In recent public speaking at a Xiaohongshu.com audio room, full of young Chinese AI practitioners from around the world, I stated these 3-phases of AI industry evolution. They’re evolving in layers that overlap rather than replace one another. Seen from the right distance, however, a clear arc emerges — and it runs through three phases: foundation models, agents, and privatized enterprise AI. Each phase builds on the last, and the timing between them shifts faster than most forecasts admit.
📜 The Three Phases at a Glance
- P1 (2023-2025): LLMs dominate. The explosive breakout of large language models and the frantic expansion of AI infrastructure. By 2025, the industry was already crossing the inflection point from raw model capability toward agentic behavior — the end of LLM dominance arrived earlier than the calendar suggests.
- P2 (2025–now): Agents take over. 2025 is widely considered the year of the AI agent. The industry’s focus has shifted from can it generate an answer to can it get a task done.
- P3 (2028 onward): Private models + private datasets. Driven by data sovereignty, security, and compliance, enterprises are moving toward private deployment at accelerating speed — this phase may arrive well before 2028.
🔄 A Finer-Grained Industry Evolution Map
Zoom in, and the three phases resolve into a more granular sequence:
- 2023-2024: Model race & capability diffusion. An arms race in model capabilities, the rise of open-source models democratizing the technology, and enterprises beginning to explore applied techniques like RAG.
- 2025-2026: Agent explosion & task automation. The core shift is from conversational AI to agentic systems that can autonomously plan, call tools, and execute tasks. This marks AI’s leap from copilot to digital employee. Gartner placed it at the peak of the hype cycle in 2025.
- 2026-2028: Enterprise deployment & value validation. AI moves into the deep end — the focus shifts to proving ROI. Per Gartner’s forecast, by 2028 at least 15% of daily work decisions will be made autonomously by agentic AI. Meanwhile, new technologies like model routers will become the key control point for managing multi-model portfolios.
- 2028 and beyond: Private models & AI systemization. Enterprises will adopt the private model + private data pattern at scale to build core competitive moats. Forrester predicts 70% of AI revenue will be driven by private models within five years. At the same time, AI will evolve into AI systems deeply embedded in business processes.
🚀 Predictions for the Future
- Privatization will accelerate — mainstream by 2028, perhaps earlier. Enterprise anxiety over data sovereignty, combined with open-source models rapidly catching up in performance, will push private deployment to mainstream status before 2028. The end state is likely a hybrid AI architecture — frontier public models for general tasks, private models for core business.
- Models themselves will quickly commoditize; value shifts to data and systems. When frontier models converge in capability, the real moat becomes unique private data and AI systems deeply embedded in workflows.
- Agents will move from solo workers to collaborative networks. Future agents won’t work in isolation — they’ll collaborate like teams. For example, a marketing agent might call a data-analysis agent, a content-generation agent, and others to accomplish complex goals together.
- Model routing will become standard in enterprise AI architecture. Enterprises will no longer depend on a single model — instead, a model router dynamically selects the best model per task type, cost, and latency.
💎 Summary
Underneath the noise, the evolution logic is consistent: from general models to specialized agents, then to privatized data. The phases aren’t sharply separated — they layer, overlap, and build on one another.
If 2023-2024 was the era of making wheels (building models) and 2025-2026 the era of mounting tires (equipping agents), then 2027 onward is the era of building the whole vehicle (building systems) — and private data is that vehicle’s engine.
中文版 / Chinese Version
最近在「小红书」（Xiaohongshu.com）音频直播间，面对来自世界各地的年轻中国AI从业者，我阐述了AI产业演进的三阶段框架。这些阶段层层递进、相互重叠，而非彼此取代。但只要拉开合适的距离，一条清晰的脉络便会浮现——它贯穿三个阶段：基础模型、智能体（Agent）、以及私有化的企业AI。每个阶段都建立在上一阶段之上，而阶段之间的节奏变化，比多数预测所承认的还要快。
📜 三阶段一览
- P1 (2023-2025): LLMs 主导。大语言模型（LLM）的爆发元年，AI基础设施的疯狂扩张期。到2025年，产业已跨越从原始模型能力向Agent化行为转变的关键拐点——LLM主导期的结束，比日历时间显示得更早。
- P2 (2025至今): Agent 接棒。2025年被广泛认为是「AI Agent元年」。产业焦点已从「能否生成答案」转向「能否完成任务」。
- P3 (2028及以后): 私有模型 + 私有数据集。在数据主权、安全与合规的驱动下，企业正加速转向私有化部署——这一阶段可能远早于2028年到来。
🔄 一个更精细的产业演进图谱
拉近镜头，三阶段会分解为更细粒度的演进序列：
- 2023-2024: 模型竞赛与能力扩散期。特点是模型能力的军备竞赛、开源模型崛起推动技术民主化，以及企业开始探索RAG等落地技术。
- 2025-2026: Agent爆发与任务自动化期。核心是从对话式AI向能自主规划、调用工具和执行任务的Agent系统转变。这标志着AI从「副驾驶 (Copilot)」向「数字员工」的跃迁。Gartner在2025年已将其置于「期望膨胀期」的顶点。
- 2026-2028: 企业落地与价值验证期。AI进入深水区，重心转向验证ROI。据Gartner预测，到2028年，至少15%的日常工作决策将由Agentic AI自主完成。同时，「模型路由器」 等新技术将成为管理多模型组合的关键控制点。
- 2028及以后: 私有模型与AI系统化时代。届时，企业将大规模采用「私有模型+私有数据」模式来构建核心竞争壁垒。Forrester预测，未来五年内，AI收入的70%将由私有模型驱动。同时，AI将演变为深度嵌入业务流程的「AI系统」。
🚀 关于未来的预测
- 「私有化」进程将加速，2028年或已成主流。企业对数据主权的焦虑，叠加开源模型性能的快速追赶，将推动私有化部署在2028年前成为主流选择。最终形态很可能是「混合AI架构」——用前沿公共模型进行通用任务，用私有模型处理核心业务。
- 「模型」本身将迅速商品化，价值向「数据」与「系统」转移。当顶尖模型能力趋同时，真正的护城河将是独特的私有数据和深度嵌入工作流的AI系统。
- Agent将从「单兵作战」走向「协同网络」。未来的Agent不会孤立工作，而是会像团队一样协作。例如，一个营销Agent可调用数据分析Agent、内容生成Agent等，共同完成复杂目标。
- 「模型路由」将成为企业AI架构的标配。企业将不再依赖单一模型，而是通过「模型路由器」根据任务类型、成本、延迟等，动态选择最合适的模型。
💎 总结
在喧嚣之下，演进逻辑始终一致：从通用模型到专用Agent，再到私有化数据。这些阶段并非截然分开——它们层层递进、相互重叠、彼此叠加。
如果说2023-2024年是「造轮子」（建模型）的时代，2025-2026年是「装轮胎」（配Agent）的时代，那么2027年及以后，就是「造整车」（建系统）的时代，而私有数据，就是这辆车的「引擎」。