{"schemaVersion":1,"generatedAt":"2026-08-20T13:21:35.716Z","horizon":{"start":"2022-08-22","end":"2026-07-13","label":"2022—今天"},"eras":[{"slug":"public-generative-ai","label":"生成式 AI 进入大众市场","period":"2022 H2","summary":"Stable Diffusion 把生成模型带入开放生态，ChatGPT 把大模型从研究能力变成大众每天可以使用的产品。","projects":[{"name":"ChatGPT","status":"active","note":"从 research preview 演化为通用 AI 产品入口。","url":"https://openai.com/index/chatgpt/"},{"name":"Stable Diffusion","status":"active","note":"开放权重持续支撑图像生成生态。","url":"https://stability.ai/news-updates/stable-diffusion-public-release"}]},{"slug":"model-platforms-and-copilots","label":"模型平台与 Copilot 成形","period":"2023","summary":"GPT-4、Claude 2、Llama 2、Qwen、Mistral 与 Gemini 建立多路线竞争，API、插件和 Copilot 开始进入软件与组织工作流。","projects":[{"name":"Microsoft 365 Copilot","status":"active","note":"将模型、组织数据和办公工作流绑定为企业产品。","url":"https://blogs.microsoft.com/blog/2023/03/16/introducing-microsoft-365-copilot-your-copilot-for-work/"},{"name":"Artifact","status":"acquired","note":"独立应用停止后，个性化技术由 Yahoo 收购并继续整合。","url":"https://www.yahooinc.com/press/yahoo-announces-the-acquisition-of-artifact-the-news-discovery-platform-created-by-instagram-cofounders-kevin-systrom-and-mike-krieger"}]},{"slug":"multimodal-and-reasoning","label":"多模态、推理与应用分化","period":"2024","summary":"原生多模态与推理时计算打开新的能力曲线，开放模型继续提升；应用市场同时进入整合，单一包装型产品开始失去空间。","projects":[{"name":"Inflection AI","status":"pivoted","note":"联合创始人与部分团队加入 Microsoft AI 后，原有消费助手路线进入组织与战略重构。","url":"https://blogs.microsoft.com/blog/2024/03/19/mustafa-suleyman-deepmind-and-inflection-co-founder-joins-microsoft-to-lead-copilot/"},{"name":"Stability AI","status":"pivoted","note":"在领导层和资本结构调整后继续聚焦开放生成媒体模型。","url":"https://stability.ai/news/stability-ai-secures-significant-new-investment"}]},{"slug":"agents-and-runtime","label":"Agent 从演示走向运行时","period":"2025 H1","summary":"DeepSeek-R1 重估效率边界，Operator、Deep Research、Claude Code 与 Agents SDK 将模型推进浏览器、代码和长任务。","projects":[{"name":"Adept","status":"pivoted","note":"部分团队加入 Amazon 后，公司转向更聚焦的 Agent 产品与企业路线。","url":"https://adept.ai/blog/adept-update/"},{"name":"Humane Ai Pin","status":"sunset","note":"消费者设备服务停止，说明新硬件入口必须同时解决产品价值与持续运营。","url":"https://support.humane.com/hc/en-us/articles/34243204841997-Ai-Pin-Consumers-FAQ"}]},{"slug":"product-and-protocol-convergence","label":"协议、产品与工作流合流","period":"2025 H2","summary":"MCP、A2A、Agent 产品与模型原生工具共同形成连接和执行层，竞争从功能演示转向真实工作流所有权。","projects":[{"name":"MCP","status":"active","note":"从单一厂商协议扩展为跨模型、工具和平台的连接生态。","url":"https://modelcontextprotocol.io/"},{"name":"Codex","status":"active","note":"编码能力从补全演化为云端与本地协同的 Agent 工作流。","url":"https://openai.com/codex/"}]},{"slug":"real-work-and-industry-concentration","label":"真实工作与产业再集中","period":"2026 H1—今天","summary":"旗舰模型围绕复杂工作、科学与长任务继续升级，收入、融资、算力和分发向头部集中，应用层开始按留存和结果重新分化。","projects":[{"name":"Claude Code","status":"active","note":"编码 Agent 向团队协作、长任务、SDK 与企业治理扩张。","url":"https://www.anthropic.com/claude-code"},{"name":"DeepSeek","status":"active","note":"开放推理与工程效率持续影响全球模型成本和部署讨论。","url":"https://api-docs.deepseek.com/news/"}]}],"tracks":[{"slug":"tech-evolution","thesis":"模型能力的主轴已从单纯扩大预训练，转向推理时计算、原生多模态、数据效率和可验证研究。","now":"旗舰模型正以可分配推理预算、多模态输入和科学评测扩展能力边界。","next":"下一阶段应关注能力增量能否跨基准复现，以及训练数据、评测污染、可靠性与成本是否透明。","stages":[{"start":"2022-08-01","end":"2022-12-31","period":"2022 H2","label":"生成能力产品化","summary":"扩散模型和对话模型把生成能力从研究展示变成公众可直接使用的产品。","interpretation":"技术竞争第一次同时受到交互体验、反馈数据和开放生态扩散速度影响。","chinaPosition":"国内团队快速进入对话、图像生成与应用验证，但尚未形成可比的开放基础模型矩阵。","nextSignal":"能力能否从流畅生成跨越到复杂推理、工具使用和稳定专业任务。"