AI drug discovery is getting expensive to train — researchers propose ways to cut computational costs by 90%

Training machine learning models to discover new drugs and materials requires massive amounts of quantum computing data, which is increasingly costly and energy-intensive. Researchers are proposing specific techniques — using simpler models first, retraining smaller versions of larger models, and letting AI actively choose which experiments to run — that could cut the compute burden dramatically while maintaining accuracy.

Right now, the bottleneck in AI-driven drug and material discovery isn't the ideas — it's the electricity and hardware needed to train the models. If these efficiency techniques work at scale, they could make drug discovery accessible to smaller labs and companies instead of just the richest institutions, while also making the whole process less wasteful. The catch: this is still mostly theoretical. The real test is whether these methods hold up when companies and research labs actually deploy them on real projects, not just benchmarks.