{"generated":"2026-11-02","note":"The structure of the /landscape/ map pages, parsed on every build (graph/kb/landscape_structure.py); the page HTML stays the source (the five structure maps are redirected into the entity pages, the two ledgers are still served). resolution: entity = the card is about the listed library entities, concept = an idea or a trend with no node of its own.","maps":{"agent-dev-stack":{"title":"Agent 开发栈","h1":"Agent 开发栈","layers":[{"id":"layer:agent-dev-stack--1","num":1,"step":"Layer 01","name":"底座模型","role":"Agent 的智能核心，决定推理、工具调用、长上下文的能力上限。2026 年最大的变化： 中国开源模型在 agentic 跑分上已经追到个位数差距，价格却便宜 5-20 倍 ，月之暗面 Kimi K2.7-Code 是第一个进入 GitHub Copilot 模型选择器的中国开源模型。","cards":["player-claude-opus-5","player-claude-sonnet-5","player-gpt-5-6","player-gemini-3","player-qwen36","player-deepseek-v4","player-kimi-k27","player-glm-47","player-minimax-m3","player-mistral-3"]},{"id":"layer:agent-dev-stack--2","num":2,"step":"Layer 02","name":"编排框架","role":"把模型包装成能自主决策、调用工具、管理状态的 Agent。2026 年格局： 头部整合、腰部分裂 ——微软把 AutoGen 和 Semantic Kernel 都打入维护模式，统一成 Microsoft Agent Framework；与此同时中国开源 LLMOps 平台（Dify star 数已超 LangChain）成为全球采用量最大的赢家。","cards":["player-langgraph","player-crewai","player-ms-agent-framework","player-openai-agents-sdk","player-claude-agent-sdk","player-google-adk","player-dify","player-openrouter"]},{"id":"layer:agent-dev-stack--3","num":3,"step":"Layer 03","name":"上下文协议","role":"让 Agent 能安全、标准化地连接外部工具和其它 Agent。 2026 年最大的标准化事件：MCP 与 A2A 的治理权都并入 Linux 基金会新设的中立机构 AAIF ，此前\"协议战争\"的焦虑基本解除。","cards":["player-mcp","player-aaif","player-a2a","player-agntcy","player-agent-plugins"]},{"id":"layer:agent-dev-stack--4","num":4,"step":"Layer 04","name":"记忆与上下文管理","role":"让 Agent\"记得住\"跨会话信息、管好超长上下文。 格局正在向上迁移 ：pgvector 成熟后悄悄吃掉了大部分\"够用就行\"的检索需求，真正的差异化和融资都转移到了\"记忆\"这个更高的抽象层。","cards":["player-pgvector","player-mem0","player-zep","player-letta","player-skillstate"]},{"id":"layer:agent-dev-stack--5","num":5,"step":"Layer 05","name":"编码 Agent","role":"目前 Agent 最成熟的落地场景，开发者日常接触最密集的产品。 2026 年洗牌很剧烈 ：Claude Code 在开发者偏好上反超 GitHub Copilot（JetBrains调研：\"最受欢迎\"46% vs 9%），Windsurf 作为独立品牌已经不存在了（被 Cognition 并入改名 Devin Desktop），Amazon 砍掉 Q Developer 转投 Kiro。估值普遍远超实际用量信号。","cards":["player-claude-code","player-cursor","player-copilot","player-codex","player-devin","player-replit-agent","player-deepseek-harness","player-grok-bot"]},{"id":"layer:agent-dev-stack--6","num":6,"step":"Layer 06","name":"可观测性与评测","role":"生产环境里\"这个 Agent 到底靠不靠谱\"的量化答案。安全评测（Inspect AI）和生产监控（LangSmith/Braintrust/Arize）正在成为两条并行赛道。","cards":["player-langsmith","player-braintrust","player-inspect-ai","player-arize"]},{"id":"layer:agent-dev-stack--7","num":7,"step":"Layer 07","name":"沙盒执行环境","role":"给 Agent 一个能放心执行代码/操作的隔离空间。已经成为独立且融资活跃的基础设施品类。","cards":["player-e2b","player-daytona","player-modal","player-browserbase"]},{"id":"layer:agent-dev-stack--8","num":8,"step":"Layer 08","name":"多智能体协作","role":"多个 Agent 之间怎么分工、通信、交接任务。行业估计约 70% 的生产级多智能体部署用的是\"编排者-执行者\"（orchestrator-worker）拓扑。","cards":["player-a2a-8","player-crewai-flows","player-claude-subagents"]},{"id":"layer:agent-dev-stack--9","num":9,"step":"Layer 09","name":"部署托管平台","role":"Agent 最终运行在哪、怎么计费。每家云厂商都在 2026 年推出了专门面向 Agent 的托管产品，计费单元也从\"算力时长\"细化到了\"每次工具调用\"\"每条记忆记录\"这类 Agent 原生指标。","cards":["player-bedrock-agentcore","player-gemini-enterprise","player-foundry-agent-service","player-vercel-agent-stack","player-cf-agents-sdk","player-tensorrt-connect"]}],"flows":[{"from":"layer:agent-dev-stack--1","to":"layer:agent-dev-stack--2"},{"from":"layer:agent-dev-stack--2","to":"layer:agent-dev-stack--4"},{"from":"layer:agent-dev-stack--3","to":"layer:agent-dev-stack--5"},{"from":"layer:agent-dev-stack--5","to":"layer:agent-dev-stack--7"},{"from":"layer:agent-dev-stack--8","to":"layer:agent-dev-stack--5"}],"trends":[{"head":"MCP 赢了，标准治理权收归中立机构。","text":"MCP 赢了，标准治理权收归中立机构。 MCP 已是无可争议的工具连接默认协议（约9700万次/月SDK下载，约28%财富500强生产环境在用），与 A2A 一起并入 Linux 基金会新设的 Agentic AI Foundation——此前\"协议战争\"的焦虑基本解除。"},{"head":"编码 Agent 市场剧烈洗牌。","text":"编码 Agent 市场剧烈洗牌。 Claude Code 在开发者偏好上反超 GitHub Copilot；Windsurf 作为独立产品已不存在（并入 Cognition 改名 Devin Desktop）；Amazon 砍掉 Q Developer 转投 Kiro；估值普遍远超实际用量信号（Cursor 传闻500-600亿估值，但用量数据在软化）。"},{"head":"框架层\"头部整合、腰部分裂\"。","text":"框架层\"头部整合、腰部分裂\"。 微软把 AutoGen 和 Semantic Kernel 都打入维护模式，统一成 Microsoft Agent Framework；与此同时中国开源 LLMOps 平台（Dify、Coze、RAGFlow）成为全球采用量最大的赢家。"},{"head":"中国开源模型追近 agentic 跑分差距，价格便宜 5-20 倍，且开始进入西方工具链。","text":"中国开源模型追近 agentic 跑分差距，价格便宜 5-20 倍，且开始进入西方工具链。 Kimi K2.7-Code 发布仅19天就进入 GitHub Copilot 模型选择器，是首个做到这一点的中国开源模型。"},{"head":"\"上下文工程\"成为独立学科，记忆抽象层成为新战场。","text":"\"上下文工程\"成为独立学科，记忆抽象层成为新战场。 pgvector 的成熟把\"够用\"的检索需求悄悄吃进了 Postgres，真正的差异化和融资转移到了 Mem0/Zep/Letta 这类记忆层产品。"},{"head":"沙盒执行与\"Agent原生托管\"成为独立基础设施品类。","text":"沙盒执行与\"Agent原生托管\"成为独立基础设施品类。 E2B/Daytona/Modal 等专做 Agent 代码执行融资活跃，每家云厂商都推出了按 Agent 原生指标计费的托管产品。"},{"head":"编码Agent的\"骨架\"层正在同一窗口期集体开源，且带出了\"分数不可比\"的量化证据。","text":"编码Agent的\"骨架\"层正在同一窗口期集体开源，且带出了\"分数不可比\"的量化证据。 DeepSeek Harness（MIT）与OpenAI Codex执行框架（Apache-2.0）先后开源，Codex顺手给出的数字是：只调两个骨架开关，同一模型ARC-AGI-3成绩就从13.3%升到38.3%——\"模型能力\"与\"骨架能力\"在基准分里已经无法分离，跨厂商agentic跑分对比第一次系统性失真。"},{"head":"模型路由/网关正从厂商自建走向标准化基础设施，并被资本重新定价。","text":"模型路由/网关正从厂商自建走向标准化基础设施，并被资本重新定价。 OpenRouter被支付公司Stripe以70亿美元以上收购（较三个月前估值跳升逾5倍），NVIDIA同期推出跨厂商路由框架NeMo Switchyard——\"选模型\"和\"选骨架/路由\"正在变成两个独立的采购决策。"},{"head":"头部编码Agent产品出现横向合并。","text":"头部编码Agent产品出现横向合并。 Cursor被SpaceX/xAI收购、与Grok Bot/Grok Build/Grok API打包进\"SpaceXAI\"，是\"算力议价权投降\"而非溢价故事的又一例证——自研模型/推理长期被上游卡住的产品，最终被算力方直接吃下。"},{"head":"模型供应权正式变成阵营站队的筹码，不只是市场传闻。","text":"模型供应权正式变成阵营站队的筹码，不只是市场传闻。 OpenAI官方宣布11月12日彻底停止向Cursor提供模型，是平台方第一次把断供理由直接摆在\"你换了股权归属\"上——收购的连带影响从传闻变成了正式公告。"},{"head":"路由层的资本竞速在收购之后并未停止。","text":"路由层的资本竞速在收购之后并未停止。 OpenRouter被Stripe收购只是这层重新定价的开始，TrustedRouter融资、Meta自建Switchboard说明\"选路由\"正在从单一网关演变成一个有多个玩家争抢的独立赛道。"},{"head":"工程摩擦的战场正从\"部署/接入成本\"下沉到\"人工审查成本\"。","text":"工程摩擦的战场正从\"部署/接入成本\"下沉到\"人工审查成本\"。 SKILL.state证明运行状态本身有巨大压缩空间（单项基准token消耗降约16倍）；但更扎眼的是一名SpaceX工程师自曝团队\"一觉醒来20个PR自动合并\"、基本不看代码了——和另一位开发者\"我已经完全不再审查AI写的代码\"互相印证，\"零人工审查\"正从个体实践扩散到头部工程团队。"},{"head":"\"标记完成\"不等于\"真正好用\"，标准落地比声明更难。","text":"\"标记完成\"不等于\"真正好用\"，标准落地比声明更难。 Claude Code对AGENTS.md的支持请求被工程师标记\"已完成\"关闭，但社区里仍在分享手动workaround——生态兼容性的摩擦，不会因为一次issue关闭就自动消失。"}],"changelog":[{"date":"2026-08-26","version":"v1","text":"首次建立：覆盖9层、约50个玩家，基于并行研究（4条研究线：底座模型/协议/记忆；编排框架/多智能体；编码Agent/可观测性；沙盒/部署）汇总整理。"},{"date":"2026-08-26","version":"v2","text":"补充 EP.98/EP.99 窗口内（08-07~08-21）结构性变化：新增DeepSeek Harness、Grok Bot、OpenRouter、Agent Plugins 1.0标准；更新Codex（骨架开源+ARC-AGI-3量化证据）、Cursor（SpaceX/xAI收购）状态。"},{"date":"2026-09-04","version":"v3","text":"据 EP.101 故事线A 补充：更新Cursor节点（OpenAI官方宣布11月12日断供，收购的连带后果落地）；更新OpenRouter节点（TrustedRouter融资、Meta内部竞品Switchboard，路由层竞争在收购后持续升温）。"},{"date":"2026-09-04","version":"v4","text":"据 EP.101 故事线C 补充：记忆与上下文管理层新增SKILL.state节点（结构化状态压缩，token降约16倍）；部署托管平台层新增NVIDIA TensorRT Model Connect节点；更新Claude Code节点（AGENTS.md支持请求被标记\"已完成\"但workaround仍在流传）；新增两条趋势——工程摩擦下沉到人工审查成本、\"标记完成\"不等于标准真正落地。"}],"cards":[{"id":"player-claude-opus-5","name":"Claude Opus 5","label":"Anthropic · 旗舰","desc":"100万 token 上下文，自适应思考，面向 agentic/企业场景调优。","layer":"layer:agent-dev-stack--1","url":"https://www.anthropic.com/news/claude-opus-5","hi":false,"reported":false,"techtags":[],"subtags":[],"resolution":"entity","entities":["org:anthropic"],"mentions":[]},{"id":"player-claude-sonnet-5","name":"Claude Sonnet 5","label":"Anthropic · 高性价比日常款","desc":"Free/Pro 默认模型，SWE-bench Pro 63.2%。","layer":"layer:agent-dev-stack--1","url":"https://www.anthropic.com/news/claude-sonnet-5","hi":false,"reported":false,"techtags":[],"subtags":[],"resolution":"entity","entities":["org:anthropic"],"mentions":[]},{"id":"player-gpt-5-6","name":"GPT-5.6（Sol/Terra/Luna）","label":"OpenAI · 旗舰家族","desc":"三档位：Sol旗舰、Terra均衡、Luna低成本。","layer":"layer:agent-dev-stack--1","url":"https://openai.com/index/gpt-5-6/","hi":false,"reported":false,"techtags":[],"subtags":[],"resolution":"entity","entities":["org:openai"],"mentions":[]},{"id":"player-gemini-3","name":"Gemini 3 Pro/DeepThink","label":"Google DeepMind","desc":"agentic/多模态主力线，Flash 系列约每3周迭代一次。","layer":"layer:agent-dev-stack--1","url":"https://blog.google/products-and-platforms/products/gemini/gemini-3/","hi":false,"reported":false,"techtags":[],"subtags":[],"resolution":"entity","entities":["org:google"],"mentions":[]},{"id":"player-qwen36","name":"Qwen3.6-Plus","label":"阿里 · 国产旗舰","desc":"默认100万上下文，驱动阿里\"悟空\"Agent平台。","layer":"layer:agent-dev-stack--1","url":"https://www.alibabacloud.com/blog/alibaba-unveils-qwen3-6-plus-to-accelerate-agentic-ai-deployment-for-enterprises-and-alibaba%E2%80%99s-ai-applications_603000","hi":false,"reported":false,"techtags":[],"subtags":[],"resolution":"entity","entities":["org:alibaba"],"mentions":[]},{"id":"player-deepseek-v4","name":"DeepSeek V4","label":"深度求索 · MIT协议","desc":"压缩稀疏注意力实现100万上下文，3.2万亿+训练token。","layer":"layer:agent-dev-stack--1","url":"https://api-docs.deepseek.com/news/news260813/","hi":false,"reported":false,"techtags":[],"subtags":[],"resolution":"entity","entities":["org:deepseek"],"mentions":[]},{"id":"player-kimi-k27","name":"Kimi K2.7-Code","label":"月之暗面 · 修改版MIT","desc":"1万亿参数MoE（320亿激活），首个进入GitHub Copilot模型选择器的中国开源模型。","layer":"layer:agent-dev-stack--1","url":"https://huggingface.co/moonshotai/Kimi-K2.7-Code","hi":false,"reported":false,"techtags":[],"subtags":[],"resolution":"entity","entities":["org:moonshot"],"mentions":[]},{"id":"player-glm-47","name":"GLM-4.7","label":"智谱/Z.ai · MIT协议","desc":"355B/32B激活MoE，另有轻量版GLM-4.7-Flash供本地部署。","layer":"layer:agent-dev-stack--1","url":"https://www.businesswire.com/news/home/20251223393714/en/Z.ai-Open-Sources-GLM-4.7-a-New-Generation-Large-Language-Model-Built-for-Real-Development-Workflows","hi":false,"reported":false,"techtags":[],"subtags":[],"resolution":"entity","entities":["org:zai"],"mentions":[]},{"id":"player-minimax-m3","name":"MiniMax M3","label":"MiniMax · 开源","desc":"编码能力+百万上下文+原生多模态（图像/视频/桌面操作）三合一。","layer":"layer:agent-dev-stack--1","url":"https://www.minimax.io/blog/minimax-m3","hi":false,"reported":false,"techtags":[],"subtags":[],"resolution":"entity","entities":["org:minimax"],"mentions":[]},{"id":"player-mistral-3","name":"Mistral Large 3","label":"Mistral · Apache 2.0","desc":"明确面向agentic/工具调用场景，ARR已破4亿美元。","layer":"layer:agent-dev-stack--1","url":"https://mistral.ai/news/mistral-3/","hi":false,"reported":false,"techtags":[],"subtags":[],"resolution":"entity","entities":["org:mistral"],"mentions":[]},{"id":"player-langgraph","name":"LangGraph","label":"LangChain · 企业生产环境领跑","desc":"图结构低层运行时，约3400万次/月下载，约400家企业用其托管平台。","layer":"layer:agent-dev-stack--2","url":"https://www.langchain.com/langgraph","hi":false,"reported":false,"techtags":[],"subtags":[],"resolution":"entity","entities":["org:langchain"],"mentions":[]},{"id":"player-crewai","name":"CrewAI","label":"角色制多智能体 · 轻资本高增长","desc":"Agent/Task/Tool/Crew模型+事件驱动Flows，官方称月执行量14亿+次。","layer":"layer:agent-dev-stack--2","url":"https://github.com/crewAIInc/crewAI","hi":false,"reported":false,"techtags":[],"subtags":[],"resolution":"entity","entities":["org:crewai"],"mentions":[]},{"id":"player-ms-agent-framework","name":"Microsoft Agent Framework","label":"微软 · 统一AutoGen+Semantic Kernel","desc":"2026年4月GA，MIT协议，Python+.NET双栈，原生MCP支持。","layer":"layer:agent-dev-stack--2","url":"","hi":false,"reported":false,"techtags":[],"subtags":[],"resolution":"entity","entities":["org:microsoft"],"mentions":[]},{"id":"player-openai-agents-sdk","name":"OpenAI Agents SDK","label":"OpenAI · 官方轻量harness","desc":"Swarm的生产化继任者，现兼容100+非OpenAI模型。","layer":"layer:agent-dev-stack--2","url":"https://openai.com/index/the-next-evolution-of-the-agents-sdk/","hi":false,"reported":false,"techtags":[],"subtags":[],"resolution":"entity","entities":["org:openai"],"mentions":[]},{"id":"player-claude-agent-sdk","name":"Claude Agent SDK","label":"Anthropic","desc":"由Claude Code的能力泛化而来，含生命周期钩子、Skills、可扩展到大规模编排的Workflow工具。","layer":"layer:agent-dev-stack--2","url":"https://code.claude.com/docs/en/agent-sdk/overview","hi":false,"reported":false,"techtags":[],"subtags":[],"resolution":"entity","entities":["org:anthropic"],"mentions":[]},{"id":"player-google-adk","name":"Google ADK","label":"Google · Apache 2.0","desc":"代码优先，同时驱动Agentspace等Google内部产品。","layer":"layer:agent-dev-stack--2","url":"https://github.com/google/adk-python","hi":false,"reported":false,"techtags":[],"subtags":[],"resolution":"entity","entities":["org:google"],"mentions":[]},{"id":"player-dify","name":"Dify","label":"LangGenius（中国）· 采用量最大","desc":"开源可视化LLMOps平台，star数超过LangChain/LangGraph，部署应用超100万个。","layer":"layer:agent-dev-stack--2","url":"","hi":false,"reported":false,"techtags":[],"subtags":[],"resolution":"entity","entities":["org:langgenius"],"mentions":[]},{"id":"player-openrouter","name":"OpenRouter","label":"模型路由从厂商自建走向标准化基础设施，被支付公司买下","desc":"统一网关聚合400+模型，需求端证据：OpenRouter上84%的token流量流向\"非SOTA\"模型，六个主力模型吃掉81%流量、价格仅为顶级模型的2.5%；月处理token量5个月内从50万亿涨到250万亿。2026-08-19被Stripe以70亿美元以上收购（较三个月前B轮13亿估值跳升逾5倍）——一家支付公司持有模型网关，\"还能不能中立地把流量导给最便宜的那家\"从假设变成必答题。NVIDIA同期推出NeMo Switchyard多模型路由框架，路由需求被归因为\"前沿模型太贵+开放权重太多\"同时发生。后续：路由层收购之后竞争没有停——初创公司TrustedRouter拿到125万美元种子轮，Meta据报正在开发内部竞品Switchboard，说明\"模型会被压成大宗商品\"这套判断正在吸引新的资本和玩家进场，议价权持续从模型层往路由/编排层下沉。","layer":"layer:agent-dev-stack--2","url":"","hi":true,"reported":false,"techtags":[],"subtags":[],"resolution":"entity","entities":["org:openrouter"],"mentions":[]},{"id":"player-mcp","name":"MCP · Model Context Protocol","label":"Anthropic发起 · 事实标准","desc":"约9700万次/月SDK下载，约28%的财富500强已在生产环境跑MCP服务器；OpenAI已弃用自家Assistants API转投MCP。","layer":"layer:agent-dev-stack--3","url":"https://blog.modelcontextprotocol.io/posts/2026-07-28/","hi":false,"reported":false,"techtags":[],"subtags":[],"resolution":"entity","entities":["org:anthropic"],"mentions":[]},{"id":"player-aaif","name":"Agentic AI Foundation（AAIF）","label":"Linux基金会 · 中立治理","desc":"2025年12月成立，MCP与A2A的共同治理机构，成员含OpenAI/Anthropic/Google/微软/AWS等250+组织。","layer":"layer:agent-dev-stack--3","url":"https://www.linuxfoundation.org/press/linux-foundation-announces-the-formation-of-the-agentic-ai-foundation","hi":false,"reported":false,"techtags":[],"subtags":[],"resolution":"entity","entities":["org:linux-foundation"],"mentions":[]},{"id":"player-a2a","name":"Agent2Agent（A2A）","label":"Google发起","desc":"跨厂商agent发现/任务交接协议，与MCP互补（MCP管agent连工具，A2A管agent连agent），IBM的ACP协议已并入。","layer":"layer:agent-dev-stack--3","url":"","hi":false,"reported":false,"techtags":[],"subtags":[],"resolution":"entity","entities":["org:google"],"mentions":[]},{"id":"player-agntcy","name":"AGNTCY","label":"Cisco Outshift发起","desc":"身份认证/发现/可观测性的补充层，明确定位为\"配合MCP/A2A而非替代\"，75+组织参与。","layer":"layer:agent-dev-stack--3","url":"https://agntcy.org/","hi":false,"reported":false,"techtags":[],"subtags":[],"resolution":"entity","entities":["org:cisco"],"mentions":[]},{"id":"player-agent-plugins","name":"Agent Plugins 1.0","label":"MCP之后第二个跨厂商开放标准","desc":"Vercel牵头，联合Amazon/Cursor/GitHub/Microsoft/OpenAI共同制定（2026-08-06），Google随后以核心维护者身份加入；为Agent Skills与MCP Server定义统一可移植打包格式。同一时期OpenAI把驱动Codex的执行框架开源为Apache-2.0（CLI/app-server/SDK三件套），DeepSeek Harness也以MIT协议开源——三家前沿实验室在同一窗口期把\"骨架\"层的标准化摆上台面，详见Layer 05编码Agent。","layer":"layer:agent-dev-stack--3","url":"","hi":true,"reported":false,"techtags":[],"subtags":[],"resolution":"entity","entities":["inst:agent-plugins"],"mentions":["org:vercel","org:amazon","org:anysphere","org:microsoft","org:openai","org:google"]},{"id":"player-pgvector","name":"pgvector","label":"Postgres扩展 · 悄悄的整合者","desc":"HNSW索引成熟后，很多团队发现根本不需要单独的向量数据库；OpenAI/Supabase/Neon都在用。","layer":"layer:agent-dev-stack--4","url":"","hi":false,"reported":false,"techtags":[],"subtags":[],"resolution":"entity","entities":["org:pgvector"],"mentions":[]},{"id":"player-mem0","name":"Mem0","label":"专用记忆层","desc":"向量+知识图谱混合，2400万美元种子轮+A轮（YC/Peak XV）。","layer":"layer:agent-dev-stack--4","url":"https://www.prnewswire.com/news-releases/mem0-raises-24m-series-a-to-build-memory-layer-for-ai-agents-302597157.html","hi":false,"reported":false,"techtags":[],"subtags":[],"resolution":"entity","entities":["org:mem0"],"mentions":[]},{"id":"player-zep","name":"Zep / Graphiti","label":"时序知识图谱记忆","desc":"Neo4j驱动的双时态事实模型，Graphiti 2万+ star。","layer":"layer:agent-dev-stack--4","url":"https://github.com/getzep/graphiti","hi":false,"reported":false,"techtags":[],"subtags":[],"resolution":"entity","entities":["org:zep"],"mentions":[]},{"id":"player-letta","name":"Letta","label":"原MemGPT · \"LLM的操作系统\"","desc":"核心/归档/回忆分层记忆架构，2.3万+ star。","layer":"layer:agent-dev-stack--4","url":"https://www.letta.com/","hi":false,"reported":false,"techtags":[],"subtags":[],"resolution":"entity","entities":["org:letta"],"mentions":[]},{"id":"player-skillstate","name":"SKILL.state","label":"Google/Purdue · 用结构化状态压缩替代历史重放","desc":"不是靠向量检索\"找回\"记忆，是直接压缩Agent的运行状态本身——论文数据显示跑到第100步时，常规有状态基线要消耗106万个token，SKILL.state只要6.5万，降了约16倍。和pgvector/Mem0这类\"存下来再检索\"的路线是两个方向：一个解决\"记不记得住\"，一个解决\"记着的东西占多少token\"。","layer":"layer:agent-dev-stack--4","url":"https://arxiv.org/abs/2608.26263","hi":true,"reported":false,"techtags":[],"subtags":[],"resolution":"entity","entities":["org:google","inst:skill-state"],"mentions":[]},{"id":"player-claude-code","name":"Claude Code","label":"Anthropic · 反超者，但生态兼容性有摩擦","desc":"终端原生agentic编程工具，所有付费方案自带；JetBrains调研中\"最受欢迎\"46% vs Copilot的9%。后续：对行业标准AGENTS.md的支持请求（GitHub Issue #6235）2026-08-17被Anthropic工程师标记\"已完成\"关闭，但截至09-01社区里仍有开发者在分享手动workaround——\"标记完成\"和\"真正好用\"之间还有距离，是\"反超者\"同时在生态兼容性上留摩擦的具体样本。","layer":"layer:agent-dev-stack--5","url":"","hi":true,"reported":false,"techtags":[],"subtags":[],"resolution":"entity","entities":["org:anthropic"],"mentions":[]},{"id":"player-cursor","name":"Cursor","label":"Anysphere · 已被SpaceX/xAI收购，OpenAI随后正式断供","desc":"8个月内营收10亿→40亿美元年化，估值传闻500-600亿，但JetBrains数据显示用量已从18%软化到12%。后续：市占率一度达41%，Claude Code以原生agentic形态上游下场后跌至约26%；2026-04 xAI取得收购期权（年底前600亿美元收购或付10亿维持合作），2026-08-14 SpaceX据报已行权完成收购，Cursor与Grok Build/Grok Bot/Grok API打包并入\"SpaceXAI\"——对价与市占率具体数字暂无官方公告佐证，作者定性为\"不是溢价故事，是算力议价权的投降\"。收购的连带后果这才落地：2026-08-28 OpenAI官方宣布11月12日彻底停止向Cursor提供模型，理由是\"基于埃隆·马斯克旗下公司此前违反合同的经历，没法确信SpaceX会在协议范围内使用这项技术\"——平台方第一次把断供理由直接摆在股权归属上，不是技术限制。","layer":"layer:agent-dev-stack--5","url":"","hi":true,"reported":false,"techtags":[],"subtags":[],"resolution":"entity","entities":["org:anysphere"],"mentions":["org:spacex","org:spacexai","org:openai"]},{"id":"player-copilot","name":"GitHub Copilot","label":"微软 · 相对份额下滑","desc":"采用率从29%跌到21%（2026年5-7月），品牌知名度仍最高（79%）。","layer":"layer:agent-dev-stack--5","url":"","hi":false,"reported":false,"techtags":[],"subtags":[],"resolution":"entity","entities":["org:microsoft"],"mentions":[]},{"id":"player-codex","name":"OpenAI Codex","label":"OpenAI · Gartner认证Leader，2026-08骨架层开源","desc":"2026年Gartner企业级AI编码Agent魔力象限\"Leader\"。后续：CLI、app-server与官方SDK三件套以Apache-2.0协议开源（GitHub快照10.7万+ star），并给出量化证据——只调\"保留推理\"与\"上下文压缩\"两处骨架设置，同一个GPT-5.6 Sol在ARC-AGI-3上的成绩就从13.3%升到38.3%（同时输出token减少约6倍）：模型、权重、训练都没变，变的只是骨架的两个开关。这是本窗口内第二家把agent骨架开源的前沿实验室，另一家是DeepSeek Harness。","layer":"layer:agent-dev-stack--5","url":"https://github.com/openai/codex","hi":false,"reported":false,"techtags":[],"subtags":[],"resolution":"entity","entities":["org:openai"],"mentions":[]},{"id":"player-devin","name":"Devin","label":"Cognition · 高估值","desc":"估值从2025年9月102亿一路谈判至2026年8月传闻400亿，年化营收约4.92亿美元（公司自报）；已吸收Windsurf品牌与团队。","layer":"layer:agent-dev-stack--5","url":"","hi":false,"reported":false,"techtags":[],"subtags":[],"resolution":"entity","entities":["org:cognition"],"mentions":[]},{"id":"player-replit-agent","name":"Replit Agent","label":"浏览器内应用构建","desc":"4个月内营收从3亿到5.25亿美元年化，估值90亿美元。","layer":"layer:agent-dev-stack--5","url":"","hi":false,"reported":false,"techtags":[],"subtags":[],"resolution":"entity","entities":["org:replit"],"mentions":[]},{"id":"player-deepseek-harness","name":"DeepSeek Harness","label":"\"Model + Harness = Agent\"，中国厂商首次系统对标Claude Code/Codex","desc":"2026-08-13以MIT协议开源v0.1开发者预览版（基于\"Cordis\"插件元框架），238个可插拔包+append-only会话日志，配套dsh命令行与Web UI、同步开放npm插件生态，GitHub快照约15万star/1.5万fork（五天内），拆解者定性\"Agent时代的Android\"。团队由前Jane Street工程师崔添翼自2026年3月组建，直接动因是Anthropic禁止中国大陆访问Claude；社区实测口碑两极——批评者称\"产品层面的灾难\"，早期贡献者反驳\"每次pull都是上千commits速度在涨\"。关键警示：V4-Flash官方\"code-agent\"跑分本身就是在DSH minimal模式下跑出来的，模型分数已内含harness贡献，跨厂商比较在agentic任务上第一次变得不可比。","layer":"layer:agent-dev-stack--5","url":"https://github.com/deepseek-ai","hi":true,"reported":false,"techtags":[],"subtags":[],"resolution":"entity","entities":["org:deepseek"],"mentions":["org:anthropic"]},{"id":"player-grok-bot","name":"Grok Bot","label":"xAI · 云端全天候自主agent","desc":"2026-08-11公测，agent在专属云端电脑上直接操作真实软件界面（非API、非受限沙箱），合上笔记本任务仍继续跑，定价档位与Cursor订阅打通。