},{"start":"2023-01-01","end":"2023-06-30","period":"2023 H1","label":"能力平台化","summary":"GPT-4 抬高复杂任务上限，API、插件和 Copilot 把模型能力接入软件。","interpretation":"评测分数开始转化为开发接口和组织工作流，模型平台成为独立技术层。","chinaPosition":"中文模型密集发布，竞争重点从参数披露转向中文能力、合规和本地交付。","nextSignal":"开放权重与多区域模型能否缩小对少数闭源 API 的依赖。"},{"start":"2023-07-01","end":"2023-12-31","period":"2023 H2","label":"开放模型成形","summary":"Llama 2、Qwen、Mistral 与 Gemini 形成开放权重、多尺寸和原生多模态路线。","interpretation":"能力不再只由单一旗舰定义，许可、部署控制和衍生生态成为技术选择的一部分。","chinaPosition":"Qwen 建立中文开放模型起点，中国团队开始通过开源和云服务同时参与全球生态。","nextSignal":"开放模型能否在推理、长上下文和多模态上持续接近闭源前沿。"},{"start":"2024-01-01","end":"2024-06-30","period":"2024 H1","label":"原生多模态","summary":"语音、图像、视频与长上下文被纳入统一模型，交互从文本框扩展到实时环境。","interpretation":"模型结构、训练数据和端到端延迟开始共同决定可用能力，而非只比较文本榜单。","chinaPosition":"国内模型在长文本、视频生成和多尺寸部署形成差异化，但独立评测仍不足。","nextSignal":"多模态增量能否稳定转化为可重复任务完成，而不是发布演示。"},{"start":"2024-07-01","end":"2024-12-31","period":"2024 H2","label":"推理时计算","summary":"o1 把推理预算变成新的 Scaling 轴，Llama 3.1 和 Qwen2.5 同时扩大开放模型边界。","interpretation":"能力、延迟和成本不再是固定点，系统需要按任务动态分配计算。","chinaPosition":"Qwen 与 DeepSeek 以开放模型、训练效率和工程透明度形成不同于封闭平台的路径。","nextSignal":"强化学习和长链推理能否在外部评测与真实任务中复现，并控制泄漏与过度思考。"},{"start":"2025-01-01","end":"2025-06-30","period":"2025 H1","label":"效率与工具融合","summary":"DeepSeek-R1、Gemini 2.5、Qwen3 与 o3/o4-mini 把推理、工具和成本放进同一能力曲线。","interpretation":"模型价值从静态回答转向在预算约束下调用工具并完成可验证步骤。","chinaPosition":"开放推理和低成本训练成为中国路线的全球影响力来源。","nextSignal":"工具增强能力是否能减少重试、人工接管和端到端总成本。"},{"start":"2025-07-01","end":"2025-12-31","period":"2025 H2","label":"统一能力路由","summary":"旗舰产品开始在速度、推理深度、工具和模态之间自动路由。","interpretation":"用户购买的不再是单个模型，而是选择模型、记忆、工具与安全策略的系统。","chinaPosition":"Kimi K2、MiniMax M1 等开放路线继续扩大长上下文和 Agent 能力供给。","nextSignal":"自动路由能否在透明成本下保持一致行为，并允许企业审计和锁定策略。"},{"start":"2026-01-01","end":"9999-12-31","period":"2026—今天","label":"可验证真实能力","summary":"能力评价转向长任务、科学、具身、多模态和部署环境中的可复现表现。","interpretation":"前沿差距越来越取决于系统可靠性、持续数据反馈和评测设计，模型规模只是其中一个因素。","chinaPosition":"不同团队在开放推理、端侧、视频和具身模型形成优势，仍需更多跨环境可比证据。","nextSignal":"关注独立复现、生产失败分布、长期记忆污染和单位有效任务成本。"}],"lenses":[{"role":"ceo","question":"能力进步会改变哪个经营控制点？","answer":"控制点正从采购某个模型，迁移到掌握任务数据、评测集、工具权限和持续反馈。模型会快速替换，组织自己的验收标准和工作流上下文更难复制。","implications":["统一模型采购应改成按任务分层路由。","高风险流程必须保留可回放证据与人工接管。"],"actions":["选择三个高价值任务，建立成功率、重试、人工分钟和单位结果成本基线。","把内部评测集、权限策略和失败分类设为平台资产。"],"watch":["跨版本真实任务胜率是否提升，而非只看厂商榜单。","模型路由是否制造不可解释的成本和行为漂移。"],"evidenceSlugs":["openai-o1-test-time-reasoning","deepseek-r1-open-reasoning","gpt-5-5-real-work"]},{"role":"investor","question":"技术价值会沉淀在哪一层？","answer":"基础模型溢价仍存在，但可持续价值更可能沉淀在独有数据、可靠执行、推理基础设施和能证明结果改善的垂直系统。