Elevated carbon dioxide (eCO2) acts as a fertiliser for photosynthesis, driving an increase in gross primary production (GPP). However, it is unclear how effectively increased GPP propagates along the ‘carbon (C) cascade’ to increase net primary production (NPP) and vegetation C stocks (Cveg) in different plant compartments. Vegetation models were criticised for being overly sensitive to photosynthesis (source-driven), neglecting sink-driven processes which may attenuate (or amplify) changes in NPP and vegetation C stocks. Here, we introduce an analytical framework to diagnose linearity (L) as ratios of relative changes in linked fluxes and pools. We then apply this framework to 16 models of the TRENDY v11 ensemble and to observation-based estimates of CO2 sensitivities. We found widely varying global patterns in L across models. Six models showed a majority of grid cells with larger relative changes in NPP than in GPP (LNPP:GPP textgreater 1 for textgreater 60% of gridcells), indicating increased vegetation carbon use efficiency under eCO2. Only three models had LNPP:GPP textless 1 for textgreater 60% of gridcells. Four models showed a majority of gridcells with larger relative changes in estimated steady-state Cveg than in NPP, while five models showed the opposite—in both cases with a large spread of LCveg*:NPP across grid cells within models. Observations-based analysis reveals median LCveg*:NPP textless 1 overall but evidence is insufficient for a conclusion. Three models showed a larger relative increase in root C than in Cveg, (LCroot:Cveg textgreater 1) while five models showed the opposite. Most field evidence shows LCroot:Cveg textgreater 1. Widely differing distributions of L among models and links in the C cascade reveal a strong influence of nonlinear behaviour in individual models. However, due to the spread in L, across the whole models ensemble, L deviations from 1 were roughly balanced, leading to an overall linear behaviour of terrestrial C cycle representations in the multi-model-mean.