, your annual DeepSeek-R1 usage consumes 264.6 kWh — a meaningful fraction of household electricity. That produces 127.8 kg of CO₂.
Key Insights
DeepSeek DeepSeek Family
How energy efficiency has evolved across versions.
What does your DeepSeek-R1 usage cost the planet?
Use our calculator to estimate your personal environmental footprint based on how often you use DeepSeek-R1.
Calculate My ComputeFrequently Asked Questions
How much energy does DeepSeek-R1 use per query?
Each DeepSeek-R1 query consumes approximately 29.0 Wh of energy. This is 97x more than a traditional Google search (~0.3 Wh).
What is DeepSeek-R1's carbon footprint?
Based on the carbon intensity of China (Hangzhou / Hainan), each query produces approximately 14.0 g of CO2. The grid in this region has a carbon intensity of 550 g CO2/kWh with 30% renewable energy.
How much water does DeepSeek-R1 use?
Each query consumes approximately 150 mL of water, primarily used for cooling the data centers that process the request.
How does DeepSeek-R1 compare to a Google search?
A DeepSeek-R1 query uses 97x more than a Google search in terms of energy. A Google search uses approximately 0.3 Wh, while DeepSeek-R1 uses 29.0 Wh.
Technical Details
Architecture
Transformer Mixture-of-Experts (decoder-only)
Parameters
671B
Context window
128,000 tokens
Release date
2025-01-20
Open source
Yes
Training data cutoff
2024-11