单纯包装最新模型的窗口持续缩短。","implications":["模型领先期缩短会压缩纯 API 转售毛利。","评测、推理优化和持续数据反馈成为新的尽调重点。"],"actions":["把收入增长拆成模型红利、工作流锁定和数据网络效应。","要求公司展示跨模型替换后的留存与毛利敏感性。"],"watch":["开放模型与闭源模型在真实任务上的价差。","能力提升是否带来更高使用频次和更低服务成本。"],"evidenceSlugs":["llama-3-1-open-frontier-model","deepseek-v3-efficient-frontier","gemma-4-open-model-efficiency"]},{"role":"cto","question":"能力与工程边界如何变化？","answer":"工程边界从单次 prompt 扩展到模型路由、上下文压缩、工具调用、状态恢复、评测和策略治理。可靠性是系统属性，不能由更强模型单独解决。","implications":["同一业务需要快速模型与深度推理模型协作。","长上下文不能替代检索、记忆清理和最小权限。"],"actions":["建立版本化评测集和回归检查。","记录每次工具调用、权限决策、重试和人工接管。"],"watch":["长任务中的错误累积和状态污染。","推理预算上升后延迟、吞吐和失败恢复的交换关系。"],"evidenceSlugs":["openai-o3-o4-mini-tools","gpt-5-unified-routing","deployment-simulation-model-behavior"]},{"role":"product","question":"下一项产品验证应该是什么？","answer":"验证重点应从用户是否喜欢回答，转向用户是否愿意把完整任务委派给系统，以及系统失败时能否被理解、修正和继续执行。","implications":["展示模型能力不等于形成可重复使用习惯。","信任来自可预览、可撤销和可验证的任务过程。"],"actions":["把一个完整任务拆成可观测的计划、执行和验收节点。","为高风险动作提供预览、确认、撤销和证据回链。"],"watch":["7/30 日任务复用率与委派深度。","用户修正后，同类错误是否持续减少。"],"evidenceSlugs":["claude-computer-use","gemini-3-5-frontier-action","blind-spots-bench-vision-language"]}]},{"slug":"agi-progress","thesis":"Agent 的关键变化是模型开始连接工具、维护任务状态，并承担跨软件的完整工作。聊天体验只是外在表现。","now":"编码、研究、浏览器操作和企业流程已经出现可重复的有限自治，软件入口正在从功能菜单转向任务委派。","next":"需要关注端到端完成率、记忆、权限、错误恢复、人工接管，以及 Agent 是否已经改变软件的交付和收费方式。","stages":[{"start":"2022-11-01","end":"2023-03-31","period":"2022 H2—2023 Q1","label":"对话委派","summary":"用户开始用自然语言委派信息整理、写作和代码建议。","interpretation":"Agent 的最初形态是可连续对话的助手，仍缺少稳定工具、状态和执行权限。","chinaPosition":"国内产品迅速验证对话入口，但主要能力仍依赖模型问答和人工复制结果。","nextSignal":"模型能否安全连接外部工具并把建议变成可执行步骤。"},{"start":"2023-04-01","end":"2023-12-31","period":"2023 Q2—Q4","label":"工具与 Copilot","summary":"插件、Assistants API 和 Copilot 把模型接入搜索、代码、文件和组织数据。","interpretation":"Agent 从生成文本转向调用受控工具，但编排仍由开发者预设。","chinaPosition":"国内云与应用团队快速接入工具链，开放协议与跨产品执行仍不成熟。","nextSignal":"系统是否能维护更长任务状态并处理工具失败。"},{"start":"2024-01-01","end":"2024-09-30","period":"2024 Q1—Q3","label":"推理与状态","summary":"长上下文和推理时计算增强计划能力，Agent 开始承担更长的知识工作。","interpretation":"任务长度提升后，记忆、验证和成本成为与模型能力同等重要的约束。","chinaPosition":"长上下文和低成本模型扩大本地 Agent 试验，但真实完成率公开证据较少。","nextSignal":"复杂计划能否跨软件执行，并在中途错误后恢复。"},{"start":"2024-10-01","end":"2025-03-31","period":"2024 Q4—2025 Q1","label":"通用界面行动","summary":"Computer Use、Operator 和 Deep Research 让模型操作浏览器、桌面与长检索链。","interpretation":"无需专用 API 的行动扩大覆盖面，也显著放大提示注入、误操作和权限风险。","chinaPosition":"开源推理缩小规划能力差距，通用浏览器和桌面 Agent 仍在验证可靠性。","nextSignal":"端到端完成率、人工接管率和安全隔离能否达到生产水位。"},{"start":"2025-04-01","end":"2025-06-30","period":"2025 Q2","label":"编码 Agent","summary":"Codex、Claude Code 和长任务模型开始连续完成仓库理解、执行和结果验证。","interpretation":"编码成为首个能用测试与 diff 验收结果的高价值 Agent 市场。","chinaPosition":"开放模型和本地工具链提供成本优势，但团队级采用、生态和分发证据仍不足。","nextSignal":"Agent 是否能稳定完成跨文件任务并减少 review 与返工总时间。"},{"start":"2025-07-01","end":"2025-12-31","period":"2025 H2","label":"协议与多 Agent","summary":"MCP、A2A 和 Agents SDK 把工具、上下文和 Agent 间协作抽象成运行时。","interpretation":"竞争从单一助手转向谁能控制连接、权限、可观测性和任务路由。","chinaPosition":"中国模型与应用进入协议兼容阶段，全球生态贡献和企业采用仍需持续观察。","nextSignal":"跨厂商互操作是否真实降低集成成本，而非增加协议层复杂度。"},{"start":"2026-01-01","end":"9999-12-31","period":"2026—今天","label":"真实工作与长任务","summary":"Agent 评价转向数小时任务、科学、具身和组织内持续运行。","interpretation":"模型峰值能力之外，恢复、记忆治理、责任边界和结果验收决定可部署性。","chinaPosition":"需观察中国 Agent 在高价值任务、企业部署和全球分发中的真实份额。","nextSignal":"长期运行是否降低单位结果成本，并在异常时留下足够审计证据。"