It remains unclear what long-term effects of substituting carbohydrates at the expense of protein or fat may have with regard to diabetes risk. Our objective was to evaluate carbohydrate intake in predicting type 2 diabetes using substitution models for fat and protein. We conducted a prospective cohort study of 9702 men and 15 365 women aged 35–65 years and free of diabetes at baseline (1994–8) who were followed for incident type 2 diabetes until 2005. Dietary intake of macronutrients was estimated with a validated FFQ. We estimated the relative risk (RR) using Cox proportional hazards analysis. During 176 117 person-years of follow-up we observed 844 incident cases of physician-confirmed type 2 diabetes. After adjustment for age, BMI, waist circumference, potential lifestyle and dietary confounders, substituting 5 % of energy intake from total, saturated, or monounsaturated fat with carbohydrates was not associated with diabetes risk. In contrast, substituting carbohydrates for protein or PUFA was inversely related to diabetes risk (RR for 5 % energy substitution of protein 0·77 (95 % CI 0·64, 0·91); RR for PUFA 0·83 (95 % CI 0·70, 0·98)). These associations appeared to be similar for men and women, but gained statistical significance only among men for protein (RR 0·78 (95 % CI 0·61, 0·99)). Restricted cubic spline regression did not indicate non-linearity of these associations (P for non-linearity in full cohort was 0·353 and 0·349). In conclusion, a higher carbohydrate intake at the expense of protein and PUFA might be associated with decreased diabetes risk.