Between them, Microsoft, Meta, Amazon and Google are expected to spend more than $600 billion on AI infrastructure this year, and global capital spending on the chips, data centers and power behind it looks set to top $1 trillion, more than the entire oil and gas industry lays out. The Economist reckons this is on course to be the largest investment boom in history, bigger than the dotcom bubble, bigger even than the railways. All of it rests on a single assumption: that generative AI is about to deliver a historic leap in economic productivity. A run of fresh analysis this week suggests the leap is proving surprisingly hard to see.
Start with how AI is actually used. ChatGPT has passed a billion active users, yet Gallup finds only 15 percent of US employees use AI at work every day, and a survey of some 6,000 executives put average usage at just 1.5 hours a week. Company spending tells the same story. The payments firm Ramp reports that while the top 1 percent of businesses spend around $7,500 per employee per month on AI, the median is $12. Adoption, as the Bennett School of Public Policy at Cambridge puts it, is "broad but shallow."
Depth turns out to be the whole game. A PYMNTS Intelligence survey of 60 large US enterprises found that the average company now uses AI across seven of eight business functions, but only one deployment in five is genuinely embedded in how the work is done. Firms that wove it into three or more functions overwhelmingly report a return; those dabbling in only one or two mostly do not. Spreading AI thinly, in other words, buys the logo on the strategy slide but not the payoff underneath it.
Where the payoff has been measured rigorously, it tends to shrink. Software development is the clear exception, with credible studies showing 20 to 40 percent gains, though even there the benefit collapses on complex legacy code. Outside coding the picture is sobering. A UK cross-government trial claimed civil servants saved 26 minutes a day, but a controlled follow-up found the savings far smaller, and one department found no gain at all, with some tasks actually slowed by the need to fix poor output. Canada's tax-agency chatbot answered correctly about a third of the time in an external audit, against the 70 to 90 percent claimed internally. KPMG finds just 7 percent of business leaders reporting a positive return so far.
The rush has even manufactured its own drag. Companies that tied performance reviews to AI usage, treating logins like sales quotas, are quietly dropping the practice; as one CTO put it, rewarding raw usage is "a really stupid way to do anything." Employees churn out "workslop," polished but hollow documents that reportedly cost each colleague nearly two hours to untangle. One widely shared figure holds that 95 percent of companies deploying generative AI have seen no meaningful revenue growth at all.
None of this means the boom is a mistake. The most encouraging result of the week, a UK Citizens Advice copilot that halved advisers' response times in a randomized trial, points to what works: AI aimed at a specific task, with a human kept firmly in the loop, backed by real changes to how the work is organized. The uncomfortable lesson is that the technology is not plug and play. To paraphrase the economist Robert Solow, AI is everywhere except in the productivity statistics. The trillion dollars is a bet that this is only a matter of time. History, where big technologies have taken decades to show up in the numbers, counsels patience, and rather more humility than the spending implies.