Demand for AI infrastructure exceeds that of the internet era, with broad programming driving ARR growth

Huo Xing Finance reports that an analysis of the AI semiconductor and infrastructure cycle highlights that AI infrastructure demand will last longer than during the internet era, as the product of user penetration and tokens per capita—converted into total tokens—significantly raises the ceiling. User penetration has already exceeded 50%, with further growth primarily driven by tokens per capita in still-early-stage markets. The median monthly AI expenditure per U.S. enterprise employee is approximately $12, potentially rising over the long term to around 10% of white-collar salaries—roughly $1,000 per month—leaving nearly two orders of magnitude of growth potential. Unlike the flat subscription model and minimal marginal hardware consumption of the internet era, high inference costs result in significantly higher marginal costs per access, demanding greater infrastructure intensity. After developers write code, the next major driver of ARR growth will come from generalized programming: non-programmers using programming infrastructure to accomplish cross-industry, non-programming tasks, making programming the default execution kernel for agents. Tasks related to generalized programming account for approximately 60–70% of ARR. As of June 2026, Codex accounted for 64% of the combined token output from Codex and ChatGPT among OpenAI’s enterprise customers; since February, Codex has seen far higher month-over-month active user growth in verticals such as legal, sales hiring, and marketing compared to engineering. Anthropic’s revenue includes approximately 40% from narrow development/software, over 20% from financial services and insurance, and significant contributions from legal, life sciences, retail, and other sectors. The demand-side logic for token growth remains intact, with strong alignment between funders and beneficiaries.