}],"lenses":[{"role":"ceo","question":"Agent 会重写哪类组织边界？","answer":"可数字验收、需要跨系统搬运信息且等待时间长的流程会最先变化。组织需要明确新的责任、审批和异常升级方式，Bot 数量本身无法形成优势。","implications":["部分软件席位会变成任务预算。","管理跨度可能扩大，但失败责任会更集中。"],"actions":["选一个跨部门流程建立人工与 Agent 对照组。","明确哪些动作可自动执行、哪些必须审批。"],"watch":["端到端周期、人工接管和返工是否同时下降。","供应商是否提供可导出的任务记录和权限审计。"],"evidenceSlugs":["claude-computer-use","openai-codex-cloud-agent","gpt-5-5-real-work"]},{"role":"investor","question":"Agent 价值向哪一层迁移？","answer":"价值从聊天入口迁向拥有工作流、权限和结果数据的运行时。能证明任务成功率和留存的垂直系统，比通用 Agent 壳层更可能保有定价权。","implications":["Agent 数量与消息量不是收入质量。","分发入口和系统记录权比单次能力领先更持久。"],"actions":["尽调真实付费任务、失败成本和续费驱动。","测试更换底层模型后产品差异是否仍存在。"],"watch":["按结果收费是否提升毛利而非转嫁不可控成本。","平台厂商是否向同一工作流下沉。"],"evidenceSlugs":["model-context-protocol","codex-general-availability","chatgpt-agent-research-to-action"]},{"role":"cto","question":"Agent 运行时的硬边界是什么？","answer":"硬边界是权限、状态一致性、幂等、回滚、可观测性和外部内容注入。模型更强不会自动解决分布式系统和安全工程问题。","implications":["工具契约必须可验证且最小权限。","长任务需要 checkpoint、预算和人工中断。"],"actions":["为工具调用定义 schema、超时、重试和补偿。","按失败类别记录环境、模型、步骤和恢复结果。"],"watch":["长任务中的重复副作用和隐性权限扩张。","MCP/A2A 兼容是否经过真实互操作测试。"],"evidenceSlugs":["model-context-protocol","google-a2a-agent-protocol","openai-responses-agents-sdk"]},{"role":"product","question":"怎样证明 Agent 不是一次性演示？","answer":"证明标准是同一用户反复委派更深任务，并愿意让系统在明确边界内采取行动。需要同时衡量完成率、信任、修正成本和结果价值。","implications":["成功演示可能掩盖长尾失败。","用户需要理解系统当前状态和下一动作。"],"actions":["把任务进度、证据、等待和失败原因做成产品界面。","为每类高风险动作设计确认与撤销。"],"watch":["30 日任务复用率和平均委派步骤数。","用户取消、接管和回退发生在哪个节点。"],"evidenceSlugs":["openai-operator-browser-agent","openai-deep-research","claude-opus-4-7-agent-work"]}]},{"slug":"commercialization","thesis":"商业化从 API token 消费，转向订阅、席位、任务结果和嵌入核心工作流。","now":"编码和知识工作率先形成高价值付费，模型厂商直接向应用和运行时扩张。","next":"关键指标将从调用量转为完成任务、留存、毛利和组织产能。","stages":[{"start":"2022-08-01","end":"2022-12-31","period":"2022 H2","label":"大众需求验证","summary":"ChatGPT 和生成图像证明用户会主动寻找生成式 AI 产品。","interpretation":"市场从技术可行性转向留存、分发和使用成本问题。","chinaPosition":"消费应用跟进迅速，但商业模式主要依赖订阅探索和流量分发。","nextSignal":"高频使用能否形成付费，以及 API 是否打开应用创业。"},{"start":"2023-01-01","end":"2023-06-30","period":"2023 H1","label":"API 与 Copilot","summary":"模型 API 降低集成门槛，Microsoft 365 Copilot 把 AI 放进组织软件。","interpretation":"商业价值从独立聊天产品扩展到已有分发和数据权限的应用。","chinaPosition":"国内大厂以云和办公入口推进，创业公司集中验证内容与效率工具。","nextSignal":"企业是否为席位付费，以及应用能否建立模型之外的壁垒。"},{"start":"2023-07-01","end":"2024-06-30","period":"2023 H2—2024 H1","label":"多模型供给","summary":"开放模型和多家 API 供应商降低单一模型依赖，产品竞争转向数据与工作流。","interpretation":"API 包装的稀缺性下降，采购开始比较成本、隐私、延迟和部署方式。","chinaPosition":"Qwen 等开放模型扩大私有化供给，价格竞争加速市场教育。","nextSignal":"产品是否能在更换模型后保持留存、质量和毛利。"},{"start":"2024-07-01","end":"2024-12-31","period":"2024 H2","label":"企业与治理门槛","summary":"推理能力、计算机使用和 AI Act 同时提高可做任务与交付责任。","interpretation":"企业采购同时评估模型能力、治理、权限、审计和长期服务。","chinaPosition":"消费入口扩张快，企业付费、合规交付和全球服务能力仍需更多证据。","nextSignal":"部署能否从试点进入重复采购，并形成可审计 ROI。"},{"start":"2025-01-01","end":"2025-12-31","period":"2025","label":"Agent 产品","summary":"研究、浏览、编码和开发平台出现可按任务验收的 Agent 产品。","interpretation":"定价从 token 和功能席位向任务额度、团队计划和结果价值移动。","chinaPosition":"模型价格竞争激烈，完整应用流程和全球分发能力仍在形成。","nextSignal":"Agent 使用是否提升留存、扩张收入和组织级标准化。"