Every quarter, the giants of cloud computing report another record haul of AI revenue, and the narrative writes itself: businesses everywhere are pouring money into artificial intelligence, and the boom has no ceiling. A new analyst note suggests the picture on the ground is more complicated. Beneath the headline numbers, ordinary companies, the ones that buy AI rather than build it, may be quietly easing off.
RBC Capital Markets, in a note led by analyst Rishi Jaluria, made a careful distinction that is easy to miss. The eye-watering demand cited by Microsoft, Amazon, Meta, Oracle and Google largely reflects spending on model training, deployment and AI-native startups, not a broad surge across traditional enterprises. In other words, a lot of the money sloshing around is the AI industry paying itself. The question is whether the rest of the economy is following.
The data point RBC leaned on is small but symbolically large. According to Ramp's Fall 2025 Business Spending Report, the share of US businesses paying for AI services slipped from 44.5 percent in August to 43.8 percent in September. That is a fraction of a percentage point, and one month is not a trend. But Ramp's figures suggest it is the first measurable pullback since enterprise AI adoption began climbing in 2023. After two years of one-way traffic, even a flat tyre is news.
RBC's explanation centres on what it calls the productivity paradox: many companies simply have not seen the gains they were promised. They bought the tools, ran the pilots, and are still waiting for the line on the chart to bend. That frustration shows up elsewhere too. A survey of workers in Singapore found that while nearly three in four employees are comfortable working alongside generative AI and most use it daily, sizeable minorities still worry about job security, accuracy of outputs, and their own lack of fluency. Enthusiasm and unease are living side by side.
Part of the problem is that, even now, nobody can say cleanly what this technology can and cannot do. As one industry observer put it, the claims made about AI four years ago are largely the same claims being made today, just pushed into new corners: better accuracy, faster reporting, more time for "other tasks." Vendors keep shipping assistants, chatbots and copilots, but adoption inside the buying companies often lags far behind the launch announcements. Buying a licence is not the same as changing how work gets done.
None of this means the AI boom is over. The infrastructure spending is real, and a single month of softer data could reverse just as quickly. But it is a useful corrective to the idea that enterprise demand is a one-way escalator. The companies writing the cheques are starting to ask a harder question than "should we adopt AI?" They are asking "what did the last round of adoption actually buy us?" For an industry that has run on momentum, that shift in tone may matter more than any one spending figure.