Will the bubble burst for AI in 2025, or will it start to deliver?

It may be the biggest gamble in business history. Today’s mania for artificial intelligence (AI) began with the launch of ChatGPT at the end of November 2022. OpenAI’s chatbot attracted 100m users within weeks, faster than any product in history. Investors also piled in. Spending on AI data centres between 2024 and 2027 is expected to exceed $1.4trn; the market value of Nvidia, the leading maker of AI chips, has increased eight-fold, to more than $3trn.

And yet most companies are still not sure what the technology can or cannot do, or how best to use it. Across the economy, only 5% of American businesses say they are using AI in their products and services. Few AI startups are turning a profit. And the energy and data constraints on AI model-making are becoming steadily more painful. The disparity between investor enthusiasm and business reality looks untenable—which means 2025 is shaping up to be a crunch year. The race to make AI more efficient and more useful, before investors lose their enthusiasm, is on.

Start at the cutting edge of innovation. Several constraints are slowing the pace at which the technological frontier is being pushed out. Training big models needs huge amounts of energy. The electricity used to train GPT-4, the large language model underpinning ChatGPT, could have powered 5,000 American homes for a year; the equivalent figure for GPT-3, its predecessor, was 100. Developing ever larger and whizzier models thus requires ever deeper pockets. By some estimates, the next generation of models could cost $1bn to train; and the larger they become, the more the cost of querying them (known as “inference”) will mount. Meanwhile, there is a looming shortage of training data. By one estimate, the stock of high-quality textual data on the internet will have run out by 2028.

Companies around the world are rushing to come up with clever fixes to these problems, from more efficient and specialised chips to more specialised and smaller models that need less power. Others are dreaming up ways of tapping new high-quality data sources such as textbooks, or generating synthetic data, for use in training. Whether this will lead to incremental improvements in the technology, or make the next big leap forward affordable and feasible, is still unclear. Investors have poured money into superstar firms like OpenAI. But in practice there is not much difference in performance and capabilities between the flagship models offered by OpenAI, Anthropic and Google. And other firms including Meta, Mistral and xAI are close behind.

For end users of AI, a different kind of struggle is under way, as individuals and companies try to work out how best to use the technology. This takes time: investments need to be made, processes rethought and workers retrained. Already some industries are further ahead in adopting AI than others: a fifth of information-technology firms, for instance, say they are using it. As the technology becomes more sophisticated—such as with the arrival in 2025 of “agentic” systems, capable of planning and executing more complex tasks—adoption may accelerate.

But culture also matters. Although few firms tell statisticians they are using AI, one-third of employees in America say they are using it for work once a week. In some roles the figure is even higher. One study found that 78% of software engineers in America are using AI at least weekly, up from 40% in 2023, as are 75% of human-resources staff, up from 35%. And OpenAI says 75% of its revenue comes, tellingly, from consumers rather than from corporate subscriptions.

All of this suggests that much use of AI is in secret, as workers use it to streamline tasks such as rewriting text or generating reports. Employees may worry that if they admit to using AI to get things done more quickly, bosses will give them more work to do, or take this as a signal that fewer workers are needed. This in turn suggests that AI adoption is as much a management challenge as a technological one. To get the most out of the technology, bosses need to create an environment that incentivises openness and experimentation, rather than secrecy and suspicion.

And AI can be used for more than office pencil-pushing. It may be that, in 2025, the most prominent AI breakthroughs come in other areas, such as drug development (the first AI-derived drugs may go into stage-three clinical trials) or defence (as intelligence is added to drones, which are emerging as key weapons systems of the future). Indeed, the West worries that China will harness AI to gain a military and economic advantage. Ironically, Chinese engineers have become particularly adept at innovating around resource constraints, in part because American export controls have curtailed their access to cutting-edge AI chips.

The AI race, then, will take many forms in 2025. Yet the point at which investors lose their nerve is often when new technologies quietly start gaining traction. Will the bubble burst, or will the technology start to deliver? The answer in 2025 may be: a bit of both.