企业AI支出从创新实验向核心运营迁移：75%年增背后，多模型、提示工程与直接采购成新标配
Based on a June 2025 survey of 100 enterprise CIOs across 15 industries, generative‑AI budgets are growing faster than expected (average ~75% YoY), with spend shifting from experimental innovation funds to core IT and business‑unit budgets. Multi‑model deployments are now standard (37% using 5+ models), driven by task‑specific model differentiation rather than just vendor‑lock‑in concerns. Model costs have dropped roughly ten‑fold per year, improving price‑performance for closed‑source small/medium models; as a result, fine‑tuning is declining in favor of prompt engineering and long‑context solutions. Reasoning models are in early testing but already showing strong adoption for OpenAI’s o3 (23% in production vs 3% for DeepSeek). Enterprises are moving toward direct procurement from model providers or via Databricks, seeking early access and lower switching costs—though agentic workflows increase lock‑in because extensive prompt tuning is model‑specific. External benchmarks (e.g., LM Arena) are now a key filter in procurement, complementing internal evaluation. The ecosystem of third‑party AI applications has matured, driving a buy‑vs‑build shift; over 90% of respondents are testing third‑party customer‑support apps, while regulated sectors like healthcare still favor in‑house development.
发布时间：2025年6月10日
英文原标题：How 100 Enterprise CIOs Are Building and Buying Gen AI in 2025
来源：查看 a16z 原文
- LLM budgets grew faster than expected; CIOs forecast ~75% YoY increase, with one noting they now spend weekly what they spent annually in 2023.
- Innovation spending dropped from 25% of LLM budgets last year to 7%, reflecting AI’s transition from experimental to essential operational spend.
- Enterprises increasingly fund AI via centralized IT and business‑unit budgets rather than project‑specific innovation funds.
- 37% of respondents now run 5 or more models in production, up from 29% a year ago, primarily due to task‑specific model strengths rather than vendor‑lock‑in avoidance.
- Anthropic’s Claude excels at fine‑grained code completion; Gemini leads in higher‑level system design; for text tasks, Anthropic outperforms in writing fluency and brainstorming, while OpenAI is stronger on complex Q&A.
- LLM预算增长超预期，平均同比增长约75%。
- 创新实验经费占比从去年25%降至7%，AI支出正从项目型实验转向常规运营投入。
- 资金来源从项目专属创新基金转向集中式IT和业务部门预算，显示AI已成核心基础设施。
- 多模型部署已成常态，37%的企业使用5个以上模型，较去年29%提升，主要基于任务专属模型优势。
- 因模型成本大幅下降，企业更倾向使用提示工程和长上下文方案，而非精细调优。
判断：在未来 6‑12 个月内，企业将把剩余的 LLM 创新专项经费全部并入核心 IT/业务部门运营预算，使得创新基金在 LLM 支出中的占比从当前的 7% 降至 2% 以下，而整体 LLM 预算通过集中采购渠道（如云服务商或模型供应商直签）完成的比例将超过 80%。
时间跨度：未来 6-12 个月
为什么是现在：模型成本每年约十倍的下降使一次性实验不再需要专项创新基金；多模型标准部署和长上下文方案强化了基于性能和外部基准的采购流程；企业内部预算治理要求所有 AI 投入统一计入运营支出；OpenAI o3 等推理模型已进入生产阶段，推动企业直接向模型供应商采购。
重点信号：在公开财报或监管文件中，企业将 LLM 支出列为“AI 平台服务”或“核心基础设施”而不是“创新项目”的比例超过 70%。、云服务商（如 AWS、Azure、Google Cloud）季度报告显示，针对 LLM 的直接采购合同数量较上年增长 50% 以上。、调研机构（如 Gartner）发布的《企业 AI 预算报告》中，LLM 创新预算占总 AI 预算的比例跌破 5%。、API 调用统计显示，精细调优 API（如 OpenAI fine‑tuning）调用频次同比下降 30% 以上，而推理模型（如 o3）调用频次增长 200% 以上。
置信度：高