The United Nations’ flagship World Economic Situation and Prospects report, released on Wednesday, also highlighted how massive investment in the US AI sector has surged in recent years. Investment in AI rose from $87 billion in 2022 to $93.1 billion in 2023 before jumping to $150.6 billion in 2024.
Jefferies noted that capital expenditure, which is the spending on assets linked to artificial intelligence, was the largest contributor to US economic growth in 2025, second only to personal consumption.
“The macro vulnerability to a sudden collapse in AI capex is clear, since that was the main driver of US economic growth last year after personal consumption,” Jefferies said, citing national accounts data.
According to the report, US real GDP expanded by $438 billion, or an annualised 2.5 per cent, in the first three quarters of 2025, with a significant share of this growth driven by investment in data centres, semiconductor manufacturing, and related digital infrastructure.
The brokerage cautioned that by mid-2026, investors are likely to start questioning the sustainability of returns from AI-related capital expenditure. It also flagged the risk of potential excess capacity in data centres which could weigh on corporate balance sheets and dampen future investment flows.
Similar risks could spill over into the US energy sector, which rallied last year on expectations of rising electricity demand driven by data centres and AI-related power needs.
While Jefferies acknowledged that AI agents could unlock significant commercial opportunities in the future, it cautioned that large investments have already been front-loaded into the current cycle. In the event of an over-investment bust, costs related to AI “inference” could fall sharply — mirroring the dot-com era, when excess fibre-optic capacity led to a collapse in broadband costs and a surge in digital commerce.
The brokerage added that the AI investment cycle is inherently more vulnerable to abrupt reversals because AI chips have a much shorter technological life compared with fiber-optic cables. The technological life of chips is around three to four years, while fiber-optic cables last nearly 25 years, making AI-related investment cycles structurally more prone to sudden downturns.
