
Introduction: A New Contender in the AI Race and What Happened?
China-based DeepSeek has taken the AI industry by storm, whilst simultaneously disrupting major U.S.-based tech firms such as NVIDIA, OpenAI, Google, Meta and Microsoft. Implementing a well-timed release, DeepSeek revealed their latest LLM to the world on the same day as Trump took his first day in Office as the 47th President of the United States. Looking back in years to come, these two events, corresponding on the same day, may be the most significant events of this decade.
DeepSeek’s R1 model has been shown to outperform or competitively match the likes of chatGPT and Claude, two AI LLMs thought to be leading AI models. What’s impressive isn’t the ability to create a competing AI model - it’s being able to do so considering the limitations in capital and hardware.
For those new to the AI space, LLMs require a lot of capital to create, maintain and support (or so we thought). OpenAI’s chatGPT is reported to have spent over $100M to train the model. Comparatively, DeepSeek’s model is reported to have required $6M to train - a significant price difference. Not only that, but Chinese firms have faced strong constraints with the U.S. increasing export controls on cutting-edge chips; a must for training AI models. As a result, R1 is reportedly trained using 2,048 H800 GPUs—a staggering 74.4% fewer GPUs than ChatGPT, which was trained on 8,000 H100 GPUs, a significantly more powerful set of chips.
As such, this has sent a powerful statement: brute forcing AI with hardware isn’t the only way to win. It’s understood that DeepSeek trained their model more smartly, utilising better chain-thinking and while this presents a competitive challenge for U.S. AI companies, there’s a bigger, more global takeaway—DeepSeek-R1 signals a shift towards more resource-efficient AI, which could be an environmental game-changer.

How DeepSeek-R1 Challenges the High-Compute Status Quo
The AI industry has been dominated by companies like OpenAI, Google, and Anthropic, all of which rely on massive computational resources—primarily powered by the latest NVIDIA A100 and H100 GPUs. These chips allow for faster training and inference but come with a major downside: they consume enormous amounts of energy.
DeepSeek, however, had to work with significantly less advanced hardware due to U.S. export restrictions on AI chips. Despite these constraints, they developed an LLM that performs on par with Western models, demonstrating that AI breakthroughs don’t have to come at the cost of extreme energy consumption.

Why This Is Good for the Planet
Lower Energy Consumption
Training and running AI models on cutting-edge GPUs requires vast amounts of electricity. If AI companies optimise models to perform well on weaker hardware, this could drastically lower AI’s carbon footprint.
Decentralization of AI Development
Open-weight AI models like DeepSeek-R1 allow more companies and researchers to use and fine-tune them, reducing the need for massive, proprietary, resource-intensive training runs.
This shift could promote localised AI solutions, reducing the carbon impact of global-scale AI training monopolies.
Extending the Lifespan of Existing Hardware
If high-performance AI can run on older or less powerful chips, it delays e-waste production and reduces demand for constant hardware upgrades.
This is particularly important given the environmental toll of mining rare earth metals for advanced semiconductors.
Encouraging a Shift in AI Priorities
For years, the AI race has been about bigger, more powerful models. DeepSeek-R1’s success proves that efficiency can be just as valuable as brute-force computing power.
If AI companies shift focus toward making models more optimised, we could see a new era of green AI that prioritises performance without excessive energy demands.
The timing of this information being released is quite comical - just last week I was calculating rough CO2e emission estimates. Specifically, comparing localised AI models to corporate monoliths such as chatGPT. My calculations led me to the following conservative, rough and certainly not 100% accurate figures:
• Centralised ChatGPT-4 (1.8T parameters): ~ 10.54 million metric tonnes of CO2e per year + 4,000 metric tonnes daily through user-based inferencing
• Localised AI model (12M parameters): ~ 36.16 metric tonnes of CO2e per year (inferencing included)

The Flip Side: A Challenge for U.S. AI Leaders and the Inevitable Jevons Paradox
While the environmental benefits of DeepSeek-R1 are clear, this shift also disrupts the current AI power balance. American AI giants have built their dominance around having access to the best hardware—but if efficiency becomes the key to AI success, this could shake up their competitive advantage.
In the immediate short term, U.S. AI businesses may struggle to adjust, as their models have been optimised with the assumption of unlimited computing resources. However, in the long run, this could force AI companies to prioritise efficiency, which tends to be more sustainable and ultimately benefits both the industry and the planet.
At first glance, DeepSeek-R1’s efficiency appears to be a clear win for sustainability. However, history suggests that greater efficiency doesn’t always lead to lower environmental impact—in fact, it can have the opposite effect due to Jevons Paradox.
Jevons Paradox occurs when advancements that improve efficiency lead to an overall increase in resource consumption rather than a decrease. In the context of AI, if models become far more energy-efficient, the cost of running them drops. This could lead to a surge in AI adoption across industries, ultimately increasing overall energy demand and emissions.
For example:
More accessible AI models could mean that businesses, governments, and individuals integrate AI into everyday processes at a much larger scale.
Cheaper AI inference costs could incentivise companies to deploy AI models for trivial or unnecessary applications, leading to overuse.
A flood of AI startups and models could increase total training runs, consuming more electricity than today’s high-compute centralized systems.
It’s also worth noting that AI, at its purest definition (and what we now refer to as ‘Artificial General Intelligence’) has a unique capability and opportunity that all previous inventions of mankind lack; it can offset its own emissions. With a genuine artificial intelligence, one capable of thinking and learning on its own, it could develop technology or systems to offset and reduce its own emissions, as well as emissions produced by humans.

Conclusion: The Future of AI Must Be More Sustainable
DeepSeek-R1’s emergence is more than just another AI breakthrough—it’s proof that efficiency matters. While this may pose a challenge for Western AI giants in the immediate future, it presents an opportunity for a more sustainable AI industry.
If AI can reach ChatGPT-level performance without relying on the latest and most power-hungry chips, then we have a clear path toward AI that is both powerful and environmentally responsible.
In the end, the AI race shouldn’t just be about who has the biggest models or the best chips—it should be about who can create the most intelligent, efficient, and sustainable systems for the future of our planet. After all, there's no point in efficiently driving ourselves to extinction.
Sign up for my newsletter (HERE) to get exclusive access to:
🌿 Actionable sustainability tips
🔍 Myth-busting insights
📽️ Behind-the-scenes stories
📈 Industry trends and innovations
📊 Real-world case studies
🤝 Spotlight on collaborations
🔗 Transparency in action
🎁 Exclusive subscriber perks
Signing off,
Sasha Higham
Co-founder & CEO
Comments