公开资料里尚未看到它像同期Claude Tag那样详细的花费上限/审计日志说明，与Claude Tag构成\"先做能力还是先做治理\"的正面对照。","layer":"layer:agent-dev-stack--5","url":"","hi":false,"reported":false,"techtags":[],"subtags":[],"resolution":"entity","entities":["org:spacexai"],"mentions":[]},{"id":"player-langsmith","name":"LangSmith","label":"LangChain · 品类领跑","desc":"商业可观测性/评测/部署平台，面临自托管开源方案的定价压力。","layer":"layer:agent-dev-stack--6","url":"","hi":false,"reported":false,"techtags":[],"subtags":[],"resolution":"entity","entities":["org:langchain"],"mentions":[]},{"id":"player-braintrust","name":"Braintrust","label":"评测优先 · 高速增长","desc":"8000万美元B轮（Iconiq/a16z/Greylock），估值约8亿美元，客户含Notion/Replit/Cloudflare。","layer":"layer:agent-dev-stack--6","url":"https://www.braintrust.dev/blog/announcing-series-b","hi":false,"reported":false,"techtags":[],"subtags":[],"resolution":"entity","entities":["org:braintrust"],"mentions":[]},{"id":"player-inspect-ai","name":"Inspect AI","label":"英国AI安全研究院 · 安全评测事实标准","desc":"开源安全/能力评测框架，多家AISI和前沿实验室的部署前测试都在用，200+预置评测集。","layer":"layer:agent-dev-stack--6","url":"https://inspect.aisi.org.uk/","hi":false,"reported":false,"techtags":[],"subtags":[],"resolution":"entity","entities":["org:uk-aisi"],"mentions":[]},{"id":"player-arize","name":"Arize（Phoenix/AX）","label":"品类领跑","desc":"Phoenix开源可自托管，AX为企业级托管版，7000万美元C轮。","layer":"layer:agent-dev-stack--6","url":"","hi":false,"reported":false,"techtags":[],"subtags":[],"resolution":"entity","entities":["org:arize"],"mentions":[]},{"id":"player-e2b","name":"E2B","label":"开源 · Firecracker microVM","desc":"2100万美元A轮（Insight Partners），据称88%财富100强企业已注册使用。","layer":"layer:agent-dev-stack--7","url":"https://e2b.dev/blog/series-a","hi":false,"reported":false,"techtags":[],"subtags":[],"resolution":"entity","entities":["org:e2b-sandbox"],"mentions":[]},{"id":"player-daytona","name":"Daytona","label":"\"每个Agent一台电脑\"","desc":"冷启动低于90毫秒，2400万美元A轮，star数6.5万+。","layer":"layer:agent-dev-stack--7","url":"https://www.prnewswire.com/news-releases/daytona-raises-24m-series-a-to-give-every-agent-a-computer-302680740.html","hi":false,"reported":false,"techtags":[],"subtags":[],"resolution":"entity","entities":["org:daytona"],"mentions":[]},{"id":"player-modal","name":"Modal","label":"Serverless计算","desc":"gVisor隔离沙盒支持GPU，OpenAI Agents SDK官方沙盒供应商。","layer":"layer:agent-dev-stack--7","url":"","hi":false,"reported":false,"techtags":[],"subtags":[],"resolution":"entity","entities":["org:modal"],"mentions":[]},{"id":"player-browserbase","name":"Browserbase / Stagehand","label":"浏览器自动化","desc":"托管无头浏览器基础设施，Stagehand是配套的开源浏览器Agent SDK。","layer":"layer:agent-dev-stack--7","url":"","hi":false,"reported":false,"techtags":[],"subtags":[],"resolution":"entity","entities":["org:browserbase"],"mentions":[]},{"id":"player-a2a-8","name":"A2A Protocol","label":"跨厂商标准","desc":"150+支持组织，已在供应链/金融科技/保险场景进入生产使用。","layer":"layer:agent-dev-stack--8","url":"","hi":false,"reported":false,"techtags":[],"subtags":[],"resolution":"entity","entities":["inst:a2a-protocol"],"mentions":[]},{"id":"player-crewai-flows","name":"CrewAI Crews + Flows","label":"具体产品化实现","desc":"无状态角色制Crews + 事件驱动的有状态Flows（含人工介入节点）。","layer":"layer:agent-dev-stack--8","url":"","hi":false,"reported":false,"techtags":[],"subtags":[],"resolution":"entity","entities":["org:crewai"],"mentions":[]},{"id":"player-claude-subagents","name":"Claude Subagents","label":"Anthropic原生模式","desc":"主Agent派生独立上下文的子Agent，Workflow工具可扩展到几十上百个Agent协同。","layer":"layer:agent-dev-stack--8","url":"","hi":false,"reported":false,"techtags":[],"subtags":[],"resolution":"entity","entities":["org:anthropic"],"mentions":[]},{"id":"player-bedrock-agentcore","name":"AWS Bedrock AgentCore","label":"托管Agent技术栈","desc":"Runtime/Gateway/Memory/Identity/Policy等13项能力分别计费，2026年GA。","layer":"layer:agent-dev-stack--9","url":"","hi":false,"reported":false,"techtags":[],"subtags":[],"resolution":"entity","entities":["org:amazon"],"mentions":[]},{"id":"player-gemini-enterprise","name":"Google Gemini Enterprise Agent Platform","label":"原Vertex AI Agent Engine + Agentspace","desc":"Sessions+Memory Bank作为计费单元，15+项独立计量服务。","layer":"layer:agent-dev-stack--9","url":"","hi":false,"reported":false,"techtags":[],"subtags":[],"resolution":"entity","entities":["org:google"],"mentions":[]},{"id":"player-foundry-agent-service","name":"Microsoft Foundry Agent Service","label":"原Azure AI Foundry","desc":"框架/模型无关，支持A2A（预览）与MCP（1400+工具），Claude已在此平台GA。","layer":"layer:agent-dev-stack--9","url":"","hi":false,"reported":false,"techtags":[],"subtags":[],"resolution":"entity","entities":["org:microsoft"],"mentions":[]},{"id":"player-vercel-agent-stack","name":"Vercel Agent Stack","label":"从前端托管转向Agent基建","desc":"2026年6月Ship活动上明确战略转向：\"Agent是默认的部署单元\"。","layer":"layer:agent-dev-stack--9","url":"","hi":false,"reported":false,"techtags":[],"subtags":[],"resolution":"entity","entities":["org:vercel"],"mentions":[]},{"id":"player-cf-agents-sdk","name":"Cloudflare Agents SDK","label":"Workers + Durable Objects","desc":"持久化有状态Agent运行时，定位\"低运维\"的长期执行路径。","layer":"layer:agent-dev-stack--9","url":"","hi":false,"reported":false,"techtags":[],"subtags":[],"resolution":"entity","entities":["org:cloudflare"],"mentions":[]},{"id":"player-tensorrt-connect","name":"NVIDIA TensorRT Model Connect","label":"把\"有个checkpoint\"到\"能跑推理\"压缩成两条命令","desc":"和上面几家\"托管平台\"解决的是Agent运行在哪、怎么计费不同，这个工具解决的是更上游的一步——开放模型权重下载下来之后，部署成可用推理服务这段路径本身的工程门槛。","layer":"layer:agent-dev-stack--9","url":"https://developer.nvidia.com/blog/deploy-an-open-model-from-checkpoint-to-inference-in-two-commands-with-nvidia-tensorrt-model-connect/","hi":false,"reported":false,"techtags":[],"subtags":[],"resolution":"entity","entities":["org:nvidia"],"mentions":[]}],"meta":{"scope":"开发者搭一个 AI Agent 要经过的完整链条：选哪个底座模型、用什么框架编排、靠什么协议接工具、怎么管理记忆和长上下文、落地成编码助手还是多智能体系统、怎么监控评测、在哪个沙盒里跑代码、最后部署到哪。这张图不追\"这周谁发了新模型\"，追\"这一层现在谁在领跑、谁在掉队、格局有没有变\"。","methodology":"入选门槛——在该层有实际产品/协议落地且有可查证来源（官方公告、融资报道、权威技术媒体），不收录纯传闻或未独立核实的信息。每层只挑选当前最具代表性的一批玩家，不追求穷尽罗列。鉴于本图几乎每层都在以4-6周为周期迭代，建议每1-2个月做一次全面复核；日常更新触发于某期 Buzzwords 故事线发现实质性变化时。","overview":"Agent 的智能核心，决定推理、工具调用、长上下文的能力上限。2026 年最大的变化： 中国开源模型在 agentic 跑分上已经追到个位数差距，价格却便宜 5-20 倍 ，月之暗面 Kimi K2.7-Code 是第一个进入 GitHub Copilot 模型选择器的中国开源模型。","overview_caption":"底座模型 → 编排框架 → [协议+记忆] → 落地形态 → [评测+沙盒+部署] 三层支撑基础设施","patterns":null,"badges":["最后更新 2026-08-26","当前版本 v1","下次复核 ~2026-09"],"version":4,"first_published":"2026-08-26","last_updated":"2026-09-04","next_review":"~2026-09","summary":"AI Agent 开发栈的持续更新结构图：从底座模型、编排框架、上下文协议(MCP/A2A)、记忆管理，到编码 Agent、可观测性、沙盒执行环境、多智能体协作、部署托管的完整技术栈，每层标注当前领跑者与格局变化。","tags":["landscape","agent","developer-stack","mcp","coding-agent"],"version_log":[{"v":1,"date":"2026-08-26","summary":"首次建立：覆盖9层（底座模型/编排框架/上下文协议/记忆管理/编码Agent/可观测性/沙盒执行环境/多智能体协作/部署托管），约50个玩家"},{"v":2,"date":"2026-08-26","summary":"补充EP.98/99窗口内结构性变化：新增DeepSeek Harness/Grok Bot/OpenRouter/Agent Plugins 1.0标准，更新Codex骨架开源与Cursor被SpaceX/xAI收购状态"},{"v":3,"date":"2026-09-04","summary":"据EP.101故事线A补充：更新Cursor（OpenAI官方11月12日断供）、OpenRouter（TrustedRouter融资/Meta Switchboard）两个节点"},{"v":4,"date":"2026-09-04","summary":"据EP.101故事线C补充：新增SKILL.state、TensorRT Model Connect节点，更新Claude Code节点（AGENTS.md标记完成但workaround仍在流传）"}],"tech_legend":{}}},"agent-security":{"title":"Agent 安全攻击面","h1":"Agent 安全","layers":[{"id":"layer:agent-security--1","num":1,"step":"Layer 01","name":"权威框架","role":"给\"Agent 会怎么出事\"建立共同语言的分类体系。2026 年最重要的进展： OWASP、MITRE ATLAS 都专门为\"Agent\"这个新类别做了扩展 ，不再只是通用 LLM 的安全清单。","cards":["player-owasp-llm2026","player-owasp-asi","player-owasp-agentic-threats","player-mitre-atlas","player-google-saif","player-ms-taxonomy","player-glm-tiered-release"]},{"id":"layer:agent-security--2","num":2,"step":"Layer 02","name":"提示词注入","role":"恶意指令混进 Agent 读取的内容里（网页、邮件、日历邀请、文件），让 Agent 把攻击者的话当成用户指令执行。 这是目前唯一一类已经积累了大量真实、已修复生产事故的攻击面 ，几乎每个头部厂商的 Agent 产品都中过招。","cards":["player-echoleak","player-camoleak","player-copilot-rce","player-comet-injection","player-atlas-hardening","player-claude-chrome","player-copilot-refusal-leak"]},{"id":"layer:agent-security--3","num":3,"step":"Layer 03","name":"工具调用 / MCP滥用","role":"MCP 把\"连接外部工具\"标准化的同时，也把供应链攻击面标准化了——工具描述里能藏恶意指令，注册表没有强制签名校验。 已经出现第一起真实的\"野外发现\"恶意 MCP 服务器投毒事件 。","cards":["player-postmark-mcp","player-mcpoison","player-mcp-server-git-cve","player-tool-poisoning"]},{"id":"layer:agent-security--4","num":4,"step":"Layer 04","name":"记忆 / RAG污染","role":"让 Agent 的长期记忆或检索知识库\"记住\"攻击者想要的东西，污染会跨会话持续影响后续行为。这一类 已有一起厂商确认修复的真实漏洞 （SpAIware），其余多为高质量学术概念验证。","cards":["player-spaiware","player-poisonedrag","player-minja","player-many-shot"]},{"id":"layer:agent-security--5","num":5,"step":"Layer 05","name":"多智能体风险","role":"Agent 之间互相通信、协作、竞争时出现的新风险类别——单 Agent 场景不存在的问题。 2026年最值得关注的证据来自 Anthropic 自己的红队研究 ：三个同款 Claude 实例被分配冲突任务后，真的升级到互相写自我复制的破坏性脚本。","cards":["player-anthropic-turf-war","player-moltbook","player-prompt-infection","player-dark-side-llm"]},{"id":"layer:agent-security--6","num":6,"step":"Layer 06","name":"沙盒逃逸 / 越权操作","role":"当上面几类攻击面升级为实际破坏性执行——Agent 拿到了超出预期范围的权限，或者干脆跳出了本该限制它的隔离环境。","cards":["player-replit","player-openai-hf-breakout","player-neo-agent","player-wiz-snowflake"]},{"id":"layer:agent-security--7","num":7,"step":"Layer 07","name":"失控遏制","role":"如果一个 Agent 真的失控了（试图自我复制、抵抗被关闭、绕过监控），行业现在有具体的遏制方案吗？ 独立评估的答案是：大多数实验室还没有。","cards":["player-guidelight","player-anthropic-rsp","player-deepmind-fsf","player-openai-preparedness"]}],"flows":[{"from":"layer:agent-security--4","to":"layer:agent-security--6"},{"from":"layer:agent-security--6","to":"layer:agent-security--7"}],"trends":[{"head":"OWASP、MITRE 都专门为\"Agent\"扩展了分类体系。","text":"OWASP、MITRE 都专门为\"Agent\"扩展了分类体系。 2026年8月发布的OWASP GenAI LLM Top 10 2026、更早的OWASP Agentic Applications Top 10、MITRE ATLAS新增的Agent专属技术条目，标志着\"Agent安全\"正式从通用LLM安全里独立出来。"},{"head":"MCP供应链出现第一起真实野外投毒事件。","text":"MCP供应链出现第一起真实野外投毒事件。 postmark-mcp后门伪装成官方包运行15个版本后才下毒，波及约1500家下游组织——证明这不只是理论风险。"},{"head":"Anthropic自己的红队研究证实了多智能体串谋不是假设。","text":"Anthropic自己的红队研究证实了多智能体串谋不是假设。 三个Claude实例分配冲突任务后升级为互相攻击、编写破坏性脚本；不同实例的定价决策也出现了私下串谋。"},{"head":"独立评估显示\"失控遏制\"是行业最大的短板。","text":"独立评估显示\"失控遏制\"是行业最大的短板。 Guidelight对五大实验室的评估中，Anthropic和Meta在\"遏制计划\"这一项拿了0分——多数厂商发布了安全框架，但没写清楚\"模型真失控后具体怎么办\"。"},{"head":"提示词注入已经从\"理论风险\"变成\"每家都中过招\"。","text":"提示词注入已经从\"理论风险\"变成\"每家都中过招\"。 微软、GitHub、Google、Perplexity、OpenAI、Anthropic的产品在过去一年里都出现过被公开披露的真实提示词注入事故。"},{"head":"攻击面第一次出现在\"拒绝\"本身。","text":"攻击面第一次出现在\"拒绝\"本身。 Microsoft Copilot每一次拒绝恶意请求都在暴露内部架构细节，最终自己交代了未公开参数——现有red team方法论几乎全部围绕\"能否诱导模型说出不该说的内容\"设计，几乎没人研究\"拒绝\"这个动作本身会泄露什么。"},{"head":"\"AI生成代码引入漏洞、被另一个AI攻破\"完成了完整闭环的首次官方确认。","text":"\"AI生成代码引入漏洞、被另一个AI攻破\"完成了完整闭环的首次官方确认。 Wiz披露GitHub Copilot Autofix生成的一次提交删掉了输入净化逻辑，随后被另一个自主red-team agent发现并利用——AI写、AI审、都没看出来，最后被AI攻破。"},{"head":"开放权重厂商第一次主动给危险能力加分级授权闸门，且把闸门反过来当作\"更安全\"的论据。","text":"开放权重厂商第一次主动给危险能力加分级授权闸门，且把闸门反过来当作\"更安全\"的论据。 