← 全部快讯
成本
kaifaai.app AI 快讯 · 2026-11-15

联想词元工厂年产能2000万亿Token

联想发布‘词元工厂’,年产2000万亿Token,成本降低70%。

联想正式推出‘词元工厂’,宣称其AI算力基础设施年Token产出达2000万亿,单位Token成本较传统方案下降70%。该数据源自官方发布,未说明具体硬件配置、能耗指标或服务对象(如是否含云租用或专属部署),但明确指向AI推理规模化带来的边际成本断崖式下降。对早期AI团队而言,这构成双重启示:一方面,底层Token供给正快速商品化,自建千卡集群的经济合理性持续削弱,按需采购高性价比推理资源将成为更优解;另一方面,成本骤降将加速‘高Token消耗型应用’的商业化尝试,例如长上下文实时分析、多模态流式生成、高频Agent决策等此前因成本过高而被搁置的场景。团队应重新评估MVP阶段的算力预算结构,将更多资源倾斜至产品交互设计与用户价值验证,而非底层性能优化。

Lenovo's 'token factory' produces 2 quadrillion tokens/year, 70% cheaper

Lenovo announced its 'token factory' infrastructure, stating it delivers an annual capacity of 2 quadrillion tokens at a 70% lower cost per token compared to conventional approaches. While hardware specs, energy efficiency, or deployment models (e.g., cloud vs. dedicated) weren't detailed, the claim underscores a tangible inflection point in AI inference economics — where massive scale drives steep marginal cost declines. For early AI teams, this means two strategic implications: first, owning large-scale inference infrastructure is becoming increasingly uneconomical versus consuming high-throughput, low-cost token services on demand; second, dramatically reduced token costs unlock previously prohibitive use cases — such as real-time long-context analysis, streaming multimodal generation, or high-frequency agent decision loops. Teams should rebalance MVP budgets away from premature infrastructure optimization and toward rapid iteration on user-facing value, interaction design, and workflow integration.

来源:量子位 — 绿电进Token出!联想重新定义词元工厂
© 2026 kaifaai.app — AI 产品技术开发 · 28 个方向明码标价