The unbridled corporate experiment with generative AI is being checked by a swift tightening of budgets. Companies that rushed to put AI tools into employees’ hands—among them Amazon, Walmart, Cisco, Uber, and Meta—have begun to limit or actively discourage use, as the cost of large-scale deployment tests corporate finances. Uber exhausted its full 2026 AI budget by April and imposed a US$1,500 monthly token cap per employee; Walmart set similar ceilings on internal agents. The trigger: a pivot by providers such as OpenAI and Anthropic from fixed subscriptions to token-based pricing, which exposes firms directly to the cost of every prompt and automated workflow. Deloitte’s global lead for generative AI notes that boards and CFOs are now scrutinising compute costs after an earlier assumption that AI was cheap or free.

The pullback coincides with a phenomenon described as AI sprawl, where workers juggle multiple tools, replicate work, and spend time ‘botsitting’ to make outputs usable. A survey of 6,000 digital workers in the US, UK, and Australia by Glean’s Work AI Institute found 77% of AI users engage with several programmes weekly, and only 13% say time savings have significantly improved company performance. The rush to signal innovation through token-intensive activity—dubbed tokenmaxxing—has produced duplicate reports and underused internal agents, burning through budgets without clear returns. In parallel, the transition from chatbots to autonomous agents that can handle complex tasks is accelerating token consumption. Goldman Sachs forecasts a 24-fold rise in token use by 2030, driven by agents, and warns of worsening chip shortages.

The pressure is also reshaping labour. Technology workers report spending evenings and weekends learning new AI tools out of fear of being left behind, according to interviews conducted in Silicon Valley and Dublin. Some 85% of US desk workers surveyed by Ernst & Young were pursuing AI education outside work hours. Meanwhile, the energy demands of AI infrastructure are colliding with grid constraints. Viewed from Hong Kong—which ranks third globally as an AI financial hub yet has run an electricity deficit for decades—the mismatch between digital ambition and power supply is acute, prompting calls to use the Greater Bay Area as an energy hinterland. European policymakers face similar bottlenecks; the EU aims to triple computing capacity but must contend with saturated electricity networks and local protests.

The next test of the sector’s momentum will arrive later this year with the planned initial public offerings of leading AI labs, including Anthropic and OpenAI, whose lofty valuations will need to weather the cost-consciousness sweeping their corporate clients. In the meantime, the focus is shifting from building at all costs to demonstrating measurable productivity gains, as companies seek to tame the sprawl while keeping innovation alive.