},{"start":"2026-01-01","end":"9999-12-31","period":"2026—今天","label":"工作流所有权","summary":"模型厂商、云、办公软件和垂直应用争夺完整任务入口与组织预算。","interpretation":"价值取决于谁控制上下文、权限、记录和结果反馈，而非谁先发布功能。","chinaPosition":"云与大厂分发占优，创业公司需要以行业数据、交付和跨境能力建立壁垒。","nextSignal":"比较任务留存、单位经济、净收入留存和供应商替换成本。"}],"lenses":[{"role":"ceo","question":"商业控制点正在迁移到哪里？","answer":"控制点从软件功能和模型 API 迁移到拥有用户任务入口、组织上下文、审批权和结果记录的一层。企业需要防止关键流程被单一模型供应商锁定。","implications":["采购应同时保留模型可替换性与数据可携带性。","AI 预算会与软件、外包和人力预算重新分配。"],"actions":["建立按任务而非按席位的 ROI 台账。","在合同中明确数据、日志、模型切换和退出边界。"],"watch":["试点转生产的周期与扩张率。","单位任务毛利是否随使用增加而改善。"],"evidenceSlugs":["microsoft-365-copilot-introduced","codex-general-availability","google-io-2026-ai-platform-bundle"]},{"role":"investor","question":"收入质量怎样与 AI 热度分开？","answer":"评估重点应放在付费任务是否高频、是否嵌入核心流程、模型成本能否被规模摊薄，以及续费是否来自真实结果。ARR 标签和补贴都可能掩盖实际采用情况。","implications":["席位增长可能掩盖低激活和高推理成本。","服务收入过高可能说明产品尚未标准化。"],"actions":["重建 cohort、毛利和使用深度三张表。","压力测试模型降价、平台捆绑和客户自建。"],"watch":["净收入留存与活跃任务增长是否同向。","头部客户集中和底层模型依赖。"],"evidenceSlugs":["chatgpt-whisper-apis","anthropic-series-h-scale","openai-acquires-ona"]},{"role":"cto","question":"商业化对技术架构提出什么要求？","answer":"多模型路由、成本归因、租户隔离、审计和可降级能力会直接决定毛利与交付速度。架构必须允许能力升级而不重写业务流程。","implications":["每个任务需要成本、质量和延迟追踪。","企业功能不能作为后期补丁。"],"actions":["把模型、工具和业务策略解耦为版本化契约。","建立按客户和任务的成本与失败观测。"],"watch":["供应商变更导致的行为回归。","高峰时延迟、降级与 SLA 违约。"],"evidenceSlugs":["openai-responses-agents-sdk","model-context-protocol","claude-fable-5-microsoft-foundry"]},{"role":"product","question":"怎样从功能使用走向可持续付费？","answer":"产品必须让用户更快得到可验收结果，并在团队中积累模板、权限、历史和协作价值。一次生成质量不足以形成续费。","implications":["使用频次需要与任务价值一起衡量。","团队协作和治理会成为扩张杠杆。"],"actions":["定义从首次成功到团队扩张的关键行为链。","展示结果证据、节省时间和可复用资产。"],"watch":["7/30/90 日留存与任务深度。","用户是否把输出带回原系统继续工作。"],"evidenceSlugs":["chatgpt-research-preview","openai-codex-cloud-agent","ai-co-clinician-healthcare-model"]}]},{"slug":"investing","thesis":"资本主线从模型融资，扩展到芯片、数据中心、电力、云承诺和高价值应用。","now":"头部模型公司正在形成云平台级资本密度，基础设施与收入规模同步放大。","next":"需要同时跟踪训练/推理 CapEx、收入质量、毛利和平台锁定。","stages":[{"start":"2022-08-01","end":"2023-06-30","period":"2022 H2—2023 H1","label":"平台联盟","summary":"模型公司与云平台形成算力、资本和分发绑定。","interpretation":"前沿模型的资本需求使云承诺和战略投资成为竞争结构，而非普通融资。","chinaPosition":"国内由云厂商与大模型团队共同投入，公开可比融资和算力数据有限。","nextSignal":"资本是否转化为模型领先、企业分发和可持续收入。"},{"start":"2023-07-01","end":"2024-06-30","period":"2023 H2—2024 H1","label":"模型融资与应用整合","summary":"基础模型融资放大，应用层同时出现收购、团队迁移和路线转向。","interpretation":"资本开始区分基础设施密集型平台与缺少工作流壁垒的应用。","chinaPosition":"模型创业密集融资，但价格竞争与备案交付提高现金消耗。","nextSignal":"应用公司能否在平台下沉前证明留存、分发和数据壁垒。"},{"start":"2024-07-01","end":"2025-06-30","period":"2024 H2—2025 H1","label":"训练与推理建设","summary":"Blackwell、AI Factory、数据中心和电力承诺成为资本主线。","interpretation":"投资对象从模型公司扩展到芯片、网络、云、电力和施工周期。","chinaPosition":"算力限制推动系统效率、国产适配和开放模型路线。","nextSignal":"CapEx 是否带来可售算力、利用率和推理收入，而非闲置承诺。"},{"start":"2025-07-01","end":"2025-12-31","period":"2025 H2","label":"资本再集中","summary":"头部模型、云和基础设施平台获得更大融资与长期合作承诺。","interpretation":"规模优势扩大，但估值开始依赖收入增速、毛利和持续融资能力。","chinaPosition":"DeepSeek 等效率路线证明资本规模不是唯一变量，融资与全球分发差距仍在。","nextSignal":"观察融资条款、云依赖、客户集中和单位推理毛利。"