曾经在硅谷被视为高效能象征的大规模AI使用正在退潮。[1]企业对AI"tokenmaxxing"（最大化token消耗）的热情日益消退，取而代之的是更精明的成本控制策略。[1]这一转变反映出，早先被技术高管推崇的无限制AI投入模式因成本飙升而逐渐失去吸引力。[1]
曾有硅谷领袖鼓励企业肆意消耗AI计算资源。Nvidia首席执行官詹森·黄曾表示"如果你的50万美元工程师没有消耗25万美元的tokens，说明有问题"。[1]然而，这种做法的代价正在显现。微软首席执行官萨蒂亚·纳德拉警告客户为AI付了两次钱——一次是token费用，二次是向AI提供商泄露专有数据。[1]Palantir首席执行官亚历克斯·卡普表示美国企业对高额token费用"感到愤怒"，称其"产生不了价值"。[1]
企业的token成本增长令人触目惊心。咨询公司贝恩的顾问王珏指出，大企业的token成本"几乎每隔一个月翻倍"，以20000名开发者计算月成本达400万美元。[1]面对这种压力，企业正在转向更经济的替代方案。它们采用"模型路由"工具，将简单查询分配给廉价模型，复杂任务分配给高能力模型，从而优化支出。[1]同时，中国初创公司Moonshot的Kimi和Zhipu的GLM等开源模型以远低的价格接近美国顶级模型性能，为企业提供了新的选择。[1]
The rush to maximize artificial intelligence usage in corporate America is cooling as companies grapple with skyrocketing expenses. What was once promoted as a mark of operational excellence by Silicon Valley executives—lavish token consumption on AI systems—has become a cautionary tale of unchecked spending.[1]
The shift reflects growing frustration with AI costs among enterprise leaders. Nvidia CEO Jensen Huang had previously stated that "if your $500,000 engineer isn't consuming $250,000 of tokens, there's a problem," establishing a cultural standard that equated heavy token usage with productivity.[1] However, this calculus is rapidly changing. Microsoft CEO Satya Nadella has cautioned customers that they are essentially "paying twice for AI"—once through token fees and again through the disclosure of proprietary data to AI vendors.[1] Palantir CEO Alex Karp went further, asserting that American enterprises are "angry" about exorbitant token costs that "don't produce value."[1]
The financial strain is substantial and accelerating. According to Bain & Company consultant Jue Wang, token expenses at major enterprises are "nearly doubling every month," with a single organization employing 20,000 developers facing monthly costs of approximately $4 million.[1] In response, companies are adopting more strategic approaches, including model routing technology that directs simple queries to cheaper models while reserving complex tasks for premium systems.[1] Additionally, Chinese startups such as Moonshot's Kimi and Zhipu's GLM are gaining traction by offering open-source alternatives that deliver performance comparable to leading American models at significantly lower prices.[1]