science & techanalysisOur World in Data
Generative AI investment has surged because training models that write and draw takes enormous computing power
Generative AI refers to models that learn statistical patterns from huge collections of text, images, audio or code, then produce new material resembling them. Language models such as GPT-4 and Claude predict the next word in a sequence, which after training lets them write essays and answer questions. Image models such as Midjourney and DALL·E turn text descriptions into pictures. Building these systems requires huge numbers of calculations on specialised chips, so the cost of training and running them has pushed spending into very large sums, mostly from big technology companies. Investment has risen sharply as firms race to build and sell these tools, and whether that spending pays off remains an open question.
Models are trained by adjusting billions of numerical weights so their predictions match examples in the data. This requires running enormous numbers of calculations on graphics processors, so most of the expense is hardware and electricity.
The investment wave decides which companies lead AI, how much electricity data centres use and whether the productivity gains justify the cost. It is one of the largest capital flows in the modern tech economy.
