Every day, it feels like we’re a step closer to a future shaped by AI—one with smarter healthcare, faster transportation, and solutions to challenges we’ve never been able to tackle before. Yet, with all the excitement about AI’s potential, there’s a pressing question we must face. Can we fully embrace AI’s potential while ensuring it doesn’t come at the expense of the environment?
AI’s energy footprint is no small matter. The rapid expansion of large-scale models leads to massive energy consumption, primarily due to the high computational power required for training. Think of it like running a data centre the size of a football field, operating 24/7—all powered by electricity, which in many cases still relies on fossil fuels. But how much energy are we really talking about?
To understand the environmental impact, let’s consider the popular AI model—ChatGPT by OpenAI—and analyse its energy consumption during training. Training GPT-4 required 10,000 Nvidia A100 GPUs over six months—consuming approximately 14.19 GWh of energy, equivalent to the electricity used in powering over 17,000 Coldplay concerts.
Amidst the energy-hungry AI revolution, a beacon of hope emerges—open-source AI. By championing efficiency and smarter resource allocation, these models prove that cutting-edge AI doesn’t have to come at an environmental cost. Unlike proprietary models, open-source AI models are developed collaboratively, allowing for more transparency, efficiency, and adaptability. These models often optimise resource use, making them a more sustainable alternative to closed-source AI.
China is emerging as a leader in open-source AI development with companies like Alibaba, Baidu and DeepSeek at the forefront. DeepSeek-V3, for example, was trained using 2,048 Nvidia A100 GPUs over two months, resulting in a total energy consumption of 0.93 GWh—15 times less than GPT-4.
The carbon footprint of AI training is heavily influenced by where the model is trained. For example, China’s reliance on coal-heavy electricity means that the same unit of energy produces higher emissions than in the U.S., where cleaner energy sources are more prevalent. This difference significantly impacts the overall environmental cost of AI training. Training models in regions with cleaner energy sources can greatly reduce their carbon emissions. When comparing emissions, GPT-4 generates nearly 10 times the CO₂ emissions of DeepSeek-V3. Embracing open-source models not only optimises resource utilisation but also leads to reduced energy consumption, thereby lowering emissions.
Beyond training, AI’s energy demands continue during real-world usage. Every query processed by ChatGPT or DeepSeek requires electricity, adding to its lifetime carbon footprint. A Google search emits approximately 0.001 kg CO₂e per query. ChatGPT-4, being more computationally intensive, consumes about 0.0029 kWh (2.9 Wh) per query, making its energy use nearly thirty times higher than a typical Google search.
The good news is that optimisation efforts, energy-efficient chips, and renewable energy integration can make AI more sustainable. This shift reconnects us with a crucial aspect of AI’s future, as discussed earlier, where collaborative innovation and transparency drive more sustainable practices.
Open-source models allow the global community to come together and contribute to more energy-efficient systems. Companies like OpenAI, while innovative, tend to operate under a capitalist model that focuses on building proprietary systems and locking down access to their models and data. This approach creates a moat, a barrier that keeps competitors at bay. However, the open-source approach flips the script. By making the AI models and the data used to train them accessible to everyone, we democratise the process. This encourages collaboration across sectors, countries, and industries to solve common problems, such as AI’s environmental impact.
But open-source isn’t just about making models publicly available—it’s also about data. Imagine if data on carbon emissions, energy use, and environmental impact was shared openly. Companies could collaborate to reduce their carbon footprints, refine their processes, and work toward a shared goal of sustainability. Open data would empower the world to see the full scope of our environmental challenges and collaborate on finding solutions.
Appendix: Our Calculations
1. GPT-4 Energy Consumption During Training
Hardware Used: 10,000 Nvidia A100 GPUs
Power Consumption per GPU: 400W (0.4 kW) at 90% utilisation → 0.36 kWh per hour
Total Power Consumption: 10,000 GPUs × 0.36 kWh = 3,600 kWh per hour
Training Duration: ~6 months = 4,380 hours
Total Energy Used: 3,600 kWh × 4,380 hours = 14.19 GWh
Equivalent to: Powering over 17,000 Coldplay concerts
2. DeepSeek-V3 Energy Consumption During Training
Hardware Used: 2,048 Nvidia A100 GPUs
Power Consumption per GPU: 0.4 kW at 90% utilisation → 0.315 kWh per hour
Total Power Consumption: 2,048 GPUs × 0.315 kWh = 645.12 kWh per hour
Training Duration: ~2 months = 1,440 hours
Total Energy Used: 645.12 kWh × 1,440 hours = 0.93 GWh
3. Carbon Footprint Comparison
Emission Factor (China): 0.5572 kg CO₂/kWh
Emission Factor (USA): 0.358 kg CO₂/kWh
DeepSeek-V3 (China)
Total Energy Consumption: 930,000 kWh (0.93 GWh)
Carbon Emissions: 930,000 × 0.5572 = 518.2 metric tonnes CO₂e