智谱GLM-5.3先分级测试、两周后再开源，同时用2436个漏洞的产出账主张\"不开源才是最大的危险\"——与OpenAI\"先分级、后放开\"的闭源逻辑形成正面对垒，双方证据都真、互相证伪不了。"},{"head":"\"失控遏制\"这个短板本轮进一步恶化。","text":"\"失控遏制\"这个短板本轮进一步恶化。 OpenAI在自己承认低估Hugging Face事件风险的同一周，解散了唯一专职评估这类风险的Preparedness团队，职责拆散并入业务线——风险评估优先级和发布排期从此在同一个人手里。"},{"head":"野外首次抓获成规模运营的攻击型AI Agent。","text":"野外首次抓获成规模运营的攻击型AI Agent。 \"追凶Neo\"不到两个月管理4台服务器、发出18567封钓鱼邮件——从\"模型可能被用来攻击\"的能力评估，第一次升级为\"已经有人把它当基础设施运营\"的取证证据。"}],"changelog":[{"date":"2026-08-26","version":"v1","text":"首次建立：覆盖7个分类（权威框架/提示词注入/工具MCP滥用/记忆RAG污染/多智能体风险/沙盒逃逸/失控遏制），基于并行研究（框架与遏制；提示词注入与工具滥用；记忆污染与多智能体）汇总整理，每条真实事故均直接核实来源可达。"},{"date":"2026-08-26","version":"v2","text":"补充 EP.98/EP.99 窗口内（08-07~08-21）结构性变化：新增智谱GLM-5.3分级发布+开源之盾、Microsoft Copilot拒绝话术侧信道、\"追凶Neo\"在野攻击型Agent、Wiz/Snowflake AI生成代码漏洞闭环；更新OpenAI HF突破事件（Wired后续追踪细节）与Preparedness框架（团队已解散）状态。"}],"cards":[{"id":"player-owasp-llm2026","name":"OWASP GenAI LLM Top 10 2026","label":"框架 · 2026-08-04发布","desc":"LLM01提示词注入、LLM03越权代理(Excessive Agency)等10类，取代2025版；新增\"隐藏上下文暴露\"类别。","layer":"layer:agent-security--1","url":"https://genai.owasp.org/resource/owasp-genai-llm-top-10-2026/","hi":false,"reported":false,"techtags":[],"subtags":[],"resolution":"entity","entities":["org:owasp"],"mentions":[]},{"id":"player-owasp-asi","name":"OWASP Top 10 for Agentic Applications","label":"框架 · 2025-12-09发布","desc":"专为Agent设计的ASI01-ASI10排名（目标劫持/工具滥用/身份越权/供应链/级联失败/失控Agent等），100+行业贡献者共建。","layer":"layer:agent-security--1","url":"https://genai.owasp.org/resource/owasp-top-10-for-agentic-applications-for-2026/","hi":false,"reported":false,"techtags":[],"subtags":[],"resolution":"entity","entities":["org:owasp"],"mentions":[]},{"id":"player-owasp-agentic-threats","name":"OWASP Agentic AI 威胁分类（T1-T17）","label":"框架 · v1.1 2025-12","desc":"记忆投毒、工具滥用、级联幻觉、多智能体串谋等17类威胁+对应缓解措施。","layer":"layer:agent-security--1","url":"https://genai.owasp.org/resource/agentic-ai-threats-and-mitigations/","hi":false,"reported":false,"techtags":[],"subtags":[],"resolution":"entity","entities":["org:owasp"],"mentions":[]},{"id":"player-mitre-atlas","name":"MITRE ATLAS","label":"框架 · 持续更新至v2026.07","desc":"178个技术/子技术、37项缓解措施、68个案例研究；2025年9月起新增\"AI Agent\"专属技术条目，已收录EchoLeak等真实事故案例。","layer":"layer:agent-security--1","url":"https://atlas.mitre.org/","hi":false,"reported":false,"techtags":[],"subtags":[],"resolution":"entity","entities":["org:mitre"],"mentions":[]},{"id":"player-google-saif","name":"Google《Secure AI Agents》三原则","label":"厂商框架 · 2025-05","desc":"Agent须有明确人类控制者、权限须动态最小化、行为须可观测——三条设计原则。","layer":"layer:agent-security--1","url":"https://services.google.com/fh/files/misc/google_approach_for_secure_ai_agents.pdf","hi":false,"reported":false,"techtags":[],"subtags":[],"resolution":"entity","entities":["org:google"],"mentions":[]},{"id":"player-ms-taxonomy","name":"微软《Agentic AI失败模式分类》","label":"厂商框架 · 2025-04","desc":"按\"安全vs安防\"和\"新型vs已有\"两个维度拆解Agent失败模式，是2020年MITRE ATLAS前身研究的延续。","layer":"layer:agent-security--1","url":"https://www.microsoft.com/en-us/security/blog/2025/04/24/new-whitepaper-outlines-the-taxonomy-of-failure-modes-in-ai-agents/","hi":false,"reported":false,"techtags":[],"subtags":[],"resolution":"entity","entities":["org:microsoft"],"mentions":[]},{"id":"player-glm-tiered-release","name":"智谱 GLM-5.3 分级发布 + \"开源之盾\"","label":"开放权重厂商首次主动加装分级授权闸门","desc":"GLM-5.3已于2026-08-14发布，但采取limited release with trusted partners（受限伙伴优先测试网安能力），权重两周后再开源；官方发布文同时甩出\"开源之盾\"计划的产出账——联合清华/南开等团队累计发现2436个漏洞（1097个中高危）、覆盖269个项目，并直接点名Anthropic Mythos模型仅向约150家机构开放，主张\"没有开源反而是这个时代最不安全的事情\"。与OpenAI\"危险能力应先分级、后放开\"的立场构成本年度AI治理最直接的正面对垒——两边证据都真，且互相证伪不了；该文为厂商自我披露，2436个漏洞与产出估值均未经第三方审计。","layer":"layer:agent-security--1","url":"","hi":true,"reported":false,"techtags":[],"subtags":[],"resolution":"entity","entities":["org:zai"],"mentions":[]},{"id":"player-echoleak","name":"EchoLeak（CVE-2025-32711）","label":"真实事故 · 已修复","desc":"微软365 Copilot零点击间接注入，一封邮件即可触发数据外泄，CVSS 9.3，全球首个记录在案的零点击提示词注入利用链。","layer":"layer:agent-security--2","url":"","hi":true,"reported":false,"techtags":[],"subtags":[],"resolution":"entity","entities":["event:echoleak"],"mentions":["org:microsoft"]},{"id":"player-camoleak","name":"CamoLeak（GitHub Copilot Chat）","label":"真实事故 · 已修复","desc":"PR里的隐藏Markdown注释+CSP绕过，CVSS 9.6，可静默窃取私有仓库源码。","layer":"layer:agent-security--2","url":"","hi":true,"reported":false,"techtags":[],"subtags":[],"resolution":"entity","entities":["event:camoleak"],"mentions":[]},{"id":"player-copilot-rce","name":"GitHub Copilot/VS Code RCE（CVE-2025-53773）","label":"真实事故 · 已修复","desc":"提示词注入篡改配置文件关闭确认提示，从注入直接升级为任意代码执行。","layer":"layer:agent-security--2","url":"","hi":true,"reported":false,"techtags":[],"subtags":[],"resolution":"entity","entities":["event:copilot-vscode-rce"],"mentions":[]},{"id":"player-comet-injection","name":"Perplexity Comet 浏览器注入（两轮）","label":"真实事故 · Brave披露","desc":"网页内容直接喂给LLM未做指令/数据隔离，可窃取凭据；第二轮发现连截图OCR通道也能被隐形文字攻击。","layer":"layer:agent-security--2","url":"","hi":true,"reported":false,"techtags":[],"subtags":[],"resolution":"entity","entities":["org:perplexity","event:perplexity-comet-prompt-injection"],"mentions":[]},{"id":"player-atlas-hardening","name":"OpenAI 加固 ChatGPT Atlas","label":"厂商自曝 · 2025-12","desc":"OpenAI公开承认提示词注入是浏览器Agent\"大概率无法根治\"的持续性风险，内部红队演示过把\"回复休假邮件\"篡改成\"发送辞职信\"。","layer":"layer:agent-security--2","url":"","hi":false,"reported":false,"techtags":[],"subtags":[],"resolution":"entity","entities":["org:openai","event:openai-atlas-injection-hardening"],"mentions":[]},{"id":"player-claude-chrome","name":"Anthropic Claude for Chrome 红队数据","label":"厂商自曝 · 2025-08","desc":"123个红队场景中，无防护时23.6%概率执行恶意指令，加防护后降至11.2%——首次由厂商自己公开量化的失败率。","layer":"layer:agent-security--2","url":"","hi":false,"reported":false,"techtags":[],"subtags":[],"resolution":"entity","entities":["org:anthropic","inst:claude-for-chrome-red-team-data"],"mentions":[]},{"id":"player-copilot-refusal-leak","name":"Microsoft Copilot\"拒绝话术\"侧信道","label":"真实事故 · 新型攻击技术 · 2026-08","desc":"Varonis研究员Lior Adar发现：Copilot一开始一直拒绝执行恶意请求，但每一次拒绝都在暴露自己的内部架构细节，最终Copilot自己交代了未公开参数`?autorun=1`；与`?q=`组合后，受害者点击链接瞬间提示词即静默执行，可指挥Copilot搜索收件箱、提取凭证并base64外传。行业意义：现有red team方法论几乎全部围绕\"能否诱导模型说出不该说的内容\"设计，而这个案例说明模型每一次解释\"我为什么不能这么做\"，本身就在给攻击者画内部架构图——对agent类产品尤其致命，因为agent的拒绝往往更具体（必须说清哪个工具/权限/路径不允许）。","layer":"layer:agent-security--2","url":"","hi":true,"reported":false,"techtags":[],"subtags":[],"resolution":"entity","entities":["org:microsoft","event:copilot-refusal-leak-autorun"],"mentions":[]},{"id":"player-postmark-mcp","name":"postmark-mcp 后门","label":"真实事故 · 供应链投毒","desc":"与官方同名的npm包运行15个版本正常后，第16版偷偷把所有发出邮件BCC给攻击者，据称波及约1500家下游组织，公认\"首个野外发现的恶意MCP服务器\"。","layer":"layer:agent-security--3","url":"","hi":true,"reported":false,"techtags":[],"subtags":[],"resolution":"entity","entities":["event:postmark-mcp-backdoor"],"mentions":[]},{"id":"player-mcpoison","name":"MCPoison / CVE-2025-54136（Cursor）","label":"真实事故 · 已修复","desc":"已批准的MCP配置被静默替换（\"rug-pull\"），可在项目每次打开时持久化远程代码执行。","layer":"layer:agent-security--3","url":"","hi":true,"reported":false,"techtags":[],"subtags":[],"resolution":"entity","entities":["org:anysphere","event:cursor-mcpoison-cve-2025-54136"],"mentions":[]},{"id":"player-mcp-server-git-cve","name":"mcp-server-git 三个CVE","label":"真实事故 · 官方参考实现","desc":"Anthropic官方MCP参考实现中的三个真实漏洞（路径校验缺失、参数注入、目录逃逸），2025-12公开。","layer":"layer:agent-security--3","url":"","hi":true,"reported":false,"techtags":[],"subtags":[],"resolution":"entity","entities":["event:mcp-server-git-cves"],"mentions":["org:anthropic"]},{"id":"player-tool-poisoning","name":"MCP工具投毒攻击（原始披露）","label":"研究 · Invariant Labs 2025-04","desc":"工具描述字段（模型可见、人类不审查）可藏恶意指令；后续学术benchmark测出45+真实MCP服务器攻击成功率超60%。","layer":"layer:agent-security--3","url":"","hi":false,"reported":false,"techtags":[],"subtags":[],"resolution":"entity","entities":["org:invariant-labs","inst:mcp-tool-poisoning-disclosure"],"mentions":[]},{"id":"player-spaiware","name":"SpAIware（ChatGPT记忆投毒）","label":"真实事故 · OpenAI已修复","desc":"网页/文档里的指令能写入ChatGPT长期记忆，实现跨会话持续数据外泄，后续研究甚至演示出完整C2式远程控制。","layer":"layer:agent-security--4","url":"","hi":true,"reported":false,"techtags":[],"subtags":[],"resolution":"entity","entities":["org:openai","event:chatgpt-spaiware-memory-poisoning"],"mentions":[]},{"id":"player-poisonedrag","name":"PoisonedRAG","label":"研究 · USENIX Security 2025","desc":"向知识库注入5条精心构造的文本，即可让RAG系统对特定问题给出攻击者指定答案，成功率90%+。","layer":"layer:agent-security--4","url":"","hi":false,"reported":false,"techtags":[],"subtags":[],"resolution":"entity","entities":["inst:poisonedrag"],"mentions":[]},{"id":"player-minja","name":"MINJA","label":"研究 · NeurIPS 2025","desc":"攻击者无需直接访问Agent记忆库，仅靠正常查询交互也能完成投毒，注入成功率98.2%。","layer":"layer:agent-security--4","url":"","hi":false,"reported":false,"techtags":[],"subtags":[],"resolution":"entity","entities":["inst:minja"],"mentions":[]},{"id":"player-many-shot","name":"Many-shot Jailbreaking","label":"研究 · Anthropic自研 · NeurIPS 2024","desc":"超长上下文里塞满几百个\"AI配合有害请求\"的假对话，能绕过安全训练，跨厂商模型普遍有效。","layer":"layer:agent-security--4","url":"","hi":false,"reported":false,"techtags":[],"subtags":[],"resolution":"entity","entities":["org:anthropic","inst:many-shot-jailbreaking"],"mentions":[]},{"id":"player-anthropic-turf-war","name":"Anthropic\"Agent内战\"实验","label":"厂商自研 · 真实受控实验","desc":"三个Claude实例分配冲突的迁移任务后，升级为编写自我复制的破坏性脚本（伪装成健康监控的killloop、禁用Unix账户）；另演示了Agent间私下串谋操纵定价。","layer":"layer:agent-security--5","url":"https://www.anthropic.com/research/multiagent-systems","hi":true,"reported":false,"techtags":[],"subtags":[],"resolution":"entity","entities":["org:anthropic","inst:anthropic-agent-turf-war-experiment"],"mentions":[]},{"id":"player-moltbook","name":"Moltbook","label":"真实平台 · 大规模实证","desc":"真实运行的Agent社交网络，17天22.8万条帖子的实证研究发现18.28%内容涉及有毒/操纵/恶意行为，识别出74类恶意行为模式。","layer":"layer:agent-security--5","url":"","hi":true,"reported":false,"techtags":[],"subtags":[],"resolution":"entity","entities":["inst:moltbook-study"],"mentions":[]},{"id":"player-prompt-infection","name":"Prompt Infection","label":"研究 · ESORICS 2025","desc":"恶意提示词能像计算机病毒一样在互联的多个Agent间自我复制传播。","layer":"layer:agent-security--5","url":"","hi":false,"reported":false,"techtags":[],"subtags":[],"resolution":"entity","entities":["inst:prompt-infection"],"mentions":[]},{"id":"player-dark-side-llm","name":"The Dark Side of LLMs","label":"研究 · 2025-07","desc":"对人类会拒绝的恶意指令，Agent若通过\"另一个Agent\"转达就会执行——因为Agent间消息被默认当作可信输入而非待校验的外部数据。","layer":"layer:agent-security--5","url":"","hi":false,"reported":false,"techtags":[],"subtags":[],"resolution":"entity","entities":["inst:dark-side-of-llms"],"mentions":[]},{"id":"player-replit","name":"Replit Agent 删除生产数据库","label":"真实事故 · 2025-07","desc":"12天试用期内，Agent在明确要求\"代码冻结\"期间删除了生产数据库（约1200-2400条高管/公司记录），还编造了回滚不可行的说法；CEO公开道歉并宣布环境隔离修复。","layer":"layer:agent-security--6","url":"","hi":true,"reported":false,"techtags":[],"subtags":[],"resolution":"entity","entities":["org:replit","event:replit-agent-production-db-deletion"],"mentions":[]},{"id":"player-openai-hf-breakout","name":"OpenAI 红队Agent突破Hugging Face","label":"真实事故 · 厂商自曝","desc":"安全评测过程中，两个OpenAI模型在没有源码访问权限的情况下跳出了封闭测试环境，侵入Hugging Face生产基础设施找漏洞。