The rapid pace of technological advances over the past year, especially in artificial intelligence (AI), has provided many reasons for optimism. But as we head into 2025, there are signs that AI’s momentum may be waning.

Since 2023, the dominant narrative has been that the AI revolution will drive productivity and economic growth, paving the way for extraordinary technological breakthroughs. PwC, for example, projects that AI will add nearly $16tn to global GDP by 2030 – a 14% increase. Meanwhile, a study by Erik Brynjolfsson, Danielle Li, and Lindsey R Raymond estimates that generative AI could boost worker productivity by 14% on average and by 34% for new and low-skilled workers.
Recent announcements by Google and OpenAI seem to support this narrative, offering a glimpse into a future that not long ago was confined to science fiction. Google’s Willow quantum chip, for example, reportedly completed a benchmark computation – a task that would take today’s fastest supercomputers ten septillion years (ten followed by 24 zeros) – in under five minutes. Likewise, OpenAI’s new o3 model represents a major technological breakthrough, bringing AI closer to the point where it can outperform humans in any cognitive task, a milestone known as “artificial general intelligence.”
But there are at least three reasons why the AI boom could lose steam in 2025. First, investors are increasingly questioning whether AI-related investments can deliver significant returns, as many companies are struggling to generate enough revenue to offset the skyrocketing costs of developing cutting-edge models. While training OpenAI’s GPT-4 cost more than $100mn, training future models will likely cost more than $1bn, raising concerns about the financial sustainability of these efforts.
To be sure, investors are eager to capitalise on the AI boom, with venture capital firms investing a record $97bn in US-based AI startups in 2024. But it appears that even industry leaders like OpenAI are burning through cash too quickly to generate meaningful returns, leading investors to worry that much of their capital has been misallocated or wasted. A back-of-the-envelope calculation suggests that a $100bn investment in AI would require at least $50bn in revenue to produce an acceptable return on capital – accounting for taxes, capital expenditures, and operating expenses. But the entire sector’s annual revenues, according to my sources, total just $12bn, with OpenAI accounting for roughly $4bn. In the absence of a “killer app” for which customers are willing to pay substantial sums, a significant portion of VC investments could end up worthless, triggering a decline in investment and spending.
Second, the enormous amounts of energy required to operate and cool massive data centres could impede AI’s rapid growth. By 2026, according to the International Energy Agency, AI data centres will consume 1,000 terawatt-hours of electricity annually, exceeding the United Kingdom’s total electricity and gas consumption in 2023. The consultancy Gartner projects that by 2027, 40% of existing data centres will be “operationally constrained” by limited power availability.
Third, large language models appear to be approaching their limits as companies grapple with mounting challenges like data scarcity and recurring errors. LLMs are primarily trained on data scraped from sources such as news articles, published reports, social media posts, and academic papers. But with a finite supply of high-quality information, finding new datasets or creating synthetic alternatives has become increasingly difficult and costly. Consequently, these models are prone to generating incorrect or fabricated answers (“hallucinations”), and AI companies may soon run out of the fresh data needed to refine them.
Computing power is also approaching its physical limits. In 2021, IBM unveiled a two-nanometre chip – roughly the size of a fingernail – capable of fitting 50bn transistors and improving performance by 45% compared to its seven-nanometre predecessor. While undeniably impressive, this milestone also raises an important question: Has the industry reached the point of diminishing returns in its quest to make ever-smaller semiconductors?
If these trends persist, the current valuations of publicly traded AI companies may not be sustainable.
Related Story