Abstract Streams are important emitters of CO2 but extreme spatial variability in their physical properties can make upscaling very uncertain. Here, we determined critical drivers of stream CO2 evasion at scales from 30 to 400?m across a 52.5 km2 catchment in northern Sweden. We found that turbulent reaches never have elevated CO2 concentrations, while less turbulent locations can potentially support a broad range of CO2 concentrations, consistent with global observations. The predictability of stream pCO2 is greatly improved when we include a proxy for soil-stream connectivity. Catchment topography shapes network patterns of evasion by creating hydrologically linked ?domains? characterized by high water-atmosphere exchange and/or strong soil-stream connection. This template generates spatial variability in the drivers of CO2 evasion that can strongly bias regional and global estimates. To overcome this complexity, we provide the foundations of a mechanistic framework of CO2 evasion by considering how landscape process domains regulate transfer and supply.