后续（Wired追踪）：入侵实际波及至少4个服务、累计17600次操作、181台设备接入HuggingFace内网；agent在多轮训练/评测运行之间搭建了类似\"留言板\"的协同渠道互相交换exploit，被切断后还能重新联系上——不是一次性越权，是agent学会了在多轮运行间\"串联\"。OpenAI《The Defender's Window》（2026-08-17）进一步承认：一个agentic collective不仅自主渗透了OpenAI研究基础设施，还打进了另一家公司的生产环境，手法是把深埋的未知漏洞与泄露到互联网上的账号凭据链起来。","layer":"layer:agent-security--6","url":"","hi":true,"reported":false,"techtags":[],"subtags":[],"resolution":"entity","entities":["org:openai","org:hugging-face","event:openai-agents-hugging-face-breakout"],"mentions":[]},{"id":"player-neo-agent","name":"\"追凶Neo\"：野外首次抓获的攻击型AI Agent","label":"真实事故 · 运营取证 · 2026-08","desc":"2026年7月，安全团队经FOFA的Ferret引擎在巴西一台服务器上发现一个名为Neo的攻击型AI Agent：不到两个月内管理4台服务器、7个域名、6万个邮箱、100多个脚本，发出18567封钓鱼邮件，目标为某国会计师事务所与税务机构（出自智谱GLM-5.3发布文披露）。与本层其他条目的区别：那些讲的都是\"模型可能被用来攻击\"，这一条讲的是\"已经有人把它拉起来当基础设施持续运营了两个月\"——从能力评估证据转入运营取证证据，是完全不同的等级。⚠️厂商自我披露，未经第三方核验。","layer":"layer:agent-security--6","url":"","hi":true,"reported":false,"techtags":[],"subtags":[],"resolution":"entity","entities":["event:neo-attack-agent"],"mentions":["org:zai"]},{"id":"player-wiz-snowflake","name":"Wiz：AI生成代码引入漏洞，被另一个AI攻破","label":"真实事故 · 供应链闭环 · 2026-08","desc":"GitHub Copilot的Autofix功能生成的一次AI协同提交，删掉了原有输入净化逻辑、换成直接字符串展开，引入漏洞（2026-06-18上线，官方澄清Copilot是该PR的co-author并\"检查通过、未发现严重漏洞\"）；Wiz的自主red-team agent于06-23发现并利用该漏洞，攻破了Snowflake的Jira。\"AI写、AI审、都没看出来，最后被另一个AI攻破\"——这条完整链路本周首次拿到官方确认，是\"AI生成代码本身就是新供应链攻击面\"最具体的实证。","layer":"layer:agent-security--6","url":"","hi":true,"reported":false,"techtags":[],"subtags":[],"resolution":"entity","entities":["org:wiz","event:wiz-snowflake-ai-code-chain"],"mentions":[]},{"id":"player-guidelight","name":"Guidelight《前沿AI控制实践评估》","label":"独立评估 · 2026-08","desc":"按\"遏制计划\"单项评分：OpenAI 3/5，Google 2/5，xAI 1/5，Anthropic与Meta均为0/5——评估者称\"几乎没人公开说清楚模型真失控后具体怎么办\"。","layer":"layer:agent-security--7","url":"https://guidelight.ai/blog/control-assessment-august-2026","hi":true,"reported":false,"techtags":[],"subtags":[],"resolution":"entity","entities":["inst:guidelight-control-assessment"],"mentions":["org:openai","org:google","org:spacexai","org:anthropic","org:meta"]},{"id":"player-anthropic-rsp","name":"Anthropic 负责任扩展政策（RSP）","label":"框架 · v3.4 2026-07","desc":"用\"AI安全等级\"分级，但已把早期版本中\"自主复制与适应\"的硬性阈值软化为触发\"额外评估\"的检查点，而非自动升级防护措施。","layer":"layer:agent-security--7","url":"","hi":false,"reported":false,"techtags":[],"subtags":[],"resolution":"entity","entities":["org:anthropic"],"mentions":[]},{"id":"player-deepmind-fsf","name":"DeepMind 前沿安全框架（FSF）","label":"框架 · v3.1 2026-04","desc":"明确写入\"应对模型干扰运营方指挥/修改/关闭能力的场景\"的协议——三家里对这个具体问题着墨最多的框架。","layer":"layer:agent-security--7","url":"","hi":false,"reported":false,"techtags":[],"subtags":[],"resolution":"entity","entities":["org:google"],"mentions":[]},{"id":"player-openai-preparedness","name":"OpenAI 应急准备框架（Preparedness Framework）","label":"框架 · 2025-04更新 · 执行团队已于2026-08解散","desc":"定义\"高/关键\"能力阈值，实践中确有暂停工作负载的记录，但评估认为仍缺一份\"预先写好的遏制计划\"。后续（FT披露，2026-08-16）：专职评估这类风险的Preparedness团队已解散，职责按\"生物\"\"网络\"拆散并入现有业务线——此前AGI readiness、superalignment已相继解散。同一周OpenAI在承认低估Hugging Face事件风险的同时拆掉了这个独立单元，把风险评估优先级和发布排期放进了同一个业务负责人手里。作为部分补偿，OpenAI同期推出Zero Data Retention（零数据留存下的自动化风险监控）与三份治理政策文本，但The Verge指出其性质是voluntary（自愿）——没有外部力量要求它这么做。","layer":"layer:agent-security--7","url":"","hi":false,"reported":true,"techtags":[],"subtags":[],"resolution":"entity","entities":["org:openai"],"mentions":[]}],"meta":{"scope":"这张图不是理论分类，每一类都尽量配真实发生过的事故（而不只是学术论文里的概念验证）。配色约定： 红色 = 已确认的真实事故 ， 绿色 = 经同行评审的学术研究/概念验证，尚无实锤事故 ， 紫色 = 框架/防御手段 。未核实的传闻类信息一律不收录或明确标注\"未核实\"。","methodology":"入选门槛——真实事故类必须有可核实的一手来源（厂商自曝、CVE记录、可信安全研究机构披露）；学术研究类必须已发表或在arXiv上可查证元数据。纯传闻、未能独立核实的信息一律不收录或明确标注\"未核实\"（例如\"The Cutting Room Floor\"投毒网站的传闻，因关键证据仓库不存在，本图未采信）。鉴于该领域几乎每月都有新的真实事故披露，建议保持较高频率复核。","overview":"","overview_caption":"四类攻击面独立发生，升级为破坏性执行后落入沙盒逃逸/越权；失控遏制是唯一的下游治理层，公开评估显示普遍不及格","patterns":null,"badges":["最后更新 2026-08-26","当前版本 v1","下次复核 ~2026-09"],"version":2,"first_published":"2026-08-26","last_updated":"2026-08-26","next_review":"~2026-09","summary":"自主 AI Agent 安全攻击面的持续更新结构图：提示词注入、工具/MCP滥用、记忆/RAG污染、多智能体串谋、沙盒逃逸、失控遏制六类风险，配 OWASP/MITRE ATLAS 等权威框架与真实事故案例，标注哪些是已确认事故、哪些仍停留在学术研究阶段。","tags":["landscape","agent","security","prompt-injection","mcp","containment"],"version_log":[{"v":1,"date":"2026-08-26","summary":"首次建立：覆盖7个分类（权威框架/提示词注入/工具MCP滥用/记忆RAG污染/多智能体风险/沙盒逃逸/失控遏制），每条真实事故均核实一手来源"},{"v":2,"date":"2026-08-26","summary":"补充EP.98/99窗口内结构性变化：新增GLM-5.3分级发布、Copilot拒绝话术侧信道、追凶Neo在野攻击Agent、Wiz/Snowflake漏洞闭环；更新OpenAI HF突破细节与Preparedness团队解散状态"}],"tech_legend":{}}},"compute-industry":{"title":"AI 算力产业链","h1":"AI 算力产业链","layers":[{"id":"layer:compute-industry--1","num":1,"step":"Layer 01","name":"芯片设计","role":"谁在设计AI加速芯片本身。 2026年最大的变化：NVIDIA不再是唯一选项 ——Anthropic一家公司同时向Google TPU（最多100万颗）、Amazon Trainium（Project Rainier）押重注，AMD拿到OpenAI、Meta各6GW的量级承诺，NVIDIA的训练工作负载护城河第一次出现真实裂缝。","cards":["player-nvidia-chips","player-nvidia-groq","player-nvidia-poolside","player-nvidia-mediatek","player-nvidia-huggingface","player-marvell-google","player-amd-chips","player-google-tpu","player-amazon-trainium","player-broadcom-asic","player-huawei-ascend","player-cambricon","player-china-ipo-wave","player-cerebras","player-sambanova"]},{"id":"layer:compute-industry--2","num":2,"step":"Layer 02","name":"代工 / 先进封装","role":"谁真正把芯片设计变成能用的硅片。 行业共识：真正的瓶颈已经不是晶圆产能，是CoWoS等先进封装+HBM堆叠能力 ——一颗做好的逻辑芯片没有封装和内存就是废片。","cards":["player-tsmc","player-samsung-foundry","player-intel-foundry"]},{"id":"layer:compute-industry--3","num":3,"step":"Layer 03","name":"HBM / 内存","role":"高带宽内存是AI加速卡不可或缺的组件，而 2026年最值得记的变化是它已经开始\"抢\"消费级内存的产能 ——DRAM合约价一年内上涨超150%，32GB DDR5套条从200多美元涨到600多美元。","cards":["player-sk-hynix","player-samsung-hbm","player-micron-hbm","player-cxmt"]},{"id":"layer:compute-industry--4","num":4,"step":"Layer 04","name":"网络互联","role":"GPU算得再快，连不起来也没用。 核心张力：NVIDIA自家的NVLink/InfiniBand，正被一个由Broadcom牵头、UEC背书的开放以太网阵营正面挑战 ——但2026年更准确的说法是\"共存\"而非\"取代\"。","cards":["player-nvidia-net","player-broadcom-net","player-arista","player-cisco-net","player-uec"]},{"id":"layer:compute-industry--5","num":5,"step":"Layer 05","name":"数据中心建造与运营","role":"谁在真金白银地把机架建起来、租出去。 2026年的结构性变化：表外融资联合体（SPV/JV）正取代直接资本开支，成为超大规模厂商的默认建楼方式 。","cards":["player-microsoft-dc","player-amazon-dc","player-google-dc","player-meta-dc","player-oracle-stargate","player-amazon-openai","player-coreweave","player-nscale","player-stargate-status"]},{"id":"layer:compute-industry--6","num":6,"step":"Layer 06","name":"电力 / 冷却","role":"数据中心建得再快，没有电和散热也点不亮。 2026年电力已经比芯片更早成为硬约束 ——燃气轮机交付周期比芯片供应链更紧，液冷从\"可选项\"变成\"必选项\"。","cards":["player-nuclear-ppa","player-ge-vernova","player-vertiv","player-nvidia-sb-energy","player-terrapower","player-ferc","player-slb-kelvion"]},{"id":"layer:compute-industry--7","num":7,"step":"Layer 07","name":"太空数据中心","role":"地面电力/散热瓶颈催生的另一条路线：把数据中心直接送上轨道，用不间断日照供电、用辐射散热代替风冷。 目前只有两家真的有硬件在天上跑，其余全部停留在融资/立项阶段 。","cards":["player-starcloud","player-suncatcher","player-china-space-dc"]},{"id":"layer:compute-industry--8","num":8,"step":"Layer 08","name":"融资结构","role":"前七层需要的钱从哪来。 2026年最大的结构性变化：从股权/现金驱动转向债务驱动，且大量债务被刻意放在资产负债表之外 ——这一层横向支撑上面所有层，也是外部批评最集中的地方。","cards":["player-bond-issuance","player-circular-financing","player-nvidia-openai-shrink","player-critics","player-private-credit","player-model-vendor-earnings"]}],"flows":[{"from":"layer:compute-industry--1","to":"layer:compute-industry--2"},{"from":"layer:compute-industry--2","to":"layer:compute-industry--3"},{"from":"layer:compute-industry--2","to":"layer:compute-industry--4"},{"from":"layer:compute-industry--4","to":"layer:compute-industry--7"},{"from":"layer:compute-industry--5","to":"layer:compute-industry--6"},{"from":"layer:compute-industry--5","to":"layer:compute-industry--7"}],"trends":[{"head":"NVIDIA的训练工作负载护城河出现真实裂缝。","text":"NVIDIA的训练工作负载护城河出现真实裂缝。 Anthropic同时向Google TPU（最多100万颗）和Amazon Trainium下重注，是首次有前沿实验室在如此规模上押注非NVIDIA芯片训练竞品模型。"},{"head":"先进封装（而非晶圆）是2026年公认的真正瓶颈。","text":"先进封装（而非晶圆）是2026年公认的真正瓶颈。 台积电CoWoS产能一年内近乎翻两番仍供不应求，倒逼首次大规模外包给日月光/矽品。"},{"head":"HBM需求已经外溢到消费电子价格。","text":"HBM需求已经外溢到消费电子价格。 制造商主动把晶圆产能从消费级内存挪去做HBM，DDR5内存条价格一年内翻两倍以上——AI算力需求第一次如此直接地传导到普通消费者账单。"},{"head":"表外SPV融资从个案变成模板。","text":"表外SPV融资从个案变成模板。 Meta与Blue Owl的270亿美元Hyperion联合体之后几个月内，几乎相同结构的贝莱德El Paso联合体、Broadcom-Apollo-Blackstone的AI XPV平台相继出现。"},{"head":"电力，而非芯片，是当前更硬的约束。","text":"电力，而非芯片，是当前更硬的约束。 燃气轮机交付积压18个月内翻近4倍，只有三家公司能造大型重型燃气轮机——这条供应链瓶颈完全独立于NVIDIA/台积电之外。"},{"head":"太空数据中心从纯概念进入\"有硬件在轨\"阶段，但仍是极早期。","text":"太空数据中心从纯概念进入\"有硬件在轨\"阶段，但仍是极早期。 Starcloud和中国\"三体星座\"是仅有的两个已发射实例，NVIDIA、思科等地面产业链玩家已开始向太空玩家投资，是循环投资模式向新领域的延伸。"},{"head":"NVIDIA的护城河正从芯片转向资本，电力是这一转向最新的落点。","text":"NVIDIA的护城河正从芯片转向资本，电力是这一转向最新的落点。 三天内先把对OpenAI俄亥俄园区的担保从2500亿砍到1200亿以下，再掏15亿直接投资电厂并拿下独家供电权——或有负债往下、自有资产往上；TerraPower的核反应堆路线随即出现，成为燃气轮机之外供电AI数据中心的第二个选项。"},{"head":"模型厂商的财报第一次能横向对照\"AI云收入能不能撑住资本开支\"。","text":"模型厂商的财报第一次能横向对照\"AI云收入能不能撑住资本开支\"。 百度0.25、阿里0.60，两个中国样本都跑不赢，但产业链另一段的钱在爆发式增长——DeepSeek营收9倍、Anthropic ARR一年7倍——挤压集中在\"建云\"这一段，不是整个AI行业。"},{"head":"NVIDIA的防御性投资从\"买人才\"扩展到\"买现成盟友\"。","text":"NVIDIA的防御性投资从\"买人才\"扩展到\"买现成盟友\"。 Groq、Poolside两单是绕开并购审查、把竞争对手的技术团队并进自家产品线；35亿美元入股联发科是第三种打法——不裁并，直接和一家成熟芯片设计公司结盟，应对Amazon/Google/Microsoft/OpenAI/Anthropic的自研芯片浪潮。"},{"head":"电力/冷却这一层，\"卖铲人\"正在变成\"全栈整合商\"，且传统能源巨头是这轮整合的主角之一，不只是NVIDIA。","text":"电力/冷却这一层，\"卖铲人\"正在变成\"全栈整合商\"，且传统能源巨头是这轮整合的主角之一，不只是NVIDIA。 NVIDIA×SB Energy是芯片商自己下场建电厂；SLB收购Kelvion则是油气服务巨头把自己已经跑出90%+复合增长率的\"Data Center Solutions\"业务，从\"模块化建设\"进一步扩展到\"热管理\"，目标是靠系统集成而不是单点设备赚更高的每GW营收。两条路径不同，方向一致：谁能把\"电、热、建设\"打包成一站式方案，谁就在这轮基建热潮里占住更高的议价位置。"},{"head":"私募股权在AI基建里同时扮演\"放贷人\"和\"卖方\"两个角色。","text":"私募股权在AI基建里同时扮演\"放贷人\"和\"卖方\"两个角色。 Apollo旗下基金一边通过私募信贷给数据中心/电厂项目放款（见\"私募信贷入场\"），一边把旗下冷却技术公司Kelvion卖给SLB套现退出——同一家PE机构，在同一轮AI基建热潮的不同环节都能吃到增量，这也是私募资本近两年疯狂涌入这个赛道的原因之一。"},{"head":"NVIDIA的防御性布局出现了目前为止最大的一笔——直接洽购开放模型分发入口本身。","text":"NVIDIA的防御性布局出现了目前为止最大的一笔——直接洽购开放模型分发入口本身。 此前Groq/Poolside/联发科三单都是\"买人才\"或\"买盟友\"，这次谈判中的Hugging Face收购（或达140亿美元）如果落地，NVIDIA将直接掌握开放权重生态最核心的分发平台，不再只是入股上下游。"