Investors have already taken a severe beating, but far more stock market carnage lies ahead as the AI investment boom wanes, according to Capital Economics, the independent economic research consultancy.

In a note published on Friday, Chief Economic Adviser Vicky Redwood said artificial intelligence still holds potential to boost productivity, but she is bearish on the amount of revenue that can be generated to sustain the current pace of investment.

A handful of hyperscalers (the largest cloud-infrastructure operators, including Microsoft, Alphabet, Amazon and Meta) are expected to spend $800 billion this year—double the total from a year ago—and to invest an estimated $1 trillion or more in 2027. Many forecasts assume annual capital expenditures will remain above $1 trillion for several years.

Redwood calculated that annual AI-related revenues earned from external customers would have to explode tenfold by 2030 to justify all of that spending. While that is possible, it would imply a far more severe and rapid disruption to the economy than is currently visible, including major displacement of existing businesses and labor. That gap between spending and revenue is the crux of her caution—and it is one that will be tested in public each quarter, as the hyperscalers update their capital-spending plans in earnings reports.

As a result, she expects investment to continue rising next year, though at a slower pace, before falling outright in 2028 by around 10%.

Higher AI investment added 0.3 percentage point to U.S. GDP growth last year, Redwood noted, so a capex slowdown should trim roughly that much from the economy's recent growth rate, while a 10% drop in capital spending might take a further 0.1% off GDP.

That would amount to only a slight haircut for the United States, especially compared with more AI-dependent economies in Asia. Taiwan's real GDP growth next year should be around 3 percentage points lower, and South Korea's about 1 point lower, with a further drag in 2028—an exposure rooted in the two countries' central roles in the global semiconductor supply chain, from leading-edge chip fabrication to memory production.

The stock market, meanwhile, will price in the looming AI capex decline even sooner. Redwood expects the S&P 500 to be 12% below its current level by the end of 2027, putting the index 17% off its early-October peak.

“We expect considerably larger falls in tech-heavy indices,” she added. “A tightening of financial conditions driven by factors such as higher energy prices or term premia could also make financing investment hard.” (Term premia are the extra yield investors demand to hold longer-dated bonds.)

Lower stock prices would also feed into the economy through the wealth effect, which fuels consumer spending. The market's rally over the past year lifted GDP by around 0.4%, according to Redwood, so a stock selloff should produce an economic drag of similar scale.

Despite the warnings on capex, GDP and stocks, AI should continue to work its way into the economy and lift productivity, helped by infrastructure that has already been built and new investment still to come, albeit at a slower rate. Redwood does not expect an abrupt collapse in AI investment or a severe economic downturn when the boom ends.

Capex could remain relatively high even after declining, she pointed out, as hyperscalers keep spending to replace depreciating equipment, exploit economies of scale, and stay ahead in the “winner-takes-most” AI race. They could also spread any investment declines over several years rather than all at once.

“The bottom line for the U.S. economy is that we expect slower GDP growth in 2028, but not a major downturn,” Redwood wrote. “Admittedly, recessions have sometimes followed previous capex booms and busts, including in the early 2000s after the dotcom bust. However, a fall in capital spending alone has generally not been sufficient to cause a recession.” The dotcom bust followed a comparable stretch of heavy infrastructure building, including telecom carriers' fiber-optic overbuild, before capital spending collapsed.

At the same time, AI adoption should continue to expand, boosting productivity and supporting output growth. Redwood estimated that AI will increase U.S. productivity by about 1% per year for several years.

“Rather, we expect investment to adjust to a more sustainable pace, with sufficient capacity still being added to support further growth in AI usage,” she concluded.

This story was originally reported by Fortune.