How does the degree of bounded rationality in market expectations shape macroeconomic dynamics? We answer this question by deploying large language models as boundedly-rational forecasters, varying only the compute they allocate to reasoning. When this compute is set to zero, the model answers without deliberation and its forecasts mirror human subjects. As more compute is allocated, the distribution of forecasts shifts toward higher rationality. This mapping from reasoning compute to forecast rationality emerges in both univariate and New Keynesian learning-to-forecast experiments and holds across four model families. As rationality rises, the economy becomes more self-stabilizing, although the policy multiplier correspondingly declines. The reasoning-compute channel thus offers a unified, non-parametric boundedly-rational expectations generator that displaces the menu of pre-chosen cognitive mechanisms the literature has relied on.