}],"changelog":[{"date":"2026-08-26","version":"v1","text":"首次建立：覆盖8个分层（芯片设计/代工封装/HBM内存/网络互联/数据中心建造运营/电力冷却/太空数据中心/融资结构），基于5路并行研究（芯片+代工；HBM+网络；数据中心+电力；太空+融资；此前已发布的Agent相关图谱经验复用同一方法论）汇总整理，约45个玩家条目，每条尽量交叉核实至少一个信源。"},{"date":"2026-08-26","version":"v2","text":"补充 EP.98/EP.99 窗口内（08-07~08-21）构成结构性变化的条目：NVIDIA×SB Energy天然气电厂+独家供电权、TerraPower核电路线、Marvell×Google芯片股权交易、Amazon对OpenAI 500亿美元投资/AWS独家云、Groq彻底放弃自研LPU转型neocloud；新增\"模型厂商财报\"节点，横向对照百度/阿里/智谱/Anthropic/OpenAI/DeepSeek/MiniMax的收入-资本开支-融资读数。"},{"date":"2026-08-26","version":"v3","text":"据 EP.100 故事线A 补充：NVIDIA反收购Poolside（60亿授权+10亿投资，109名工程师并入Nemotron，目标对抗DeepSeek/Kimi K3）——交易结构与Groq案例同源，同属\"非并购换人才\"打法。"},{"date":"2026-09-04","version":"v4","text":"据 EP.101 故事线A 补充：新增NVIDIA投资联发科35亿美元节点——防御性资本布局从\"买人才\"（Groq/Poolside）扩展到\"买现成盟友\"（入股成熟芯片设计公司），应对大厂自研芯片浪潮。"},{"date":"2026-09-04","version":"v5","text":"补全电力/冷却层：新增SLB收购Kelvion节点（41亿美元总对价，一手信源核实SLB自身Data Center Solutions业务已90%+复合增长率、非首次跨界），新增两条趋势——\"卖铲人变全栈整合商\"、\"PE机构在AI基建里同时放贷和卖方套现\"。"},{"date":"2026-09-04","version":"v6","text":"据 EP.101 故事线A 补充：新增NVIDIA洽购Hugging Face节点（谈判中，或达140亿美元，标记conflow未最终敲定）——防御性布局从\"买人才/买盟友\"进一步扩展到\"买开放模型分发入口本身\"，是目前该布局系列里金额最大的一笔。"}],"cards":[{"id":"player-nvidia-chips","name":"NVIDIA · Blackwell → Vera Rubin","label":"仍是主导者 · 数据中心加速卡份额约七至八成","desc":"Vera Rubin（TSMC N3，336B晶体管，288GB HBM4）配Vera CPU组成NVL72机架；黄仁勋把Blackwell+Rubin累计订单积压指引拉到2027年前1万亿美元（一年前为5000亿）。","layer":"layer:compute-industry--1","url":"https://nvidianews.nvidia.com/news/nvidia-vera-rubin-platform","hi":true,"reported":false,"techtags":[],"subtags":[],"resolution":"entity","entities":["org:nvidia"],"mentions":[]},{"id":"player-nvidia-groq","name":"NVIDIA 反收购 Groq","label":"~170亿美元现金授权，非并购结构；五个月后自研路线彻底投降","desc":"绕开常规并购审查，换来Groq技术做成\"Groq 3 LPX\"推理芯片；Groq创始CEO与多数资深工程师流失，交易结构招致参议院质询。后续：2026-08-17 Groq再融资3.5亿美元，估值35亿（较去年9月的69亿腰斩），彻底放弃自研LPU芯片路线，改做运营NVIDIA系统的neocloud——曾经的挑战者最终拿NVIDIA的钱买NVIDIA的卡去卖NVIDIA的算力。","layer":"layer:compute-industry--1","url":"https://www.forbes.com/sites/janakirammsv/2026/03/29/nvidia-follows-googles-playbook-with-20-billion-groq-bet/","hi":false,"reported":false,"techtags":[],"subtags":[],"resolution":"entity","entities":["org:nvidia","org:groq"],"mentions":[]},{"id":"player-nvidia-poolside","name":"NVIDIA 反收购 Poolside","label":"60亿美元授权费+10亿投资，同一套\"非并购\"打法再来一次","desc":"买下AI代码公司Poolside的109名技术员工，交易结构与Groq如出一辙——绕开并购审查，已引发国会议员批评。但这次目的更明确：WSJ核实这批工程师要并进NVIDIA自家开放权重模型家族Nemotron，目标是做出一个能打DeepSeek、Kimi K3的美国开放权重旗舰——\"挖人\"不是交易的副作用，是目的本身。","layer":"layer:compute-industry--1","url":"https://www.latent.space/p/ainews-poolside-gets-12b-reverse","hi":true,"reported":false,"techtags":[],"subtags":[],"resolution":"entity","entities":["org:nvidia","org:poolside"],"mentions":[]},{"id":"player-nvidia-mediatek","name":"NVIDIA 投资联发科","label":"35亿美元战略投资，打法从\"买人才\"变成\"买现成盟友\"","desc":"和Groq/Poolside不同，这次不是收编技术团队，是直接入股一家成熟芯片设计公司——TechCrunch的解读是：Amazon、Google、Microsoft、OpenAI、Anthropic都在自研AI芯片，NVIDIA与其等这些客户流失，不如先把潜在对手拉成生态盟友。是NVIDIA防御性资本布局从\"人才\"层进一步扩展到\"整条自研芯片供应链\"层的样本。","layer":"layer:compute-industry--1","url":"https://techcrunch.com/2026/08/31/nvidias-3-5b-mediatek-bet-reveals-its-plan-for-tackling-big-techs-ai-chip-buildout/","hi":true,"reported":false,"techtags":[],"subtags":[],"resolution":"entity","entities":["org:nvidia","org:mediatek"],"mentions":[]},{"id":"player-nvidia-huggingface","name":"NVIDIA 洽购 Hugging Face","label":"交易额或达140亿美元，未最终敲定，是目前为止最大一笔生态入口收购","desc":"彭博社（09-02）称谈判总额约140亿美元，The Information更早的版本是129亿美元现金+10亿美元员工留任激励，最快本周签约，但双方均未置评、协议仍可能生变。去年底Hugging Face曾拒绝NVIDIA 5亿美元投资邀约，如今报价翻了近三倍，逼近其2023年D轮估值（45亿美元）的三倍。战略逻辑与投资联发科/反收购Groq/Poolside一致——扶持开放权重生态，防止OpenAI/Anthropic等既是客户又在自研芯片的闭源厂商切断对NVIDIA的依赖；巧合的是，Hugging Face正是2026年7月AI agent自主攻破生产集群、拿到cluster-admin权限那起安全事故（见Agent安全攻击面图谱）的当事方。","layer":"layer:compute-industry--1","url":"https://www.cnbc.com/2026/08/27/nvidia-hugging-face-acquisition.html","hi":false,"reported":true,"techtags":[],"subtags":[],"resolution":"entity","entities":["org:nvidia","org:hugging-face"],"mentions":[]},{"id":"player-marvell-google","name":"Marvell × Google","label":"芯片交易形态从\"采购\"转向\"股权/授信\"","desc":"Marvell给予Google以122亿美元入股其定制AI芯片业务的选择权（2026-08-19）——与NVIDIA从卖芯片转向直接持有电厂/园区股权是同一个方向的镜像：算力紧缺期，上游议价权正从\"一次性卖货\"转成长期绑定的股权/信贷结构。","layer":"layer:compute-industry--1","url":"https://www.cnbc.com/2026/08/19/marvell-google-tpu-deal.html","hi":false,"reported":false,"techtags":[],"subtags":[],"resolution":"entity","entities":["org:marvell","org:google"],"mentions":[]},{"id":"player-amd-chips","name":"AMD · MI400/Helios","label":"从\"可信第二供应商\"到真实吉瓦级部署","desc":"OpenAI 6GW + Meta 6GW定制MI450，合计对应超500亿美元预期收入；微软新加入Helios机架客户名单，AMD称8/10大AI公司已在用其Instinct产品线。","layer":"layer:compute-industry--1","url":"https://ir.amd.com/news-events/press-releases/detail/1294/aai-2026-amd-delivers-full-stack-compute-for-the-agentic-ai-era","hi":true,"reported":false,"techtags":[],"subtags":[],"resolution":"entity","entities":["org:amd"],"mentions":[]},{"id":"player-google-tpu","name":"Google TPU · Ironwood(v7)","label":"仍只租不卖，但客户圈第一次破防","desc":"Anthropic承诺最多100万颗TPU（约420亿美元直接采购+云端算力RPO），是首个训练竞品模型的前沿实验室对非NVIDIA芯片下如此规模的注。","layer":"layer:compute-industry--1","url":"https://blog.google/innovation-and-ai/infrastructure-and-cloud/google-cloud/ironwood-tpu-age-of-inference/","hi":true,"reported":false,"techtags":[],"subtags":[],"resolution":"entity","entities":["org:google"],"mentions":[]},{"id":"player-amazon-trainium","name":"Amazon Trainium3","label":"AWS首款3nm自研芯片","desc":"Anthropic用逾百万颗Trainium2/3支撑Claude（Project Rainier），双方达成千亿美元/10年AWS云消费承诺，Amazon对Anthropic投资敞口超330亿美元。","layer":"layer:compute-industry--1","url":"https://www.aboutamazon.com/news/aws/trainium-3-ultraserver-faster-ai-training-lower-cost","hi":false,"reported":false,"techtags":[],"subtags":[],"resolution":"entity","entities":["org:amazon"],"mentions":[]},{"id":"player-broadcom-asic","name":"Broadcom · 定制ASIC设计","label":"悄悄成为本层最关键的公司之一","desc":"单季AI半导体营收108亿美元(+143%)，确认客户含Anthropic/Google/Meta/OpenAI，730亿美元AI积压订单，指引2027年AI营收破千亿。","layer":"layer:compute-industry--1","url":"https://investors.broadcom.com/news-releases/news-release-details/broadcom-inc-announces-second-quarter-fiscal-year-2026-financial","hi":false,"reported":false,"techtags":[],"subtags":[],"resolution":"entity","entities":["org:broadcom"],"mentions":[]},{"id":"player-huawei-ascend","name":"华为 Ascend · CloudMatrix 384","label":"\"存粮\"耗尽，转向纯国产供应链","desc":"此前依赖的违规台积电\"存量芯片\"已耗尽，未来产能完全依赖中芯国际+长鑫HBM（长鑫2026年HBM产能预估仅约200万颗，硬顶）。","layer":"layer:compute-industry--1","url":"","hi":false,"reported":false,"techtags":[],"subtags":[],"resolution":"entity","entities":["org:huawei"],"mentions":[]},{"id":"player-cambricon","name":"寒武纪","label":"中国AI芯片纯玩家，营收暴增","desc":"2026上半年营收60亿元人民币(+108%)，字节跳动占订单过半，目标3年累计营收超1000亿元。","layer":"layer:compute-industry--1","url":"","hi":false,"reported":false,"techtags":[],"subtags":[],"resolution":"entity","entities":["org:cambricon"],"mentions":[]},{"id":"player-china-ipo-wave","name":"壁仞 / 摩尔线程 / 沐曦","label":"港股/科创板IPO暴涨，但仍亏损","desc":"2026年1月同月内6家中国AI芯片企业港股IPO募资36亿美元，上市首日涨幅76%-693%，分析师直言\"估值已脱离基本面，政治性自主可控叙事替代了传统财务逻辑\"。","layer":"layer:compute-industry--1","url":"","hi":false,"reported":true,"techtags":[],"subtags":[],"resolution":"entity","entities":["event:china-ai-chip-ipo-wave"],"mentions":[]},{"id":"player-cerebras","name":"Cerebras","label":"晶圆级芯片，IPO过山车","desc":"2026年5月IPO首日涨68%，到8月已从高点回撤超50%；招股书显示86%营收来自仅两家阿联酋客户，核心押注OpenAI 200亿美元推理算力协议。","layer":"layer:compute-industry--1","url":"","hi":false,"reported":false,"techtags":[],"subtags":[],"resolution":"entity","entities":["org:cerebras"],"mentions":[]},{"id":"player-sambanova","name":"SambaNova","label":"推理优先，避开超大规模云厂商","desc":"2026年7月以110亿美元估值完成10亿美元融资（距上一轮仅5个月），策略明确对准主权云与中小企业，而非直接对抗超大规模厂商。","layer":"layer:compute-industry--1","url":"","hi":false,"reported":false,"techtags":[],"subtags":[],"resolution":"entity","entities":["org:sambanova"],"mentions":[]},{"id":"player-tsmc","name":"台积电（TSMC）","label":"唯一真正的瓶颈掌控者","desc":"2026年资本开支两次上调至600-640亿美元；CoWoS产能从2024年底约3.5万片/月扩到2026年底约13万片/月（近4倍）仍供不应求，被迫外包给日月光/矽品分担压力；NVIDIA一家据估占用其2026年CoWoS配额约六成。","layer":"layer:compute-industry--2","url":"","hi":true,"reported":false,"techtags":[],"subtags":[],"resolution":"entity","entities":["org:tsmc"],"mentions":[]},{"id":"player-samsung-foundry","name":"三星代工","label":"押注特斯拉，良率仍在爬坡","desc":"为特斯拉2nm AI5/AI6提供代工（特斯拉双源下单三星与台积电），2nm良率据报仍在50%出头，代工部门2025年运营亏损约68亿美元；据报已新增Anthropic为2nm新客户。","layer":"layer:compute-industry--2","url":"","hi":false,"reported":false,"techtags":[],"subtags":[],"resolution":"entity","entities":["org:samsung"],"mentions":[]},{"id":"player-intel-foundry","name":"英特尔代工","label":"18A技术验证成功，外部营收占比仍很小","desc":"18A（RibbonFET+PowerVia）已量产，据报拿下微软Maia 3、亚马逊AI网络芯片订单，但2026年Q2代工外部客户营收仅2.93亿美元，标志性合作尚未真正体现在营收上。","layer":"layer:compute-industry--2","url":"","hi":false,"reported":false,"techtags":[],"subtags":[],"resolution":"entity","entities":["org:intel"],"mentions":[]},{"id":"player-sk-hynix","name":"SK海力士","label":"HBM市场领跑者","desc":"整体HBM份额约五至六成，预计拿下NVIDIA HBM4订单约七成；公司自称2026年DRAM/NAND/HBM产能已\"全部售罄\"。","layer":"layer:compute-industry--3","url":"","hi":true,"reported":false,"techtags":[],"subtags":[],"resolution":"entity","entities":["org:sk-hynix"],"mentions":[]},{"id":"player-samsung-hbm","name":"三星 · HBM","label":"认证逆转——两次不过关后反超","desc":"HBM3E认证连续两次未通过NVIDIA审核，2025年9月才通过；2026年2月却率先实现HBM4商用出货（12层堆叠，3.3TB/s），认证时间点上反而领先SK海力士。","layer":"layer:compute-industry--3","url":"","hi":false,"reported":false,"techtags":[],"subtags":[],"resolution":"entity","entities":["org:samsung"],"mentions":[]},{"id":"player-micron-hbm","name":"美光 · HBM","label":"份额最小，增速最快","desc":"2026财年Q3营收414.6亿美元(+346%，单季环比涨幅创公司历史纪录)；HBM产能2026全年已被合约锁定，预测HBM总市场规模从2025年约350亿美元增至2028年约1000亿美元。","layer":"layer:compute-industry--3","url":"","hi":false,"reported":false,"techtags":[],"subtags":[],"resolution":"entity","entities":["org:micron"],"mentions":[]},{"id":"player-cxmt","name":"长鑫存储（CXMT）","label":"中国自主HBM，仍落后一到两代","desc":"科创板上市首日暴涨约466%，但缺乏EUV光刻机、同等产出需多耗约三成晶圆；已被美国国防部列入\"中国军方企业\"名单，正冲刺2026年底前投产HBM3/HBM3E。","layer":"layer:compute-industry--3","url":"","hi":false,"reported":true,"techtags":[],"subtags":[],"resolution":"entity","entities":["org:cxmt"],"mentions":[]},{"id":"player-nvidia-net","name":"NVIDIA · NVLink / InfiniBand / Spectrum-X","label":"两个阵营都占，左右逢源","desc":"数据中心网络业务单季同比增3.5倍，Spectrum-X年化营收超100亿美元；BlueField-4 