核心要点：每百万token的有效价格自2026年3月峰值以来下降41%，从1.15美元跌至0.68美元（Ramp报告）。企业正从高价前沿模型“降级”迁移到更高效的标准化模型。8月，Anthropic占美国付费使用AI企业份额的43.8%，OpenAI占39.8%（Binance对Ramp数据的分析）。
企业人均AI支出随Token成本缩水而下降
据Ramp AI Index数据，2026年8月，头部企业公司的AI支出出现显著收缩：最高支出1%企业的人均月度AI中位支出下降9.7%，从7976美元降至7205美元。这标志着前沿模型供应商与超大规模云厂商所处经济格局的一次重要转变。
尽管整体商业AI采用仍在上升——8月仅增长0.4个百分点——但重度使用者的人均支出下降，表明效率提升与市场竞争正在压低成本。对基础设施投资者而言，这一趋势尤其重要：它凸显了AI硬件的巨额资本开支，与实际从企业token消耗中捕获的营收之间可能存在的落差。
市场动态与技术规格
支出下降由两个主要的技术与经济因素驱动：token价格的激进通缩，以及企业用户在模型选型上的策略转向。
Token价格通缩：每百万token的有效价格自2026年3月峰值以来已下降41%，从1.15美元降至0.68美元（Ramp报告）。
模型档位迁移：企业正越来越多地从高成本前沿模型“降级”到更高效的标准化模型。据Ramp AI Index，Claude Opus、Fable与Sol等前沿模型的token消耗份额，从8月初的53%降至9月初的45%。
供应商市场份额：据Binance对Ramp数据的分析，8月Anthropic在美国付费使用AI服务的企业中占据43.8%的份额，OpenAI为39.8%。
对AI基础设施的影响
向更便宜、更专用模型的迁移，对行业构成一体两面的现实。对企业而言，智能成本的下降使更大范围的部署和更好的单位经济性成为可能。但对模型开发者而言，使用量的当前增长未必足以抵消每token价格的快速下滑。
此外，开放权重模型在企业端的采用仍属小众：Ramp平台上8月仅6.4%的AI支出企业使用了开放权重模型，说明闭源专有模型仍主导专业生产工作流（据TechCrunch）。
随着企业把成本效率与可衡量结果放在首位，行业关注点正从原始模型能力，转向优化RAG（检索增强生成）流水线，以及调用最具性价比模型的智能体工作流。
本文来源：https://thenextgentechinsider.com/pulse/enterprise-ai-spending-per-employee-drops-as-token-costs-deflate
若有版权问题，请来信告知：13951431913#139.com 或在右方提交评论。
Key Takeaway: The effective price per million tokens has dropped by 41% since peaking in March 2026, falling from $1.15 to $0.68 (Ramp report). Enterprises are "downgrading" from high-cost frontier models to more efficient, standardized options. Anthropic held 43.8% of the U.S. paid enterprise AI market in August, with OpenAI at 39.8% (Binance analysis of Ramp data).
Corporate AI Spending Per User Shrinks as Token Costs Plunge
According to the Ramp AI Index, the top 1% of enterprises by spending saw a significant contraction in August 2026: the median average AI spend per user per month dropped by 9.7%, from $7,976 to $7,205. This marks a critical shift in the economic landscape for frontier model providers and hyperscale cloud vendors.
While overall commercial AI adoption continues to rise — ticking up just 0.4 percentage points in August — the drop in per-user spending among heavy adopters indicates that efficiency gains and market competition are driving down costs. For infrastructure investors, this trend is particularly relevant: it highlights a potential disconnect between the massive capital expenditures on AI hardware and the actual revenue capture from token consumption by businesses.
Market Dynamics and Technical Specifications
The spending decline is driven by two primary factors: aggressive token price deflation and a strategic shift in model selection by enterprise users.
Token Price Deflation: The effective price per million tokens has fallen sharply, dropping 41% from a peak of $1.15 in March 2026 to $0.68 (Ramp report).
Model Downgrading: Enterprises are increasingly "downgrading" from expensive frontier models to more cost-effective, standardized options. The Ramp AI Index shows that token consumption for frontier models like Claude Opus, Fable, and Sol dropped from 53% of total consumption at the start of August to 45% by early September.
Supplier Market Share: In August, Anthropic accounted for 43.8% of the U.S. enterprise market for paid AI services, with OpenAI at 39.8%, according to Binance's analysis of Ramp data.
Implications for AI Infrastructure
The migration toward cheaper, more specialized models presents a dual reality for the industry. For enterprises, the lower cost of intelligence enables broader deployment and improved unit economics. However, for model developers, the current growth in usage may not be sufficient to offset the rapid decline in per-token pricing.
Additionally, the adoption of open-weight models remains niche in enterprise settings: only 6.4% of companies spending on AI in August used open-weight models via the Ramp platform, indicating that proprietary, closed-source models still dominate professional production workflows (source: TechCrunch).
As businesses prioritize cost efficiency and measurable outcomes, the focus is shifting away from raw model capability toward optimized RAG (Retrieval-Augmented Generation) pipelines and agent workflows that call the most cost-effective models for specific tasks.