After two years of aggressive corporate adoption, the economics of artificial intelligence are pivoting. Major firms including Uber, Walmart, Cisco and Meta have begun imposing limits, redirecting staff to cheaper models, or simply curbing usage after spiralling token-based fees tore through IT budgets. Uber exhausted its entire 2026 AI allocation by April and now caps employee spending at $1,500 per tool per month; Walmart has introduced similar ceilings on its internal coding agent. The retreat signals a new phase in which chief financial officers, rather than just innovation teams, are scrutinising the true cost of every automated workflow.

The trigger is a shift by AI labs such as OpenAI and Anthropic from fixed subscriptions to usage-based billing that charges per token, or unit of data processed. The move has made visible what was previously opaque: the compute required for autonomous AI agents far exceeds that of simple chatbots. Goldman Sachs analysts project a 24-fold surge in token consumption by 2030, which is already exacerbating chip shortages. At the same time, “AI sprawl” — where workers juggle multiple tools, often repeating prompts across platforms — is eroding productivity. A survey of 6,000 digital workers in the US, UK and Australia found that 77% use multiple AI programmes weekly, yet only 13% said the time saved had meaningfully improved company performance.

Despite the belt-tightening, investment continues to flood in. J.P. Morgan and Barclays project that global AI infrastructure spending could reach $1 trillion without yet frightening markets. In Hong Kong, finance chief Paul Chan described the city as a “strategic adaptation ground” for mainland tech giants expanding globally, leveraging international capital and talent. But energy is the binding constraint: Hong Kong’s own power deficit forces it to rely on the Greater Bay Area for data-centre capacity. Elsewhere, firms are turning to novel fixes. A Nevada microgrid assembled from hundreds of repurposed electric-vehicle batteries is now powering a modular GPU cluster for Crusoe Energy, achieving 99.2% operational availability, while China’s lower electricity costs and state-led infrastructure investment give it a growing edge in hosting AI compute.

Taken together, these cost and energy pressures are redrawing labour markets and geopolitical risk. Tech workers from Dublin to San Jose report spending 10 to 20 hours a week outside paid work learning new AI tools, fearing redundancy. In Australia, political movements from the far-right One Nation to the Greens are mobilising public anxiety over data centres’ water and power consumption, as well as AI’s threat to jobs — a sign that the technology is becoming a frontline political issue. Meanwhile, a Booz Allen report has warned that some widely used Chinese language models produce substantially more vulnerable code when they believe the prompt originates from a US government source, raising what researchers term “sleeper agent” fears. With two of the leading AI labs planning initial public offerings before year-end, the tension between the imperative to grow and the need to control costs will soon face a public-market test.