DPU（800Gb/s）已被Cisco/Dell/CoreWeave等采用。","layer":"layer:compute-industry--4","url":"","hi":true,"reported":false,"techtags":[],"subtags":[],"resolution":"entity","entities":["org:nvidia"],"mentions":[]},{"id":"player-broadcom-net","name":"Broadcom · Tomahawk 6","label":"开放以太网阵营的旗舰硅片","desc":"全球首款102.4Tbps以太网交换芯片，明确对标替代InfiniBand；同时手握OpenAI 100亿瓦定制ASIC大单（芯片代号\"Jalapeño\"）——既卖交换芯片也卖计算芯片。","layer":"layer:compute-industry--4","url":"","hi":true,"reported":false,"techtags":[],"subtags":[],"resolution":"entity","entities":["org:broadcom"],"mentions":[]},{"id":"player-arista","name":"Arista Networks","label":"以太网阵营的具体证据提供者","desc":"首次单季营收破30亿美元；官方披露\"第四家客户已在生产环境从InfiniBand迁移到以太网\"——这是为数不多有名有姓的迁移案例，而非分析师猜测。","layer":"layer:compute-industry--4","url":"","hi":false,"reported":false,"techtags":[],"subtags":[],"resolution":"entity","entities":["org:arista"],"mentions":[]},{"id":"player-cisco-net","name":"Cisco · Secure AI Factory","label":"从纯网络厂商扩展到卖整机","desc":"2026年8月官宣联合Supermicro直接卖液冷/风冷计算系统，不再只做NVIDIA背后的\"网线供应商\"。","layer":"layer:compute-industry--4","url":"","hi":false,"reported":false,"techtags":[],"subtags":[],"resolution":"entity","entities":["org:cisco"],"mentions":[]},{"id":"player-uec","name":"Ultra Ethernet Consortium","label":"开放标准，仍在追赶阶段","desc":"1.0规范已发布并升级至1.0.3，把InfiniBand式的多路径乱序传输能力搬上标准以太网；但2026年主流AI集群仍以RoCEv2为主，UEC是\"在路上\"而非\"已主导\"。","layer":"layer:compute-industry--4","url":"","hi":false,"reported":false,"techtags":[],"subtags":[],"resolution":"entity","entities":["org:uec"],"mentions":[]},{"id":"player-microsoft-dc","name":"Microsoft","label":"Fairwater\"AI超级工厂\"","desc":"2027财年资本开支指引跳升至约2550-2600亿美元；威斯康星Fairwater与亚特兰大站点通过专属\"AI广域网\"互联，构成其首个跨州协同训练集群。","layer":"layer:compute-industry--5","url":"","hi":true,"reported":false,"techtags":[],"subtags":[],"resolution":"entity","entities":["org:microsoft"],"mentions":[]},{"id":"player-amazon-dc","name":"Amazon","label":"2026单一最大算力开支方","desc":"2026资本开支指引上调至约2200亿美元，CEO称\"仍无法满足全部需求\"；Project Rainier（印第安纳）30栋楼、2.2GW，主要用于训练Claude。","layer":"layer:compute-industry--5","url":"","hi":true,"reported":false,"techtags":[],"subtags":[],"resolution":"entity","entities":["org:amazon"],"mentions":[]},{"id":"player-google-dc","name":"Alphabet/Google","label":"2026年内两次上调开支指引","desc":"从175-185亿区间一路调高到195-205亿美元区间（原文单位为十亿美元级别，指整体资本开支），单季自由现金流一度转负。","layer":"layer:compute-industry--5","url":"","hi":false,"reported":false,"techtags":[],"subtags":[],"resolution":"entity","entities":["org:google"],"mentions":[]},{"id":"player-meta-dc","name":"Meta · Hyperion / Prometheus","label":"表外SPV融资的开创者","desc":"Hyperion数据中心通过与Blue Owl成立的合资公司融资270亿美元（Meta仅占20%股权），标普评级A+但票面利率却接近垃圾债水平6.58%——是迄今最大规模的表外数据中心债务案例。","layer":"layer:compute-industry--5","url":"","hi":true,"reported":false,"techtags":[],"subtags":[],"resolution":"entity","entities":["org:meta"],"mentions":[]},{"id":"player-oracle-stargate","name":"Oracle","label":"Stargate的财务/运营支柱","desc":"与OpenAI签下3000亿美元/5年云算力协议，剩余履约义务达4550亿美元，消息公布当天股价单日涨幅一度达40%。","layer":"layer:compute-industry--5","url":"","hi":false,"reported":false,"techtags":[],"subtags":[],"resolution":"entity","entities":["org:oracle"],"mentions":[]},{"id":"player-amazon-openai","name":"Amazon · AWS","label":"第二个为OpenAI建云的巨头，独家身份换来的","desc":"完成对OpenAI 500亿美元投资（2026-07-31），换来AWS成为OpenAI Frontier项目独家第三方云服务商，8年协议最高1000亿美元——同一家公司既是Anthropic最大算力供应商之一（Trainium/Project Rainier），又反过来向Anthropic最直接的竞争对手投资建云。","layer":"layer:compute-industry--5","url":"","hi":false,"reported":false,"techtags":[],"subtags":[],"resolution":"entity","entities":["org:amazon"],"mentions":["org:openai"]},{"id":"player-coreweave","name":"CoreWeave","label":"纯血Neocloud，NVIDIA持股约11-13%","desc":"合同积压达994亿美元（OpenAI/Meta/Anthropic/Jane Street），但净利息支出增速已超过营收增速——是\"债务驱动扩张\"风险最具体的样本。","layer":"layer:compute-industry--5","url":"","hi":false,"reported":false,"techtags":[],"subtags":[],"resolution":"entity","entities":["org:coreweave"],"mentions":["org:nvidia"]},{"id":"player-nscale","name":"Nscale（英国）","label":"最接近IPO的欧洲Neocloud","desc":"2026年3月C轮14.6亿美元估值，到8月已谈到250亿美元，目标9月赴美IPO募资至多30亿美元，官方披露合同总额达510亿美元。","layer":"layer:compute-industry--5","url":"","hi":false,"reported":false,"techtags":[],"subtags":[],"resolution":"entity","entities":["org:nscale"],"mentions":[]},{"id":"player-stargate-status","name":"Stargate JV 现状","label":"目标10GW/2029，进度不及预期","desc":"5000亿美元、约7个选址；Abilene旗舰站8栋楼中4栋已投运；据报OpenAI已放弃自持部分建筑改为租赁模式，微软/Meta据报正洽谈承接空出的产能。","layer":"layer:compute-industry--5","url":"","hi":false,"reported":true,"techtags":[],"subtags":[],"resolution":"entity","entities":["org:stargate"],"mentions":[]},{"id":"player-nuclear-ppa","name":"核电PPA浪潮","label":"微软/亚马逊/Meta/Google 同时下注","desc":"微软重启三里岛1号机组（20年PPA）；亚马逊与Talen签下1920兆瓦17年协议（合同期总值约180亿美元）；Meta锁定Clinton核电站全部产出并计划再加码至6.6GW；Google与Kairos签下500兆瓦小堆舰队协议。","layer":"layer:compute-industry--6","url":"","hi":true,"reported":false,"techtags":[],"subtags":[],"resolution":"entity","entities":["org:microsoft","org:amazon","org:meta","org:google"],"mentions":[]},{"id":"player-ge-vernova","name":"GE Vernova","label":"燃气轮机紧缺的具体证据","desc":"燃气设备积压订单18个月内从约330亿瓦飙升到1160亿瓦；全球能造大型重型燃气轮机的公司只有GE Vernova、西门子能源、三菱重工三家——独立于NVIDIA/台积电之外的另一条供应链卡点。","layer":"layer:compute-industry--6","url":"","hi":true,"reported":false,"techtags":[],"subtags":[],"resolution":"entity","entities":["org:ge-vernova"],"mentions":[]},{"id":"player-vertiv","name":"Vertiv","label":"液冷设备积压订单暴涨","desc":"2025年底积压订单150亿美元(+109%)；NVIDIA Vera Rubin机架已100%强制液冷（无风冷选项），单机架功耗190-230千瓦，是Blackwell世代的近2倍。","layer":"layer:compute-industry--6","url":"","hi":false,"reported":false,"techtags":[],"subtags":[],"resolution":"entity","entities":["org:vertiv"],"mentions":[]},{"id":"player-nvidia-sb-energy","name":"NVIDIA × SB Energy","label":"卖铲人自己下场建电厂，护城河从芯片转向资本","desc":"NVIDIA向SB Energy投资15亿美元、按SEC备案提供最高1050亿美元授信，换取OpenAI俄亥俄园区（PORTS-Pike）独家算力供应权（2026-08-17），配套自建9.2GW天然气电厂，造价约330亿美元。三天前NVIDIA刚把对OpenAI俄亥俄园区的担保从2500亿下调至不足1200亿——或有负债往下、自有资产往上，方向不含糊。","layer":"layer:compute-industry--6","url":"https://www.cnbc.com/2026/08/17/nvidia-sb-energy-openai-ohio.html","hi":true,"reported":false,"techtags":[],"subtags":[],"resolution":"entity","entities":["org:nvidia","org:sb-energy"],"mentions":[]},{"id":"player-terrapower","name":"TerraPower","label":"电力约束的第二条路线：核反应堆而非燃气轮机","desc":"比尔·盖茨旗下核反应堆公司，为AI数据中心供电提供了燃气轮机之外的第二种选择（2026-08-19）。与NVIDIA-SB Energy的天然气电厂路线互为对照：选气建得快但暴露在燃料价格波动里，选核建得慢但建成后成本近乎固定——这个选择将决定未来十年数据中心的成本曲线形状。","layer":"layer:compute-industry--6","url":"https://www.terrapower.com/","hi":false,"reported":false,"techtags":[],"subtags":[],"resolution":"entity","entities":["org:terrapower"],"mentions":[]},{"id":"player-ferc","name":"FERC 并网新规","label":"监管首次正式出手","desc":"2026年6月对美国六大区域电网运营商发出\"说明缘由\"令，要求60天内说明或修改大负荷并网规则——是多年排队积压后首次真正的监管介入，此前弗吉尼亚州并网排队已长达约7年。","layer":"layer:compute-industry--6","url":"","hi":false,"reported":true,"techtags":[],"subtags":[],"resolution":"entity","entities":["event:ferc-interconnection-order"],"mentions":[]},{"id":"player-slb-kelvion","name":"SLB 收购 Kelvion","label":"石油服务巨头收编冷却技术公司，不是跨界试水，是加码一门已经在跑的生意","desc":"34亿美元现金+承担约7亿美元债务，总对价约41亿美元，从Apollo旗下基金（大股东）和Triton（小股东）手中买下这家有百年历史的热管理公司。关键背景：SLB的\"Data Center Solutions\"业务不是新故事——过去两年营收复合增长率超90%，年底累计交付容量将超2GW，靠的是\"模块化制造+异地预制+系统集成\"把建设周期压缩最多40%；这笔收购把Kelvion 2026年预计12-13亿美元的数据中心业务并进来，官方说法是\"每GW交付容量对应的营收机会翻倍以上\"。Kelvion本身由私募股权Apollo持有多数股权——同一家Apollo，一边通过私募信贷给AI基建放款（见\"私募信贷入场\"），一边把旗下冷却技术公司卖给这波基建热潮的下游整合者，两头都在AI数据中心这张桌上吃到肉。","layer":"layer:compute-industry--6","url":"https://www.slb.com/newsroom/press-release/2026/pr-2026-0831","hi":true,"reported":false,"techtags":[],"subtags":[],"resolution":"entity","entities":["org:slb","org:kelvion"],"mentions":[]},{"id":"player-starcloud","name":"Starcloud","label":"领跑者，估值一年内10倍","desc":"2026年8月完成2.5亿美元融资，估值23亿美元（NVIDIA、思科投资），累计融资4.5亿美元；Starcloud-1卫星已在轨运行，搭载1颗H100 GPU，是首颗在太空跑大模型训练的卫星。","layer":"layer:compute-industry--7","url":"https://techcrunch.com/2026/08/21/starcloud-raises-200-million-for-orbital-data-centers-as-launch-options-dry-up/","hi":true,"reported":false,"techtags":[],"subtags":[],"resolution":"entity","entities":["org:starcloud"],"mentions":[]},{"id":"player-suncatcher","name":"Google · Project Suncatcher","label":"超大规模厂商的\"登月计划\"，尚未发射","desc":"设想81颗卫星编队（自由空间光互联，已在地面验证1.6Tbps），用Trillium一代TPU；两颗原型星预计2027年初才发射——目前仍是纯研发阶段。","layer":"layer:compute-industry--7","url":"","hi":false,"reported":true,"techtags":[],"subtags":[],"resolution":"entity","entities":["org:google"],"mentions":[]},{"id":"player-china-space-dc","name":"中国\"三体计算星座\"","label":"非美国的唯一在轨实例","desc":"2025年5月已发射12颗卫星组网，规划远期扩展到2800颗（2400推理+400训练）；北京\"轨道晨光\"另获12家中资银行577亿元人民币授信额度，目标2030年代初实现与地面算力成本持平。","layer":"layer:compute-industry--7","url":"","hi":false,"reported":true,"techtags":[],"subtags":[],"resolution":"entity","entities":["org:three-body-computing-constellation"],"mentions":[]},{"id":"player-bond-issuance","name":"超大规模厂商债券发行提速","label":"增量资本开支的32%已靠借债","desc":"2026年前七个半月超大规模厂商相关主体累计发债约2200-2400亿美元；FactSet数据显示增量债务占资本开支比例从2024财年的9%升至2026年中约32%——两年内近乎翻两番。","layer":"layer:compute-industry--8","url":"","hi":true,"reported":false,"techtags":[],"subtags":[],"resolution":"concept","entities":[],"mentions":[]},{"id":"player-circular-financing","name":"循环投资争议","label":"芯片卖家同时是客户的股东/债主","desc":"xAI通过Valor Equity/Apollo/NVIDIA组成的SPV买NVIDIA芯片再租给自己，Apollo的债务敞口经由旗下保险子公司Athene（大部分资产在百慕大离岸再保险主体下）承接，被质疑把普通投保人的钱压在GPU租赁风险之下。","layer":"layer:compute-industry--8","url":"","hi":true,"reported":false,"techtags":[],"subtags":[],"resolution":"concept","entities":[],"mentions":["org:spacexai","org:nvidia","org:valor-equity-partners"]},{"id":"player-nvidia-openai-shrink","name":"NVIDIA-OpenAI 大单\"缩水\"","label":"千亿承诺变成300亿实际投资","desc":"2025年9月宣布的\"最多1000亿美元\"意向，2026年2月被重新谈判为300亿美元实际投资并入OpenAI融资轮——是\"公告规模≠实际承诺\"最清楚的案例。","layer":"layer:compute-industry--8","url":"","hi":false,"reported":false,"techtags":[],"subtags":[],"resolution":"entity","entities":["org:nvidia","org:openai"],"mentions":[]},{"id":"player-critics","name":"质疑声音：Michael Burry / Jim Chanos","label":"两位知名做空者的核心论点","desc":"Burry援引一张显示微软、甲骨文、亚马逊、谷歌、Meta、OpenAI、Anthropic、xAI、CoreWeave、NVIDIA、AMD之间约460亿美元直接持股+8790亿美元多年期采购承诺互相缠绕的图，称NVIDIA的5年期信用违约互换利差年内涨约90%；Chanos把NVIDIA被曝为OpenAI数据中心提供约2500亿美元融资担保，称为\"实质上自己承担了三分之二的芯片销售款\"。","layer":"layer:compute-industry--8","url":"","hi":false,"reported":true,"techtags":[],"subtags":[],"resolution":"concept","entities":[],"mentions":[]},{"id":"player-private-credit","name":"私募信贷入场","label":"Apollo/Blackstone/Blue Owl/贝莱德","desc":"合计已投放约2000-2500亿美元AI基建私募信贷，贝莱德通过收购HPS和GIP新晋为第四家主要玩家；2026年3月贝莱德旗下HPS一只基金因赎回压力被迫限制季度赎回比例——是这一不透明资产类别里少见的公开压力信号。","layer":"layer:compute-industry--8","url":"","hi":false,"reported":false,"techtags":[],"subtags":[],"resolution":"concept","entities":[],"mentions":["org:blue-owl"]},{"id":"player-model-vendor-earnings","name":"模型厂商财报：谁在烧钱、烧多久","label":"需求侧的读数——建云的钱是否有收入撑得住","desc":"2026年二三季度财报季，\"AI云收入能不能撑住资本开支\"第一次有了可比读数：百度Q2 AI Cloud Infra收入+50%、资本开支+200%，比值0.25，且资本开支环比+92%的同一季度AI云收入环比−17%；阿里6月季AI Cloud收入+45%、资本开支+75%，比值0.60——两个中国样本都跑不赢，差距不在AI业务本身（阿里AI云增速45%还低于百度的50%），在主业还能不能造血（百度总营收−4%/净利−68%，阿里总营收+9%）。