DeepSeek’s Disruption: Rethinking AI’s Energy Consumption and Investment Strategies

The belief behind billions of AI investment is simple: more computing power means better AI. DeepSeek, a highly efficient model from Chinese researchers, challenged that. Could this be an Emperor’s New Clothes moment for AI? Let’s break it down.

01The energy demands of traditional AI

Traditional AI models, such as OpenAI’s GPT-4, need substantial computing resources:

• High energy consumption: early estimates put each ChatGPT query at around 0.0029 kWh, nearly ten times a standard Google search at around 0.0003 kWh (Goldman Sachs). Boiling a full kettle takes roughly 0.1 kWh, so that’s about 1/35 of a kettle per AI query against about 1/300 per search.
• Infrastructure: these demands need massive data centres with thousands of high-performance GPUs, driving up costs and environmental impact.

02DeepSeek: a shift in AI efficiency

• Optimised performance: streamlined algorithms let DeepSeek achieve results comparable to larger models with far less computing power.
• Reduced energy use: exact figures are proprietary, but early estimates suggested 10-35x less energy per query, closer to a Google search than a traditional AI query.

03What it meant for big tech investments

• NVIDIA: the leading supplier of AI chips saw a 17% drop in its share price after DeepSeek’s announcement, on fears that demand for high-end GPUs could fall (Barron’s).
• Energy providers: Vistra, Constellation Energy and Talen Energy, which power data centres, saw share price falls of more than 20%.
• Stargate: the US-backed $500 billion Stargate Project assumes infrastructure-heavy AI will remain the standard. DeepSeek suggested that assumption may not hold.

04The Emperor’s New Clothes

• Questioning assumptions: the belief that better AI needs immense power and energy is being challenged. Are existing models over-engineered?
• Strategic shifts: companies may need to pivot to more efficient models, reducing demand for expansive data centres and high-end hardware.

05The accelerating pace of AI

• Innovation: a focus on efficiency opens new avenues for AI that is both powerful and sustainable.
• Accessibility: models like DeepSeek bring AI to a wider range of people and organisations, lowering barriers to entry.

06Energy use in everyday terms

How each compares with boiling a full kettle (about 0.1 kWh):

07Don’t underestimate AI’s speed

• Efficiency over power: we can achieve more with less energy and computing.
• Global accessibility: lower costs bring AI to individuals, small businesses and regions without advanced infrastructure.
• Sustainability: lower energy per query means a smaller footprint per task, provided total use doesn’t simply balloon.

DeepSeek is a catalyst for re-evaluating where AI is heading. Efficiency doesn’t just save resources. It democratises AI, making it more accessible and more sustainable.