Oops, something went wrong
Corporate AI spending has surpassed the peak of the late 1990s dotcom boom in scale relative to the US economy, but a backlash against soaring costs is beginning to emerge as companies discover that encouraging staff to use AI freely is proving far more expensive than anticipated, according to Jefferies.
Investment in US information processing equipment and software hit a record 4.91% of gross domestic product in the first quarter of 2026, exceeding the previous peak of 4.46% reached at the height of the dotcom bubble in 2000, with AI-related spending contributing 1.34 percentage points to annualised US economic growth of 2% in the same period.
Yet Jefferies, writing in its Greed & Fear note, flags growing signs of a corporate rethink, pointing to reports that Microsoft has begun cancelling internal licences for Claude Code, an AI-powered coding tool made by Anthropic, and reverting to its own GitHub Copilot product.
Uber's chief technology officer reportedly warned internally that the company burned through its entire 2026 AI budget in just four months.
The culprit is a phenomenon Jefferies dubs "tokenmaxxing", a term derived from the way AI models charge for their output, measured in units called tokens, similar to how a printer charges per page.
As companies rewarded staff for high AI usage through internal leaderboards, employees began running up AI bills on unnecessary tasks simply to inflate their scores, generating costs without productivity gains.
Cloud service providers are meanwhile raising prices by around 30% year on year, compounding the pressure on corporate budgets at a time when the cost of AI output per token has been falling much faster than the underlying cost of computing power.
Jefferies notes that Anthropic, the AI company behind Claude, has seen its annualised revenue run rate surge from $9 billion at the end of 2025 to over $44 billion by early May 2026, illustrating both the scale of corporate AI adoption and the associated cost burden being absorbed across the economy.
Despite the cost pressures, there is no sign of the broader infrastructure spending wave slowing, with the Philadelphia Semiconductor Index, a gauge of chip company stocks, up 80% year-to-date and trading 64% above its 200-day moving average.
Jefferies' base case is that a significant amount of capital will ultimately be destroyed in the current AI spending cycle, drawing a parallel with the dotcom boom-bust and the 19th-century railway mania.
The bank also invoked Amara's Law, the principle that people tend to overestimate a technology's short-term impact while underestimating its long-term consequences.
The biggest winners so far remain the makers of memory chips, with SK Hynix and Micron both surpassing $1 trillion in market capitalisation this week.
Sign in to access your portfolio
Leave a Reply