Renewable-energy forecasting studies usually rank models by predictive error, although model development and deployment also consume electricity and produce greenhouse-gas emissions. This study asks whether additional predictive sophistication creates positive net carbon value after computational emissions are deducted. The empirical analysis uses the frozen October 6, 2020 release of Open Power System Data for Germany, covering hourly solar and wind generation from January 2015 through September 2020. Generation is normalized by contemporaneous installed capacity. Persistence, ridge regression, random forest, and extremely randomized trees are evaluated for one-hour and 24-hour forecasting under a strict temporal design, with 2020 reserved for out-of-time testing. Forecast error is observed directly. Computational electricity is estimated from measured wall-clock runtime, a documented 65 W processor-power assumption, and a power-usage-effectiveness factor of 1.2. Operational emissions benefits are not treated as observed. They are estimated under low, central, and high scenarios that translate reductions in absolute forecast error into changes in balancing energy. Ridge regression generated the lowest mean absolute error in three of four tasks; persistence remained best for 24-hour solar forecasting. Its strongest advantage occurred for one-hour solar forecasting, where mean absolute error fell from 0.0321 under persistence to 0.0081. Greater algorithmic complexity did not produce better results. Tree ensembles consumed more training energy and underperformed ridge regression, and no learned model improved 24-hour solar mean absolute error relative to persistence. Under the central operational scenario, ridge regression also produced the highest net carbon value. The ranking was driven primarily by operational assumptions, however, because estimated computational emissions were small for the tested data and hardware scale. The findings support carbon-aware model selection based on practical accuracy, deployment burden, and operational value rather than accuracy alone.