People walk past an AI sign at the All In artificial intelligence conference in Montreal, on Sept. 28, 2023.Ryan Remiorz/The Canadian Press
When generative artificial intelligence burst on to the market in late 2022, the technology seemed to be magic. A user could type in a few words, and out would come an automatically generated piece of text or an image that was, if not perfect, often good enough. It promised to quickly make a number of professions – from writers to paralegals – more efficient, if not obsolete.
Almost two years later, generative AI now resembles not-so-much a fairy godmother as a ravenous beast gorging on investments, energy and intellectual property, and creating little of value in return.
The companies behind this technology need to dramatically rein in their ambitions and deal with the enormous costs and large carbon footprint, or otherwise demonstrate their technology can live up to the economy-transforming promises they’ve made.
First is the fiscal cost. Investors have, in recent months, become more concerned with just how quickly spending on AI infrastructure has gone up. Dario Amodei, chief executive officer of Anthropic, which has received investments from Amazon and Google, told Time that the cost of training generative AI models had increased from the tens of millions of dollars in past years, to US$1-billion this year, and expected it to rise to US$10-billion in the near future.
Goldman Sachs released a research report in late June saying that companies would spend a combined US$1-trillion in coming years on capital expenditures, such as data centres, computer chips and electricity.
Jim Covello, Goldman Sachs’ head of global equity research, said in the report that there is no evidence yet that the tech companies making generative AI – or the companies using it – had seen any noticeable increases in revenue despite the spending. The only companies making profit, he noted, were those supplying the infrastructure.
Mr. Covello also did not expect those costs to come down any time soon, as the number of suppliers – such as chipmaker Nvidia – is so small that they have the market power to keep their prices high.
It would be one thing if it was only investor money that was being shovelled into the furnace. (Not a great thing for those investors, mind you, or anyone with a pension plan buying into the hype.)
But generative AI is quickly proving to have a nearly unquenchable thirst for water and power.
The International Energy Agency says data centres are a significant driver of higher energy usage worldwide. Data centres used 460 terawatt-hours of electricity in 2022 and, with demand growing, are projected to consume 1,000 terawatt-hours in 2026 – roughly equivalent to Japan’s annual energy needs.
Those electronics run hot and need water to cool down. AI computing could drink up 6.6-billion cubic metres of water by 2027, equivalent to the needs of six Denmarks, according to one study. Those researchers had previously estimated that generative-AI tool ChatGPT used half a litre of water for every 10 to 50 queries it received – the equivalent of repeatedly dumping out bottles of water on the ground as you work.
AI “is one of the most energy-intensive, water-intensive, mineral-intensive infrastructures at planetary scale that we have ever built as a species,” AI researcher Kate Crawford recently told The Globe and Mail’s podcast Machines Like Us. “It is truly astronomical.”
All this comes, of course, as Earth continues to warm and countries work to reduce their greenhouse gas emissions – not find innovative new ways to increase them. Higher demand from air-conditioning during relentlessly hotter summers will put increased strain on power grids, and water will become a more scarce resource as droughts become more common and widespread. When looking at the enormity of environmental challenges humanity has coming up, a chatbot tool does not look like it belongs high on the priority list.
There may be focused applications of the technology that can be designed to be more efficient and use fewer resources, while still providing benefits. For example, there is promise in generative AI that cuts down on voluminous paperwork for busy professionals, such as doctors and lawyers, or helps coders speed up their work.
But as a general-use tool, the magic of generative AI so far seems to be little more than an enormously expensive parlour trick.