Berkeley researchers argue smaller open-source AI models can replace massive data centers

UC Berkeley researchers Carl Boettiger and Fernando Pérez argue in a Nature commentary that the current race to build ever-larger AI models and data centers is driven by business incentives rather than technical necessity.

They note that open-source large language models have closed the performance gap with proprietary systems such as ChatGPT, Claude and Gemini to roughly six months, while the hardware required to run them has shrunk from server racks to laptops.

Boettiger, an associate professor of environmental science, policy and management, says data centers create concentrated energy spikes that raise local power rates and are often sited in communities least able to absorb the costs.

Pérez, an associate professor of statistics, adds that centralized models allow companies to collect user data for further training, whereas open-source alternatives let scientists and the public retain control over their tools and data.

The authors cite the 2023 leak of Meta's Llama model and the 2025 release of China's DeepSeek model as turning points that demonstrated high performance with far lower energy use.

Boettiger predicts that within six months open models will run on common consumer hardware, pointing to laptop manufacturers already shipping chips designed for local AI inference.

They compare today's massive models to Formula One cars — expensive, complex and overkill for most tasks — while open models are more like reliable, affordable sedans that meet everyday needs.

The researchers urge environmental scientists and other researchers to adopt lower-footprint open-source tools rather than abandoning AI altogether, arguing that the environmental cost stems from current business models, not from the physics of computation itself.