},{"start":"2026-01-01","end":"9999-12-31","period":"2026—今天","label":"收入、退出与纪律","summary":"行业同时出现大额融资、并购、人才交易与对真实收入质量的追问。","interpretation":"资本叙事从拥有模型转向能否控制工作流、基础设施现金流和可持续退出路径。","chinaPosition":"开放生态与工程成本仍有优势，但上市、并购和跨境资本路径更受约束。","nextSignal":"把增长、毛利、CapEx、债务、客户集中和退出可行性放进同一模型。"}],"lenses":[{"role":"ceo","question":"资本集中会怎样影响经营选择？","answer":"企业应假设模型与算力供应将长期集中，同时保持关键任务的供应商替换与成本可见性。过早押注单一平台会把采购优惠变成未来迁移成本。","implications":["长期云承诺需要与真实需求爬坡匹配。","小团队应避开纯资本规模竞争。"],"actions":["对关键供应商做价格、容量和退出压力测试。","把自建、托管和 API 的总拥有成本放在同一模型。"],"watch":["承诺用量与实际利用率偏差。","融资或并购是否改变产品路线和服务连续性。"],"evidenceSlugs":["nvidia-blackwell-ultra-reasoning-factory","anthropic-series-h-scale","openai-acquires-ona"]},{"role":"investor","question":"怎样识别资本支出背后的价值？","answer":"必须把算力规模拆成可用容量、利用率、单位收入、合同期限和融资结构。大额 CapEx 只有在需求、毛利和现金回收可验证时才是壁垒。","implications":["芯片采购不等于可售推理产能。","云关联融资可能同时带来分发与锁定。"],"actions":["建立 CapEx 到收入的滞后桥接表。","分别估值模型、基础设施和应用现金流。"],"watch":["推理收入增速与折旧、能源成本。","客户集中、关联交易和再融资需求。"],"evidenceSlugs":["nvidia-blackwell-ultra-reasoning-factory","deepseek-v3-efficient-frontier","anthropic-series-h-scale"]},{"role":"cto","question":"资本周期会约束哪些技术路线？","answer":"可获得芯片、网络、电力和云额度会反向决定模型规模、训练频率与推理架构。效率工程既是成本优化，也是供应安全能力。","implications":["硬件路线需要软件栈和模型结构共同适配。","推理增长会把容量规划从训练峰值转向持续负载。"],"actions":["记录有效利用率、故障恢复和每任务资源消耗。","验证替代硬件与模型压缩的质量边界。"],"watch":["容量瓶颈从芯片转向网络、电力或冷却。","新硬件交付时间与软件成熟度错配。"],"evidenceSlugs":["nvidia-blackwell-ultra-reasoning-factory","decoupled-diloco-resilient-distributed-training","gemma-4-open-model-efficiency"]},{"role":"product","question":"资本变化如何传导到产品？","answer":"融资与基础设施扩张会带来更低价格和更强能力，也可能推动供应商优先服务高价值工作流。产品需要把价格红利转成用户价值，而非无限增加生成量。","implications":["降价会加速功能同质化。","平台并购可能改变 API、配额和路线图。"],"actions":["把成本下降投入更深任务或更好验证。","为供应商变更设计用户可感知的质量回归检查。"],"watch":["模型降价是否改善留存和毛利。","并购后产品承诺与接口稳定性。"],"evidenceSlugs":["deepseek-r1-open-reasoning","openai-acquires-ona","google-io-2026-ai-platform-bundle"]}]},{"slug":"global-innovation","thesis":"全球 AI 创新由多条路线共同推动：中国的开放模型、工程效率、应用分发和本土算力，与美国的平台能力、欧洲治理和全球开源生态相互影响。","now":"中国团队已在开放推理、模型效率、视频生成和应用规模上形成独立影响，差异更多来自生态、算力、市场与分发路径。","next":"观察跨区域开发者采用、全球收入、开放生态贡献、国产算力的端到端验证和产品长期留存。","stages":[{"start":"2022-08-01","end":"2022-12-31","period":"2022 H2","label":"开放生成扩散","summary":"Stable Diffusion 证明开放权重可以在全球开发者和创作者中快速扩散。","interpretation":"创新速度开始由社区改造、工具链和本地部署共同推动。","chinaPosition":"中国开发者快速进入开放生成生态，但基础模型贡献和许可影响仍有限。","nextSignal":"开放生态能否从图像扩展到通用语言模型。"},{"start":"2023-01-01","end":"2023-06-30","period":"2023 H1","label":"区域模型起跑","summary":"全球团队围绕语言、监管和本地部署建立区域模型路线。","interpretation":"单一美国平台之外，数据主权、语言与产业环境开始塑造技术路线。","chinaPosition":"国内大模型密集发布并进入备案与产品验证。","nextSignal":"是否形成可下载、可评测、可持续维护的开放模型。"},{"start":"2023-07-01","end":"2023-12-31","period":"2023 H2","label":"开放模型多中心","summary":"Llama 2、Qwen、Mistral 与 Gemini 形成美国、中国、欧洲多路线供给。","interpretation":"全球竞争从追随单一榜单转向许可、语言、尺寸、部署和生态的组合。","chinaPosition":"Qwen-7B 开放使中国模型开始直接进入全球开发者比较。","nextSignal":"多区域模型能否在前沿能力和生态采用上持续迭代。"},{"start":"2024-01-01","end":"2024-06-30","period":"2024 H1","label":"多模态与应用分化","summary":"不同区域围绕长上下文、视频、端侧和行业应用形成专长。","interpretation":"创新不再只有一个能力轴，市场与数据条件决定路线优先级。","chinaPosition":"中国团队在视频生成、长文本和消费分发形成独立产品节奏。","nextSignal":"产品热度能否转化为全球留存、收入与开发者生态。"