当前人工智能基础模型行业正面临严重的 Token 供应短缺，但这一供需失衡局面预计将在未来几年内随着市场出清而达到新的平衡 [1]。在供给侧，数万亿美元的数据中心资本支出正在持续涌入，同时推理效率也在快速提升，不同新模型之间的 Token 使用效率存在巨大差异 [1]。需求侧的产能紧张目前主要由软件开发领域单一用例的产品 - 市场契合度所驱动，该领域的市场规模相对较小 [1]。
就利润率现状而言，AI 推理业务的毛利率约为 40% 至 50%，但这一数据尚未包含高昂的模型训练成本，且相关资产的寿命存在不确定性 [1]。文章分析认为，除非出现网络效应、监管壁垒或技术断层等变量改变趋势，否则前沿 AI 模型最终将演变为低利润的商品化基础设施，其价值将被上层应用捕获 [1]。这种未来格局类似于移动通信数据或半导体制造行业，其中基础设施的价值主要被上层的软件或服务所获取 [1]。此外，潜在的政策变量如中国对开源模型的监管以及美国的出口管制也可能改变这一市场走向 [1]。
An article published on Hacker News analyzes the current pricing mechanisms and future market landscape for AI foundation models against a backdrop of constrained token supply [1]. The author argues that while severe supply-demand imbalances exist today, this situation is temporary; it will resolve into a new equilibrium over the coming years as massive capital expenditures flood data centers and inference efficiency improves significantly [1]. Currently, capacity constraints are primarily driven by product-market fit in software development—a relatively small use case—rather than broader demand saturation .
Regarding profitability, AI inference operations currently boast gross margins of approximately 40% to 50%, though these figures do not yet account for the high costs associated with model training or uncertainties regarding asset lifespans . The core thesis posits that unless variables such as network effects, regulatory barriers (including China's regulation of open-source models and US export controls), or technological discontinuities intervene, leading AI models will eventually evolve into low-margin commoditized infrastructure similar to mobile data networks or semiconductor manufacturing . In this future scenario, the primary value capture is expected to shift from model providers to upper-layer applications and services .