Why So Much Money Is Being Spent on AI

Summary

This opinion essay argues that the scale of current AI spending is driven less by a credible path to artificial general intelligence than by a structural feature of neural networks: they can almost always be given more neurons, layers, training epochs, data, and hardware. Conventional software generally reaches a point where enough hardware has been purchased to run it adequately, while neural-network systems can plausibly be expected to improve with further investment. Earlier neural-network applications such as translation or speech recognition could eventually reach an acceptable performance level, but the current generation of hyperscaler-backed systems is pursuing AGI, which the author describes as the ability to think, solve unfamiliar problems, and avoid confidently producing nonsense. Because the author believes large language models cannot reach that goal, there is no clear point at which companies can declare that they have enough data centers. The essay says the spending is amplified by the wealth of companies able to sustain losses at enormous scale and by the reputational cost of admitting failure after so much money has been committed. In the author’s view, this creates a feedback loop in which additional hardware is repeatedly justified as the path to success, potentially expanding until it strains the semiconductor sector and the finances of major technology companies. The author acknowledges that LLMs are useful for some tasks, but argues that their usefulness will not produce the level of productivity needed to justify the investment.