Cursor团队单月消耗1.5万亿token,年化成本约1.32亿美元。
According to InfoQ China, Cursor’s engineering team consumes 1.5 trillion tokens per month — translating to approximately $132 million annually at prevailing API pricing. Over 73% of this usage occurs in CI/CD pipelines (test generation, build failure analysis) and PR review (diff summarization, vulnerability scanning), not just interactive coding. Crucially, this figure excludes self-hosted model inference costs — it reflects pure cloud API spend. This reveals a critical inflection: for AI-native dev tools, token consumption shifts from ‘experimental budget line’ to ‘non-negotiable infrastructure cost’ once embedded in daily workflows. Early-stage AI application teams must treat token economics like database query optimization — measuring cost-per-action, caching high-value outputs, and aggressively pruning low-ROI calls (e.g., redundant re-embedding of unchanged files). Ignoring this leads to unsustainable unit economics long before product-market fit.