Why we don’t need more data centers to build better AI
UC Berkeley researchers argue that smaller, energy-efficient open-source AI models can match the performance of large proprietary systems, reducing environmental and economic costs while preserving scientific control.
UC Berkeley experts Carl Boettiger and Fernando Pérez, along with collaborator Cassie Buhler, contend in a Nature commentary that AI development should prioritize efficiency over scale to curb energy and water demands from data centers. They note that while data centers represent a concentrated energy burden, smaller open-source models are rapidly closing the performance gap with proprietary systems like ChatGPT or Gemini. The authors suggest that shifting focus to these alternatives could mitigate environmental harm without sacrificing AI’s scientific and social benefits.
Boettiger highlights how the current AI industry’s race to build ever-larger models has led to unsustainable energy demands, often concentrated in underserved communities. He argues that historical trends in computing—where innovations eventually become miniaturized—should apply to AI, but economic pressures prioritize speed over efficiency. Pérez adds that centralized, proprietary models centralize data control, raising privacy concerns, while open-source tools empower users to adapt technology to their specific needs rather than relying on subscription-based platforms.
The researchers point to recent milestones, such as Facebook’s Llama model and China’s DeepSeek, which demonstrated that open-source AI can achieve competitive performance with far less energy. They note that open models now lag only six months behind frontier systems but require increasingly modest hardware, enabling local execution on devices like laptops equipped with NVIDIA RTX Spark chips. This shift could reduce reliance on energy-intensive data centers, which face growing political and community opposition due to their environmental and social impacts.
Pérez emphasizes the importance of scientific autonomy, warning against a future where researchers are locked into proprietary tools that restrict innovation. He argues that the AI revolution was built on decades of open infrastructure and data, and that reverting to closed systems undermines progress. Boettiger concludes by urging a redefinition of AI beyond its association with massive data centers, comparing it to distinguishing between a bicycle and an airplane in terms of carbon footprints—both enable transportation, but their impacts differ vastly.