卖Token还是卖结果：AI商业模式的几个悖论

人工智能产业的快速扩张掩盖了其商业生态中的深层矛盾。2026年3月，全球单周Token调用量达到20.4万亿，中国日均调用量突破140万亿，较2024年初增长超过千倍1。然而这种惊人的使用量增长并未转化为均衡的利润分布。芯片层在生成式AI生态年化收入约4000亿美元的规模中独占约七成收入和近八成毛利，而应用层收入仅约600亿美元、毛利率多数徘徊在0%-30%的低位1。这种失衡反映了AI行业面临的第一个悖论：成本下降与账单反增的错位。

GPT-4的Token价格三年内下跌超过95%，从2023年3月的输入30美元/百万、输出60美元/百万大幅下降1。尽管单位成本急剧下落，企业和用户的总支出却在增加，形成了"越便宜越贵"的怪象。

敢于承担结果责任的厂商获利能力凸显。OpenAI的年化收入从2023年的20亿美元增至2026年2月底的超过250亿美元1；Anthropic的年化收入在2024年12月约10亿美元的基础上，到2026年5月飙升至470亿美元以上，其推理基础设施毛利率也从38%提升至70%以上1。与之形成对比的是，企业对开源模型的支出占比从前年的19%降至今年的11%，转而投向闭源方案，表明市场对承诺结果的提供商更有信心1。

新型计费模式的出现则反映了行业的转变。Intercom推出的智能客服产品Fin采取按实际解决问题计费的模式，每单收费0.99美元，未解决问题则不收取费用1。这种结果导向的定价方式体现了技术方对自身能力的承诺。OpenRouter平台上，Anthropic虽只占Token份额的12%，却获得了金额份额的46%，进一步验证了高价值厂商的溢价能力1。

The artificial intelligence sector is experiencing a fundamental disconnect between unit economics and overall profitability, according to analysis of market trends through mid-2026. As token prices have collapsed by over 95% since March 2023—dropping from $30 per million input tokens and $60 per million output tokens—the industry paradoxically faces rising total customer expenditures1. Global weekly token invocations reached 20.4 trillion in March 2026, while China's daily token consumption surged to 140 trillion, representing growth exceeding 1,000% from early 20241.

The generative AI ecosystem illustrates a striking concentration of value capture. With approximately $400 billion in annualized revenue across the sector, chip manufacturers capture roughly 70% of total revenue and nearly 80% of gross margins, leaving application-layer companies to divide approximately $60 billion in revenue with profit margins typically ranging from 0% to 30%1. This structural imbalance reflects what analysts term a "hierarchy paradox," where despite rapid value creation at application levels, capital flows persistently toward infrastructure layers. Anthropic exemplifies this dynamic, growing annualized recurring revenue from approximately $1 billion in December 2024 to over $47 billion by May 2026, with reasoning infrastructure gross margins expanding from 38% to above 70% during the same period1. OpenAI's trajectory similarly demonstrates infrastructure's gravitational pull on profits: the company achieved $2 billion in annualized revenue in 2023, $6 billion in 2024, exceeding $20 billion in 2025, and surpassing $25 billion by February 2026, though free cash flow profitability remains projected for approximately 20301.

A fourth emerging paradox involves the divergence between open-source adoption and monetization. Enterprise spending on open-source models declined from 19% of AI spending budgets to 11%, while closed-source captured 89% of expenditures1. On the OpenRouter platform, Anthropic commands 46% of transaction value despite representing only 12% of token volume, underscoring how proprietary solutions extract disproportionate revenue1. New result-based pricing models are also emerging, exemplified by Intercom's Fin product, which charges $0.99 per successfully resolved customer issue, with no fee charged when problems remain unresolved1.