},{"start":"2024-07-01","end":"2025-03-31","period":"2024 H2—2025 Q1","label":"开放效率转折","summary":"Qwen2.5、DeepSeek-V3/R1 把模型矩阵、训练效率和开放推理带入全球主线。","interpretation":"资本规模之外，架构、数据和工程效率成为可观察的创新变量。","chinaPosition":"中国模型从本地替代叙事进入影响全球成本、开源和推理路线的阶段。","nextSignal":"独立复现、海外开发者采用和衍生生态是否持续。"},{"start":"2025-04-01","end":"2025-12-31","period":"2025 Q2—Q4","label":"Agent 与开放生态","summary":"Qwen3、MiniMax M1、Kimi K2 等模型转向混合推理、长上下文和 Agent。","interpretation":"竞争已经从模型发布扩展到工具兼容、全球分发和完整应用流程。","chinaPosition":"开放模型供给保持活跃，企业采用、国际服务和商业收入成为新门槛。","nextSignal":"观察全球 API 使用、开源贡献、企业案例和区域合规。"},{"start":"2026-01-01","end":"9999-12-31","period":"2026—今天","label":"全球采用与产业落地","summary":"各区域围绕真实工作、具身、科学、算力和治理形成不同组合优势。","interpretation":"领先不应由单一榜单定义，而要比较能力、成本、生态、收入和供应链韧性。","chinaPosition":"竞争进入产品留存、企业采用、全球分发与国产算力协同。","nextSignal":"用同维度证据比较开发者份额、全球收入、生产部署和国产算力的端到端表现。"}],"lenses":[{"role":"ceo","question":"全球创新版图会改变哪些区域决策？","answer":"模型选择应按语言、数据边界、供应安全、生态和任务表现组合决策，而不是默认单一全球冠军。多区域供应可以提高韧性，也会增加治理复杂度。","implications":["中国与海外业务可能需要不同模型组合。","开源模型提供控制力但增加运维责任。"],"actions":["为关键区域建立同任务、同成本评测。","明确跨境数据、模型许可和退出方案。"],"watch":["区域供应商的长期服务与全球节点覆盖。","监管变化是否影响模型、数据或芯片可得性。"],"evidenceSlugs":["qwen-2-5-open-model-matrix","deepseek-r1-open-reasoning","kimi-k2-agentic-open-model"]},{"role":"investor","question":"怎样避免把全球比较做成单一排名？","answer":"应分别比较能力、成本、开放生态、分发、收入和供应链。中国团队的效率与开放影响力可以领先，但全球企业收入和分发仍可能受限。","implications":["技术领先与商业领先可能分离。","区域政策和云生态会改变可实现市场。"],"actions":["建立同维度区域 scorecard。","拆分国内收入、海外收入、开发者采用和生态贡献。"],"watch":["海外留存与付费是否同步增长，避免只看下载量。","国产算力适配能否通过完整生产验证。"],"evidenceSlugs":["qwen-3-hybrid-thinking","minimax-m1-long-context-reasoning","deepseek-v3-efficient-frontier"]},{"role":"cto","question":"多区域模型供给怎样进入架构？","answer":"需要标准化模型接口、评测、内容安全和数据路由，同时保留各模型在语言、工具和部署上的差异。最低公分母抽象会损失能力。","implications":["多模型不是简单 endpoint 切换。","许可、权重来源和区域部署需要可审计。"],"actions":["建立模型能力 manifest 与区域策略。","对同一任务持续回归质量、成本和安全。"],"watch":["模型升级造成的语言与工具行为漂移。","开源依赖、权重和训练数据许可变化。"],"evidenceSlugs":["llama-2-open-model-release","qwen-7b-open-release","mistral-7b-release"]},{"role":"product","question":"全球产品应该先验证什么？","answer":"先验证不同地区用户是否在同一核心任务上持续获得价值，再决定统一产品或区域化。翻译界面不能替代本地工作流、支付、合规和分发。","implications":["模型语言能力不等于市场进入能力。","本地生态合作可能比功能复制更重要。"],"actions":["按地区追踪激活、任务留存和付费原因。","为区域差异建立产品假设而非只改文案。"],"watch":["海外用户是否形成自然复用和口碑。","分发成本、支付与合规是否吞噬单位经济。"],"evidenceSlugs":["qwen-2-5-open-model-matrix","kimi-k2-agentic-open-model","deepmind-apac-climate-accelerator"]}]},{"slug":"model-economics","thesis":"单位 token 价格仍在下降，企业评估成本时开始转向“一次任务可靠完成需要多少钱”。","now":"MoE、FP8、推理预算和模型路由共同决定成本曲线。","next":"采购需要比较成功率、重试、延迟、人工接管和总拥有成本。","stages":[{"start":"2022-11-01","end":"2023-06-30","period":"2022 H2—2023 H1","label":"API 成本下降","summary":"对话与语音 API 标准化并快速降价，应用验证门槛下降。","interpretation":"模型成本首次成为可直接进入产品单位经济的变量。","chinaPosition":"国内团队通过云服务和价格竞争快速扩大可及性。","nextSignal":"开放权重是否进一步压低推理与锁定成本。"},{"start":"2023-07-01","end":"2023-12-31","period":"2023 H2","label":"开放部署选择","summary":"Llama 2、Qwen、Mistral 提供不同尺寸和本地部署路径。","interpretation":"采购从 API 单价比较扩展到硬件、运维、许可和定制成本。","chinaPosition":"开放中文模型扩大私有化选择，真实运维成本仍缺少统一口径。","nextSignal":"同质量下的吞吐、显存和部署复杂度能否持续改善。"