Artificial intelligence was presented as the ultimate efficiency machine, a technology that would remove repetitive work, cut costs and deliver unprecedented gains in productivity. Executives, investors and consultants repeated the promise with remarkable confidence: software would replace expensive employees, organisations would become leaner and economic output would increase. Yet a striking reversal is now taking shape across the technology industry. The machines that were supposed to reduce costs are generating expenses so large that many companies are confronting an uncomfortable reality. In a growing number of cases, artificial intelligence is proving more expensive than the people it was meant to replace.
Signs of this reversal are emerging from the very companies that championed automation most enthusiastically. Writing in Forbes, Tim Bajarin observed that businesses are increasingly shifting spending towards artificial intelligence computing workloads and away from traditional labour costs. Bajarin cited comments from Nvidia executive Bryan Catanzaro, who noted that compute expenditure had exceeded workforce costs in parts of his operation. Such remarks are significant because they come from industry insiders rather than critics. The assumption that automation automatically lowers costs begins to weaken when its strongest advocates acknowledge that machines are consuming resources faster than human workers.
At the heart of this shift lies the economics of computation. Large language models depend on vast data centres, specialised processors, electricity, cooling systems and continuous infrastructure expansion. Every query, prompt and automated task carries a cost. What appears effortless on a computer screen is supported by an industrial system of enormous scale. The perception that artificial intelligence exists in a weightless digital world obscures the reality that every token processed requires energy, hardware and investment.
Recent reporting has highlighted the emergence of what some technology workers call “tokenmaxxing”, a culture in which employees are encouraged to maximise artificial intelligence usage regardless of whether the results justify the expense. Coverage in Business Insider described concerns within major firms that usage statistics were increasingly being treated as measures of productivity. When organisations reward activity rather than outcomes, waste can easily masquerade as efficiency. Employees may use artificial intelligence for routine or trivial tasks, generating substantial computational costs without creating corresponding value.
The consequences are becoming visible in company accounts. Reporting in Forbes and The Verge noted that Uber exhausted its annual artificial intelligence budget within months after expanding the use of coding assistants and related tools. Executives openly questioned whether rising token consumption was translating into better products, stronger services or measurable financial returns. Productivity is not determined by the number of interactions with a machine but by whether those interactions create genuine economic value.
A deeper irony lies beneath the surface. Many organisations justified workforce reductions by arguing that artificial intelligence would eventually take over tasks previously performed by employees. Yet reports discussed by Neil C. Hughes in Cybernews suggest that some firms are discovering that advanced artificial intelligence systems can cost more than the labour they were intended to replace. Rather than eliminating expenses, businesses often face rising subscription fees, infrastructure demands and continuing requirements for human oversight.
This contradiction reflects a familiar pattern in the history of technology. New innovations are often accompanied by ambitious promises before economic realities emerge. Railways, telecommunications networks and internet platforms all passed through periods of intense speculation. The difference is that those technologies eventually demonstrated clear productivity gains. Artificial intelligence occupies a more uncertain position. It can generate text, summarise information and automate selected tasks, but evidence of broad economic transformation remains limited.
The economist Joseph Schumpeter described capitalism in Capitalism, Socialism and Democracy as a process of creative destruction, in which older systems are replaced by more productive ones. The concept is frequently used to defend technological disruption. Yet disruption alone is not progress. If a new system consumes more resources while delivering uncertain benefits, destruction ceases to be creative and becomes merely costly. There is a growing risk that parts of the artificial intelligence boom are moving in that direction, with vast sums invested in infrastructure before clear returns have been demonstrated.
Public discussion has also been shaped by the assumption that large-scale job displacement is inevitable. Debates on BBC Question Time have often treated the replacement of workers by artificial intelligence as a future certainty rather than a proposition to be tested against evidence. Yet the evidence remains mixed. Research associated with Massachusetts Institute of Technology has shown that some forms of automation are not always economically viable when compared with human labour. Matt Stoller, writing in Think of X, has argued that artificial intelligence markets frequently reward consumption metrics rather than productive outcomes. A technology can be impressive in theory while remaining inefficient in practice.
The environmental implications add another layer of concern. Every expansion of artificial intelligence infrastructure requires more electricity, more water and more construction. Data centres have become strategic assets precisely because computation carries substantial physical costs. Public discussion often treats digital services as though they exist without material constraints, but behind every artificial intelligence application lies a network of energy systems, supply chains and industrial infrastructure.
The social consequences are equally significant. Workers increasingly find themselves defending their value against systems that still depend heavily on human oversight. Errors, fabricated information and inaccuracies continue to require correction by skilled employees. In The Age of Surveillance Capitalism, Shoshana Zuboff argued that digital technologies often evolve into systems of extraction. Her warning appears increasingly relevant as artificial intelligence companies draw upon vast quantities of data, attention and capital while concentrating many of the benefits within a small number of firms.
Investors are beginning to recognise these tensions. For years, financial markets rewarded almost any company associated with artificial intelligence, often regardless of profitability. That period appears to be ending. Shareholders are increasingly demanding evidence of return on investment rather than demonstrations of technological novelty. Boards want measurable results: savings, new revenue streams or genuine operational improvements.
None of this suggests that artificial intelligence lacks potential. The technology will continue to improve and may eventually deliver many of the efficiencies its supporters promised. However, the assumption that its success is inevitable has been weakened. Economic realities have not disappeared simply because algorithms have become more sophisticated. Businesses cannot spend indefinitely without demonstrating value. The central lesson of the present moment is straightforward: human labour was declared obsolete long before the economics of replacement had been properly tested. Until artificial intelligence consistently creates more value than it consumes, the paradox will remain: the machines designed to replace people are proving more expensive than the people themselves.
