Core Conclusion: AI has evolved from “edge projects” to “core business,” requiring enterprises to build sustainable competitiveness across technical, cost, compliance, and talent dimensions.

1. Differentiated Paths for AI-Native and AI-Enabled Enterprises

  • AI-Native Companies: Leverage technical advantages for rapid iteration but face scaling cost and compliance challenges.
  • AI-Enabled Companies: Embed AI into existing products (e.g., customer service, recommendation systems), balancing user experience with cost control.
  • Model Selection: A multi-model strategy has become mainstream (e.g., open-source models + customized training), balancing performance and cost.

2. Cost Management: The Core Challenge of Scaling

  • API Usage Fees: 70% of companies list this as their largest cost, with user fluctuations (e.g., promotional campaigns) causing exponential fee increases.
  • Inference and Training Costs: During scaling, high-growth companies face inference costs of $2.3 million monthly, while average companies see $1.1 million. Model training costs rise from $163,000 pre-launch to $1.5 million during scaling.
  • Optimization Strategies:
    • Model Compression: Reducing large model parameters from 70 billion to 700 million cuts inference costs by 70%.
    • Open-Source Alternatives: 41% of companies use Llama 3 or similar open-source models; 37% optimize inference efficiency (e.g., quantization, distillation).

3. Compliance and Pricing: Escalating Complexity

  • Compliance Needs: AI tools must meet data security and privacy regulations, requiring early compliance frameworks.
  • Pricing Strategy: Balance user experience, cost, and regulatory requirements to avoid profit erosion from cost overruns.

4. Talent Bottleneck: Scarcity of Composite Skills

  • Demand: AI engineers with both technical and business expertise account for 20%-30%, with high-growth companies reaching 37%.
  • Mitigation Measures:
    • Internal Training: Enhance existing employees through training and university partnerships.
    • Talent Competition: Top AI engineers are frequently “poached,” with salary increases exceeding 50%.

5. Internal AI Tools: Key to Efficiency

  • Use Cases: Code assistance (65% of companies report maximum productivity gains), content generation, and knowledge retrieval.
  • Adoption Rates: 70% of employees engage with AI tools, but only 50% continue using them; 44% of large companies have persistent users at 44%.
  • Challenges: 46% of companies struggle to find use cases, 42% find it hard to prove ROI.
  • Measurement: 75% track productivity gains, 51% track cost savings, and only 20% focus on revenue growth.

6. Technology Stack and Tools: Ecosystem Differentiation and Collaboration

  • Frameworks: PyTorch/TensorFlow dominate (over 50%), with hosted platforms (e.g., AWS SageMaker) also popular.
  • Development Tools: LangChain, Hugging Face toolkits simplify prompt chain management; private API encapsulation of public models.
  • Security and Monitoring: Guardrails prevent harmful content generation, LangSmith/Weights & Biases track model drift.
  • Inference Optimization: NVIDIA TensorRT/Triton account for 60%+, ONNX Runtime offers cross-hardware compatibility.

7. Key Lessons and Strategic Recommendations

  1. Clarify the Problem: AI must address specific business pain points, not “AI for AI.”
  2. Control Costs: Avoid profit erosion from API fees during scaling.
  3. Compliance First: Preemptively build data security and regulatory frameworks.
  4. Internal Implementation: Executive support and employee training are critical for AI tool adoption.
  5. Talent Strategy: Build composite teams, balancing technical and business capabilities.

Summary

In 2025, AI has entered the “deep water zone,” with competition focusing on stability, compliance, and economic viability. Enterprises must build systematic capabilities across technology, cost, compliance, and talent to remain competitive in the AI浪潮.

Translation

2025年AI开发趋势与挑战：从技术到战略的全面洞察

核心结论：AI已从“边缘项目”演变为“核心业务”，企业需在技术、成本、合规、人才等多维度构建可持续竞争力。

1. AI原生与赋能企业的差异化路径

  • AI原生公司：凭借技术优势快速迭代，但需应对规模化成本和合规挑战。
  • AI赋能公司：通过现有产品嵌入AI功能（如客服、推荐系统），需平衡用户体验与成本控制。
  • 模型选择：多模型策略成为主流（如开源模型+定制化训练），以平衡性能与成本。

2. 成本管理：规模化的核心挑战

  • API使用费：70%企业将其列为最大成本，用户量波动（如促销活动）导致费用激增。
  • 推理与训练成本：规模化阶段，高增长公司推理成本达每月230万美元，普通公司为110万美元；模型训练成本从发布前的16.3万美元增至规模化阶段的150万美元。
  • 优化手段：
    • 模型压缩：将大模型参数从700亿降至70亿，推理成本降70%。
    • 开源替代：41%企业使用Llama 3等开源模型，37%优化推理效率（如量化、蒸馏）。

3. 合规与定价：复杂性升级

  • 合规需求：AI工具需符合数据安全、隐私保护等监管要求，企业需提前搭建合规体系。
  • 定价策略：需平衡用户体验、成本与监管要求，避免因成本失控导致利润缩水。

4. 人才瓶颈：复合型人才稀缺

  • 需求：既懂技术又懂业务的AI工程师占比20%-30%，高增长公司甚至达37%。
  • 应对措施：
    • 内部培养：通过培训、高校合作提升现有员工能力。
    • 人才争夺：顶级AI工程师常被“挖墙脚”，薪资涨幅超50%。

5. 内部AI工具：效率提升的关键

  • 应用场景：编码辅助（65%企业认为提升生产力最大）、内容生成、知识检索。
  • 使用率：70%员工接触AI工具，但仅50%持续使用；大公司员工持续使用率仅为44%。
  • 挑战：46%企业难以找到使用场景，42%难证明ROI。
  • 衡量方式：75%跟踪生产力提升，51%跟踪成本节省，仅20%关注收入增长。

6. 技术栈与工具：生态分化与协作

  • 框架：PyTorch/TensorFlow主导（占50%以上），托管平台（如AWS SageMaker）也受青睐。
  • 开发工具：LangChain、Hugging Face工具集简化提示词链管理；私有API封装公开模型。
  • 安全与监控：Guardrails防有害内容生成，LangSmith/Weights & Biases跟踪模型漂移。
  • 推理优化：英伟达TensorRT/Triton占60%以上，ONNX Runtime提供跨硬件兼容性。

7. 关键教训与战略建议

  1. 明确问题：AI需解决具体业务痛点，而非“为AI而AI”。
  2. 控制成本：避免规模化后利润被API费用吞噬。
  3. 合规先行：提前布局数据安全与监管框架。
  4. 内部落地：高管支持与员工培训是推动AI工具落地的关键。
  5. 人才战略：构建复合型团队，平衡技术与业务能力。

总结

2025年AI已进入“深水区”，竞争焦点从“功能创新”转向“稳定性、合规性与经济性”。企业需在技术、成本、合规、人才等维度构建系统性能力，方能在AI浪潮中持续领先。

Reference:

https://www.iconiqcapital.com/growth/reports/2025-state-of-ai, https://cdn.prod.website-files.com/65d0d38fc4ec8ce8a8921654/685ac42fd2ed80e09b44e889_ICONIQ%20Analytics_Insights_The_AI_Builders_Playbook_2025.pdf

<
Previous Post
Context Engineering - a new made-up AGI term
>
Next Post
Dylan Patel interview with Matthew Berman