A November 2025 study from Stanford and Together AI introduces a metric called 'intelligence per watt' and finds that local, efficient models handled 71% of queries by 2025 with 88.7% accuracy, up from 23% in 2023. The author argues that US export controls and data-center buildout focus on frontier compute, but the structural bottleneck is grid power and efficiency gains that outpace infrastructure. For a financial institution, this implies that the expected returns on hyperscaler data-center capex may be overstated if demand shifts to efficient local inference instead of centralized compute.
268 other tracked developments touching the same claims.