Nvidia provides its engineers with an AI token budget of US$250,000 per person per year, while Uber has set a monthly cap of US$1,500, highlighting the significant disparities in AI spending among top tech companies. Overseas investment institutions point out that companies follow two sets of economic logic when purchasing AI services: first, those viewing AI as an expansionary market for growth capital are willing to pay any cost for the best models to generate more revenue, ultimately moving towards outcome-based pricing; second, those seeing AI as an efficiency market for cost control stop additional investment once intelligence meets standards, tending towards infrastructure logic with pricing based on tokens or API calls. The same task may fall into different markets depending on a company’s goals—for example, software engineering is an efficiency task for Uber but an expansion task for startups. The same company may also adopt both logics simultaneously. Currently, most companies are efficiency buyers, scaling back excessive AI spending, which has given rise to four trends: companies allocate AI as a portfolio; open-source models gain development opportunities; closed-source frontier labs need to escape the efficiency market and shift to areas like finance and drug development; the efficiency market prices based on infrastructure logic, with competition descending to inference costs and other links. In expansionary markets, the value of intelligence has no upper limit, with competition taking two forms: zero-sum and positive-sum. In zero-sum games, intelligence merely redistributes a limited pie, with model providers effectively collecting a competition tax—this spending is intense but fragile. In positive-sum games, intelligence expands the pie, resembling investment—sustainable but encountering non-intelligence bottlenecks. For frontier labs, customer composition matters more than total volume. Whether AI investment is worthwhile hinges on whether intelligence is the bottleneck for output: if so, frontier AI is worth the cost; if not, choosing the cheapest model that meets standards is more rational. The goal of most AI markets is cost reduction, while the true frontier lies in areas where the value of higher intelligence continues to climb.