The Great Hallucination: Behind the AI Hype

In the past five years, over two trillion dollars have been invested in AI with roughly 75% going into US companies like Anthropic or OpenAI. But to date, there is still no viable business model for all this investment. Is the AI bubble about to burst?

The release of ChatGPT in November 2022 turned the world upside down. For the first time, millions of people experience a machine that not only strings sentences together, but also conducts conversations in real time, asks questions, writes texts and solves problems as a matter of course.  Within five days, ChatGPT had 1 million users; within two months, 100 million. And suddenly, a question that had long sounded theoretical arises: Could this be the moment when humanity has closed in on the great science fiction dream of “artificial general intelligence” (AGI)? Is this the big breakthrough for an AI that would surpass human capabilities everywhere?

For Silicon Valley, the release couldn’t have come at a better time, because the very place that makes a business out of thinking and designing futures had run out of ideas. Two years earlier Facebook had rebranded itself Meta projecting a great departure into the metaverse, a move that backfired spectacularly. The company lost two-thirds of its market value. Self-driving robo-taxis were still not ready for the market, and the market for digital collectible images, NFTs, had also collapsed. Were large language models (LLMs) like ChatGPT seemed to be ‘the next big thing’ everyone had been waiting for? 

Investors jumped on the new hype and in no time at all, hundreds of billions of dollars flowed into the feverish development of large language models (LLMs), such as ChatGPT. But it didn’t take long for a phenomenon to emerge that irritated many users: LLMs simply invented stories, people or facts that had never existed – they hallucinated. At the same time, new problems arose because the huge computational models ran out of data, the fuel from which they learn. By the end of 2024, the entire internet had been exhausted in terms of data. It became known that the language models had been trained illegally with copyrighted material. Instead of a huge business opportunity, there was a barrage of billion-pound lawsuits from authors, publishers, media companies and artists. 

Costs skyrocketed, not least because LLMs require enormous computing power for every single user query. This in turn requires data centers that rely on highly specialized chips, which require rare earths and metals, the mining of which transforms entire regions into lunar landscapes. Chip manufacturing itself is highly toxic and extremely harmful to the environment: it requires aggressive chemicals, produces large amounts of toxic waste and is only possible at a few locations worldwide, with Taiwan as the epicenter. New, so-called “giga chip factories” are supposed to get the problems under control – but their power consumption is so enormous that new power plants must first be built to cover it, most of them based on natural gas. And what is it all for? To date, there is no viable business model for large language models – because they hallucinate.

In the meantime, China has developed its own LLMs at a fraction of the cost incurred by the US and with a real business model behind them. So where’s Europe in this story? Does it want to follow the American path and dig a trillion-dollar investment grave? Or take the Chinese path of pragmatic upgrading of industrial goods?