From Fortune: 2024-04-04 08:18:37

Training AI models like GPT-4 and Gemini is an expensive business, with costs possibly exceeding $100 million for some models. The amount of computational power needed to train AI models has doubled roughly every six months since 2010. Hardware costs for training the most expensive AI models could reach $140 billion by 2030, posing challenges for sustainability.

As AI models become larger and more capable, the cost of training them continues to rise exponentially, potentially reaching the point where no company can bear the cost. The growth in computational power required for training AI models is predicted to outpace efficiency gains, raising concerns about the sustainability of the current trend. Companies may need to explore alternatives like using smaller, fine-tuned models for specific use cases to mitigate rising training costs.

Hyperscalers like Microsoft are exploring new energy sources like small modular nuclear reactors for data centers powering AI models, as the demand for computational power increases. Resistance to the expansion of data centers due to environmental concerns is growing, posing a challenge to the scalability of AI infrastructure. The future impact of training costs on the development of artificial general intelligence (AGI) remains uncertain, with potential benefits and challenges to consider.



Read more at Fortune: The cost of training AI could soon become too much to bear