Large-scale deep learning training requires substantial energy consumption, driven by many factors: model size, training duration, number of experiments, and the electricity source that powers data centers. Public analyses show that compute demand for state-of-the-art models has grown rapidly, increasing the total electricity required for training. OpenAI reports a long-term trend of exponential growth in compute used to train the largest models, which has direct implications for energy use and costs.
Causes and scale
The primary technical causes are larger neural architectures, longer pretraining on massive datasets, and extensive hyperparameter or architecture searches that multiply compute runs. These multiply electricity consumption because each experiment may require thousands of GPU or TPU hours. Emma Strubell, Ananya Ganesh, and Andrew McCallum University of Massachusetts Amherst examined natural language processing experiments and found that certain large-scale searches can produce greenhouse gas emissions comparable to the lifetime emissions of several cars, highlighting how methodological choices amplify energy and carbon impacts. Model efficiency varies widely; two similarly capable models can have very different energy footprints depending on search strategy and hardware.
Environmental and social consequences
Energy costs translate into direct financial cost for researchers and companies and into environmental cost through carbon emissions, which depend heavily on regional electricity mixes. Training in regions that rely on coal yields higher emissions than running the same computation where grids are low-carbon. There are also territorial and cultural dimensions: data centers and high-performance computing facilities cluster in particular regions, affecting local water use for cooling and local economies. The concentration of compute resources favors well-funded organizations, shaping who designs and benefits from these models and raising questions of equity in research access.
Efficiency improvements in hardware, software, and scheduling can reduce costs and emissions: more efficient chips, mixed-precision training, algorithmic advances that require fewer training steps, and shifting workloads to low-carbon times or locations. Reporting and benchmarking energy use for model training enables comparisons and encourages responsible choices. As compute demand continues to grow, transparent measurement and targeted reductions will determine whether large AI models scale sustainably across environmental and social contexts.