The next computing supercycle is being defined by agentic AI, which the analysis argues is more economically impactful than generative AI and far more compute-intensive. The spending figures are striking: data center AI capital expenditure is expected to reach about $1 trillion this year and $3–4 trillion by 2030, compared with projected world GDP of $123 trillion. The investment case rests on returns: Morgan Stanley sees paths to 25–50% ROIC for major hyperscalers, Google says TPUs recover cash outlay in one year and Axion CPUs in two, and Amazon says parallel construction of chips, racks, and data centers lets Trainium systems reach revenue-generating operation within months.
Demand is also shifting from humans to agents. Cisco's Jeetu Patel said agent token consumption now exceeds human token use of generative AI, while fewer than 2% of humans use agents as power users; the analysis extrapolates that agent token demand could grow 50-fold. Falling token costs from advanced process nodes and packaging are expected to increase demand rather than reduce it, which is why TSMC, builder of more than 90% of AI compute devices, is expanding fabs and advanced packaging capacity.
Because this article draws on a single source, there are no differing estimates to reconcile. The projection is not independently verified, but the source expects pauses and consolidation even as it describes a 10-plus-year supercycle.