},{"start":"2024-01-01","end":"2024-06-30","period":"2024 H1","label":"MoE 与推理优化","summary":"稀疏模型、量化和推理框架降低单位激活计算。","interpretation":"参数规模不再等于每次请求成本，系统软件和硬件利用率影响扩大。","chinaPosition":"国内价格战与工程优化同时推进，形成低价 API 供给。","nextSignal":"低价是否在可靠性、峰值容量和服务质量下仍成立。"},{"start":"2024-07-01","end":"2024-12-31","period":"2024 H2","label":"推理预算","summary":"o1 让推理成本随任务难度和思考长度变化。","interpretation":"固定 token 单价无法解释复杂任务的真实成本与价值。","chinaPosition":"Qwen 与 DeepSeek 加速转向推理与效率路线。","nextSignal":"更多计算是否稳定提高可验证结果，还是带来过度思考和泄漏。"},{"start":"2025-01-01","end":"2025-03-31","period":"2025 Q1","label":"效率冲击","summary":"DeepSeek-V3/R1 把训练与推理效率带入全球成本讨论。","interpretation":"架构、数据和工程能力可以改变对前沿模型资本门槛的估计。","chinaPosition":"DeepSeek 把效率、开放权重和价格变成全球竞争力。","nextSignal":"外部复现、真实部署成本和持续服务能力。"},{"start":"2025-04-01","end":"2025-09-30","period":"2025 Q2—Q3","label":"工具与 Agent 成本","summary":"模型开始为每个任务调用搜索、代码、浏览器和多个子步骤。","interpretation":"重试、工具费用、长上下文和等待时间使单次请求成本失真。","chinaPosition":"低价模型有利于扩大尝试，但端到端成功率决定实际优势。","nextSignal":"比较一次成功任务的总调用、重试和人工接管。"},{"start":"2025-10-01","end":"2026-03-31","period":"2025 Q4—2026 Q1","label":"系统级 TCO","summary":"自动路由、缓存、压缩和多模型组合成为成本控制核心。","interpretation":"采购单位从模型 token 迁移到包含平台、工程和治理的总拥有成本。","chinaPosition":"开放模型和本地部署提供控制力，SLA 与运维成熟度决定企业价值。","nextSignal":"模型切换、容量峰值和治理要求下的真实年度成本。"},{"start":"2026-04-01","end":"9999-12-31","period":"2026 Q2—今天","label":"单位有效结果","summary":"行业开始按可靠完成的任务、节省的人工和业务结果衡量成本。","interpretation":"便宜但失败的调用可能比昂贵但一次完成的模型更贵。","chinaPosition":"需要补齐真实 ROI、SLA、全球交付和企业续费证据。","nextSignal":"公开成功率、失败分布、人工分钟、毛利和长期留存。"}],"lenses":[{"role":"ceo","question":"应该用什么单位管理 AI 成本？","answer":"用一次被业务接受的结果作为成本单位。计算时应包含模型、工具、重试、人工复核、失败损失和治理成本，不能只看 token 或调用次数。","implications":["低单价可能被低成功率抵消。","不同任务应配置不同质量与预算上限。"],"actions":["为核心任务建立 cost-per-accepted-outcome。","把人工接管和失败损失纳入采购复盘。"],"watch":["规模增加后单位结果成本是否下降。","供应商降价是否伴随容量或质量变化。"],"evidenceSlugs":["chatgpt-whisper-apis","deepseek-r1-open-reasoning","gpt-5-unified-routing"]},{"role":"investor","question":"成本下降会把利润留给谁？","answer":"底层降价会扩大需求，但利润更可能留在拥有分发、专有工作流、数据和基础设施利用率的一层。纯模型转售会持续承压。","implications":["推理毛利需要结合折旧和容量利用率。","应用毛利要剔除模型补贴和人工服务。"],"actions":["按任务拆解收入、推理成本和服务成本。","压力测试价格每年下降时的竞争位置。"],"watch":["降价是否带来需求弹性和留存。","自建与 API 的临界规模是否真实到达。"],"evidenceSlugs":["deepseek-v3-efficient-frontier","nvidia-blackwell-ultra-reasoning-factory","gemma-4-open-model-efficiency"]},{"role":"cto","question":"怎样工程化控制单位任务成本？","answer":"通过模型路由、缓存、上下文压缩、批处理、工具预算和失败早停共同优化；任何单点优化都必须在质量回归下验证。","implications":["长上下文可能增加无效计算。","重试策略既影响可靠性也影响尾部成本。"],"actions":["记录每任务 token、工具、延迟、重试和结果。","为不同风险等级设模型与预算策略。"],"watch":["缓存命中、路由漂移和峰值容量。","量化或替代模型导致的长尾质量损失。"],"evidenceSlugs":["openai-o1-test-time-reasoning","decoupled-diloco-resilient-distributed-training","gemma-4-12b-unified-multimodal"]},{"role":"product","question":"如何把成本约束转成更好的产品？","answer":"让用户感知任务预算、等待和结果置信度，并把高成本计算留给高价值步骤。盲目追求即时或无限生成会伤害毛利和信任。","implications":["速度、质量和价格需要按场景表达。","用户应能选择快速草稿或深度执行。"],"actions":["设计明确的任务模式与预算提示。","用结果接受率而非生成次数优化体验。"],"watch":["高成本模式的付费意愿和复用率。","等待时间是否换来可感知的结果提升。"],"evidenceSlugs":["openai-o1-test-time-reasoning","gpt-5-unified-routing","gemini-2-5-thinking-model"]}]}]}