Intelligence per Watt: Measuring Intelligence Efficiency of Local AI
A new research paper introduces "intelligence per watt" (IPW) as a metric to evaluate the efficiency of local AI models. The study found that local models can accurately answer 88.7% of real-world queries and have shown a 5.3x improvement in IPW from 2023 to 2025. Local accelerators also demonstrated at least 1.4x lower IPW compared to cloud-based solutions, suggesting local inference can significantly offload demand from centralized infrastructure. AI
IMPACT Introduces a new metric to track the viability and efficiency of local AI inference, potentially shifting demand from cloud infrastructure.