Characterization of uncertainty (variance) in ecosystem projections under climate change is still rare despite its importance for informing decision-making and prioritizing research. We developed an ensemble modeling framework to evaluate the relative importance of different uncertainty sources for food web projections of the eastern Bering Sea (EBS). Specifically, dynamically downscaled projections from Earth System Models (ESM) under different greenhouse gas emission scenarios (GHG) were used to force a multispecies size spectrum model (MSSM) of the EBS food web. In addition to ESM and GHG uncertainty, we incorporated uncertainty from different plausible fisheries management scenarios reflecting shifts in the total allowable catch of flatfish and gadids and different assumptions regarding temperature-dependencies on biological rates in the MSSM. Relative to historical averages (19942014), end-of-century (20802100 average) ensemble projections of community spawner stock biomass, catches, and mean body size (±standard deviation) decreased by 36% (±21%), 61% (±27%), and 38% (±25%), respectively. Long-term trends were, on average, also negative for the majority of species, but the level of trend consistency between ensemble projections was low for most species. Projection uncertainty for model outputs from ?2020 to 2040 was driven by inter-annual climate variability for 85% of species and the community as a whole. Thereafter, structural uncertainty (different ESMs, temperature-dependency assumptions) dominated projection uncertainty. Fishery management and GHG emissions scenarios contributed little (<10%) to projection uncertainty, with the exception of catches for a subset of flatfishes which were dominated by fishery management scenarios. Long-term outcomes were improved in most cases under a moderate mitigation relative to a high business-as-usual GHG emissions scenario and we show how inclusion of temperature-dependencies on processes related to body growth and intrinsic (non-predation) natural mortality can strongly influence projections in potentially non-additive ways. Narrowing the spread of long-term projections in future ensemble simulations will depend primarily on whether the set of ESMs and food web models considered behave more or less similarly to one another relative to the present models sets. Further model skill assessment and data integration are needed to aid in the reduction and quantification of uncertainties if we are to advance predictive ecology.