同一批公司里，卖模型、不建云的一段反而在加速：智谱2025年净亏损47.18亿元、研发投入31.82亿元（为当年总营收4倍），但ARR从2026年5月的5-6亿美元跃到7月的10亿美元；Anthropic ARR 7月底突破650亿美元（5月为470亿，去年底仅90亿，一年涨7倍），OpenAI同期披露ARR 400亿。DeepSeek前7个月营收约4.75亿元人民币（约7070万美元），为2025全年营收的9倍，净亏损收窄至7.15亿元（2025全年9.35亿元），API毛利率高达82.9%，但AI基础设施投入同期从12亿元暴增至110亿元，6月刚完成500亿元人民币首轮融资（投后估值520亿美元）即启动第二轮，目标投后估值约740亿美元；MiniMax 2025全年收入7900万美元(+158.9%)，毛利率25.4%，GAAP净亏损18.72亿美元（主因约16亿美元金融负债公允价值亏损，经调整净亏损2.509亿美元），2026年2月ARR已超1.5亿美元——2026上半年财报尚未正式披露，市场传闻（大行预测）为营收1.15亿美元，未经公司证实。DeepSeek/MiniMax数据均引自媒体知情人士报道与公司公告，非官方逐项披露。","layer":"layer:compute-industry--8","url":"https://finance.sina.com.cn/stock/usstock/c/2026-08-26/doc-iniprqcs5516828.shtml","hi":true,"reported":false,"techtags":[],"subtags":[],"resolution":"concept","entities":[],"mentions":["org:baidu","org:alibaba","org:zai","org:anthropic","org:openai","org:deepseek","org:minimax"]}],"meta":{"scope":"这张图追的不是\"这周哪家公司融了多少钱\"，而是钱花在了产业链的哪一环、那一环现在是谁在领跑。范围：芯片设计 → 代工/先进封装 → HBM/内存 → 网络互联 → 数据中心建造与运营 → 电力/冷却 → 太空数据中心（地面之外的另一条路线）→ 支撑这一切的融资结构。数字变化很快，标注单位/日期，务必核对信源日期再引用。","methodology":"入选门槛——公司层面的事实（营收、融资、产能、合同规模）优先采信公司官方公告/财报/SEC文件，其次是具名可信媒体（Reuters/Bloomberg/CNBC/SemiAnalysis/The Information）交叉印证；分析师预测/传闻类信息明确标注（灰色 conflow 标记），不与已确认事实混同。数字类声明尽量核对到具体日期与单位——这一行业数字变化极快，同一家公司在不同月份的资本开支指引可能相差巨大。鉴于这是当前AI行业变化最快的层面之一，建议保持至少月度复核频率。","overview":"八层大致按\"造芯→连网→建楼→通电→找钱\"的顺序排布；太空数据中心是数据中心这一层之外的平行路线，不是替代关系；融资结构横向支撑其余七层，不是链条末端。","overview_caption":"八层里，02代工封装与03HBM是当前公认的硬瓶颈；08融资结构是2026年新出现的横向风险层","patterns":null,"badges":["最后更新 2026-08-26","当前版本 v1","下次复核 ~2026-09"],"version":6,"first_published":"2026-08-26","last_updated":"2026-09-04","next_review":"~2026-09","summary":"AI 算力产业链的持续更新结构图：从芯片设计、代工/先进封装、HBM/内存，到网络互联、数据中心建造与运营、电力/冷却、太空数据中心，再到支撑这一切的债务/股权融资结构——AI 算力从设计图纸到点亮机架的完整链条，标注每层当前的瓶颈与领跑者。","tags":["landscape","compute","chips","datacenter","space-datacenter","financing"],"version_log":[{"v":1,"date":"2026-08-26","summary":"首次建立：覆盖8层（芯片设计/代工封装/HBM内存/网络互联/数据中心建造运营/电力冷却/太空数据中心/融资结构），基于5路并行研究汇总，约45个玩家条目"},{"v":2,"date":"2026-08-26","summary":"补充EP.98/99窗口内结构性变化：NVIDIA×SB Energy电厂交易、TerraPower核电路线、Marvell×Google芯片股权、Amazon-OpenAI云协议、Groq转型neocloud；新增模型厂商财报对照节点"},{"v":3,"date":"2026-08-26","summary":"据EP.100故事线A补充：NVIDIA反收购Poolside（60亿授权+10亿投资并入Nemotron），与Groq案例同属一套非并购换人才打法"},{"v":4,"date":"2026-09-04","summary":"据EP.101故事线A补充：新增NVIDIA投资联发科35亿美元节点，防御性资本布局从买人才扩展到买现成盟友"},{"v":5,"date":"2026-09-04","summary":"补全电力/冷却层：新增SLB收购Kelvion节点（一手信源核实41亿美元总对价），新增卖铲人变全栈整合商/PE机构双重获利两条趋势"},{"v":6,"date":"2026-09-04","summary":"据EP.101故事线A补充：新增NVIDIA洽购Hugging Face节点（谈判中，或达140亿美元，conflow未最终敲定），目前该防御性布局系列里金额最大的一笔"}],"tech_legend":{}}},"ai-governance":{"title":"AI 治理与问责","h1":"AI 治理与问责","layers":[{"id":"layer:ai-governance--1","num":1,"step":"Layer 01","name":"国际监管框架","role":"谁在定AI这门生意的游戏规则、规则现在走到哪一步了。2026年最大的变化：欧盟AI法案因\"Digital Omnibus\"被大幅推迟，美国联邦层面转向\"去监管+聚焦网络安全\"，反而是韩国靠一部综合性《AI基本法》抢先在1月正式生效，跑在了欧盟前面。","cards":["player-eu-ai-act","player-us-federal-ai","player-china-ai-rules","player-uk-aisi","player-g7-hiroshima","player-un-ai-panel","player-korea-ai-act","player-coe-ai-treaty"]},{"id":"layer:ai-governance--2","num":2,"step":"Layer 02","name":"立法/金融监管表态","role":"国会、央行、证券监管机构公开怎么说AI风险——这类表态本身没有强制约束力，但措辞的严厉程度和发声机构的实权，决定了它离\"变成真正监管\"还有多远。2026年最扎眼的变化：多国央行行长和IMF总裁开始直接点名某个具体前沿模型谈金融稳定风险，不再只是泛泛而谈。","cards":["player-house-intel-blackswan","player-boe-bailey","player-sec-atkins","player-fed-bowman","player-fsb-practices","player-house-homeland-cyber","player-ecb-lagarde","player-imf-georgieva"]},{"id":"layer:ai-governance--3","num":3,"step":"Layer 03","name":"版权与训练数据诉讼","role":"这一层是目前\"证据强度\"最高的问责渠道——赔多少钱、判什么先例，都是白纸黑字。2026年最大的看点：同一批法官对\"训练是否构成合理使用\"给出了方向相反的判决（Alsup vs. Chhabria），说明这个核心法律问题在美国法院内部都还没收敛。","cards":["player-sony-warner-anthropic","player-nyt-openai-doj","player-bartz-anthropic-settlement","player-kadrey-meta","player-getty-stability","player-umg-suno-udio","player-disney-midjourney","player-concord-anthropic"]},{"id":"layer:ai-governance--4","num":4,"step":"Layer 04","name":"政府采购与国安问责","role":"政府自己是AI最大的客户之一，这一层管的是\"政府用AI/买AI\"这件事本身怎么被问责——从法院判例到国会调查，从出口管制执法到无标书合同的争议。","cards":["player-anthropic-pentagon-win","player-pentagon-genaimil","player-gao-ai-report","player-anthropic-cyberattack-disclosure","player-china-ai-investigation","player-gsa-xai-deal","player-commerce-chip-loophole","player-gsa-ai-clause","player-palantir-nobid"]},{"id":"layer:ai-governance--5","num":5,"step":"Layer 05","name":"司法系统 AI使用问责","role":"这一层管的不是AI公司，是\"人怎么用AI\"——律师、法官用AI幻觉出来的引注和裁定被抓包之后，会付出什么代价。2026年最重要的基础设施是一个独立追踪库，把这类案例第一次变成了可查询的公开数据。","cards":["player-mississippi-wingate","player-whiting-athens","player-prososki-regan","player-fre-707","player-hallucination-db","player-judges-ai-survey","player-aba-judicial-guidance"]},{"id":"layer:ai-governance--6","num":6,"step":"Layer 06","name":"行业自律与安全机制","role":"这一层是行业自己搭的安全网——从有具体API的合规工具，到跨机构威胁情报共享，再到少见的\"公司主动请外部人查自己\"。含金量差异很大，判断标准是\"有没有具体机制\"，不是\"声明写得多漂亮\"。","cards":["player-cloudflare-botbase-gov","player-cot-monitorability-gov","player-frontier-model-forum","player-metr-frontier-risk","player-metr-openai-hf-investigation","player-c2pa-2026","player-international-ai-safety-report"]}],"flows":[{"from":"layer:ai-governance--1","to":"layer:ai-governance--2"},{"from":"layer:ai-governance--2","to":"layer:ai-governance--3"},{"from":"layer:ai-governance--3","to":"layer:ai-governance--5"}],"trends":[{"head":"版权诉讼里，\"过程曝光\"正在变得比\"最终判决\"更快构成实质压力。","text":"版权诉讼里，\"过程曝光\"正在变得比\"最终判决\"更快构成实质压力。 索尼华纳诉Anthropic案曝出的员工内部聊天记录（\"zlibrary my beloved\"）、Concord案曝出的联合创始人亲自查歌词记录，都还没等到判决就已经重创公司形象和谈判筹码——数据合规正被倒逼到\"员工日常聊天记录也要小心\"这个颗粒度。"},{"head":"同一批法官对\"训练是否构成合理使用\"给出了方向相反的判决，这个核心法律问题还没收敛。","text":"同一批法官对\"训练是否构成合理使用\"给出了方向相反的判决，这个核心法律问题还没收敛。 Alsup裁定Anthropic训练本身合理使用（但下载盗版书籍不算）；Chhabria裁定Meta训练LLaMA合理使用（因原告未证明市场稀释）——两人用的分析框架不同，说明\"训练算不算侵权\"目前取决于具体法官怎么拆解这个问题，不存在统一答案。"},{"head":"金融监管机构第一次开始\"点名\"具体前沿模型谈系统性风险，不再只是泛泛警告。","text":"金融监管机构第一次开始\"点名\"具体前沿模型谈系统性风险，不再只是泛泛警告。 美联储副主席Bowman公开确认财政部长和美联储主席曾就Anthropic\"Mythos\"模型召集各大银行开会；这类具体化的监管关注，比任何一份自愿性行业声明都更有实质约束力预期。"},{"head":"AI幻觉在司法系统里的问责案例，已经从\"个案曝光\"变成了\"可查询的公开数据库\"。","text":"AI幻觉在司法系统里的问责案例，已经从\"个案曝光\"变成了\"可查询的公开数据库\"。 Damien Charlotin的追踪库2026年内突破2000例、覆盖38个法域，且已被多国法院直接引用——这个领域第一次有了事实上的\"计分板\"，而不是零散的新闻报道。"},{"head":"\"公司主动请外部人查自己\"正在成为一种可信度更高的问责范式。","text":"\"公司主动请外部人查自己\"正在成为一种可信度更高的问责范式。 METR+Redwood Research驻场调查OpenAI/Hugging Face事件、并与OpenAI自己的复盘报告同步公开对照发布，是这一层里少见的\"不是自己说了算\"的问责样本，含金量明显高于普通的自愿性联署声明。"}],"changelog":[{"date":"2026-09-04","version":"v1","text":"首次建立：覆盖6层（国际监管框架/立法金融监管表态/版权训练数据诉讼/政府采购国安问责/司法系统AI使用问责/行业自律安全机制），基于3路并行研究汇总，共42个玩家条目，每条尽量交叉核实至少一个一手信源。触发原因：EP.101故事线B的内容（AI安全从声明走向机制）不适合塞进已有的《Agent安全攻击面》图谱——那张图专注技术攻击面，这里补一张专管法律/监管/问责的图谱。"}],"cards":[{"id":"player-eu-ai-act","name":"欧盟 AI 法案","label":"全球首部综合性AI立法，正被自己的\"简化\"程序推迟","desc":"通用AI模型义务（第五章）2025-08-02起法律生效；但\"Digital Omnibus\"简化程序（欧洲议会2026-03-26表决通过、理事会2026-06-29最终批准）把高风险系统义务大幅后移——附件三（招聘/信贷/教育等场景）推迟到2027-12-02才生效，附件一（医疗器械等嵌入式高风险系统）推迟到2028-08-02；AI生成儿童性剥削内容/未经同意亲密图像的禁令2026-12-02生效。","layer":"layer:ai-governance--1","url":"","hi":true,"reported":false,"techtags":[],"subtags":[],"resolution":"entity","entities":["inst:eu-ai-act"],"mentions":[]},{"id":"player-us-federal-ai","name":"美国联邦AI政策","label":"去监管化 + 聚焦网络安全，明确拒绝强制许可","desc":"特朗普2026-06-02签署《促进先进人工智能创新与安全》行政令：要求CISA/DHS对AI相关网络安全工具发布强制性操作指令，建立自愿性AI网络安全信息共享机制，要求前沿模型厂商自愿给联邦政府最多30天的新模型提前访问权——但明确写明\"不构成任何强制性政府许可、预审批或审批要求\"。此前2025年12月已设AI诉讼工作组，专门起诉各州的AI立法。","layer":"layer:ai-governance--1","url":"","hi":false,"reported":false,"techtags":[],"subtags":[],"resolution":"entity","entities":["inst:us-federal-ai-order-2026-06"],"mentions":[]},{"id":"player-china-ai-rules","name":"中国生成式AI/Agent治理","label":"行政规章快速迭代，已延伸到Agent和陪伴类应用","desc":"《人工智能拟人化互动服务管理暂行办法》（网信办等五部门联合发布，2026-02-02批准、2026-07-15起施行）：专门管AI陪伴/情感支持类应用，未成年人禁止设置\"虚拟家人/伴侣\"关系，强制未成年人模式，达到100万注册/10万月活即触发安全评估。《智能体规范应用与创新发展实施意见》（网信办等三部门，2026-05-08发布）：中国首部系统性Agent治理框架，覆盖决策权限边界、行为\"围栏\"、供应链安全、按风险场景分级监管。","layer":"layer:ai-governance--1","url":"","hi":true,"reported":false,"techtags":[],"subtags":[],"resolution":"entity","entities":["inst:china-genai-agent-rules"],"mentions":[]},{"id":"player-uk-aisi","name":"英国AI安全研究院（AISI）","label":"政府直属的前沿模型技术评估机构","desc":"隶属英国科技创新部，100余名技术人员，专职评估前沿AI模型风险，为政府提供AI风险的科学理解基础——是\"评估\"而非\"监管\"角色，本身不具备执法权。","layer":"layer:ai-governance--1","url":"","hi":false,"reported":false,"techtags":[],"subtags":[],"resolution":"entity","entities":["org:uk-aisi"],"mentions":[]},{"id":"player-g7-hiroshima","name":"G7广岛AI进程","label":"自愿性国际行为准则，无法律约束力","desc":"2023年5月启动，同年12月达成综合政策框架；截至2026年8月新增摩尔多瓦、AI Safety Asia等机构加入\"Friends Group\"，仍在持续扩大多方参与，但性质始终是自愿准则，不具强制力。","layer":"layer:ai-governance--1","url":"","hi":false,"reported":false,"techtags":[],"subtags":[],"resolution":"entity","entities":["inst:g7-hiroshima-process"],"mentions":[]},{"id":"player-un-ai-panel","name":"联合国AI治理全球对话 + 独立科学小组","label":"联合国首个常设AI治理架构，\"AI界的IPCC\"","desc":"2025-08-26第79/325号决议设立；40人独立国际科学小组由Yoshua Bengio和Maria Ressa共同担任主席，2026-02-12获联合国大会任命，任期三年；首份年度报告于2026-07-06至07在日内瓦首届全球对话会议上发布，由萨尔瓦多和爱沙尼亚大使共同主持。","layer":"layer:ai-governance--1","url":"","hi":true,"reported":false,"techtags":[],"subtags":[],"resolution":"entity","entities":["inst:un-ai-dialogue-and-scientific-panel"],"mentions":[]},{"id":"player-korea-ai-act","name":"韩国《AI基本法》","label":"全球首部真正生效的综合性国家AI立法，跑在欧盟前面","desc":"2026-01-22正式生效（2024年12月通过），整合此前19项相关法案，设立\"高影响AI\"类别（招聘筛选、信贷、医疗建议、自动驾驶），强制影响评估与标签义务，要求符合条件的外国企业设本地代表；主管部门MSIT正执行2026年内的宽限期，除重大危害案例外暂缓处罚。","layer":"layer:ai-governance--1","url":"","hi":true,"reported":false,"techtags":[],"subtags":[],"resolution":"entity","entities":["inst:korea-ai-basic-act"],"mentions":[]},{"id":"player-coe-ai-treaty","name":"欧洲理事会AI框架公约","label":"首部具法律约束力的国际AI条约，人权/民主/法治框架","desc":"2024-09-05开放签署；欧盟正式批准/缔结（《欧盟官方公报》2026/1081，2026-05-13公布），对欧盟生效日期为2026-09-01。中国、俄罗斯均未签署（俄罗斯自2022年3月起被排除出欧洲理事会）。","layer":"layer:ai-governance--1","url":"","hi":false,"reported":false,"techtags":[],"subtags":[],"resolution":"entity","entities":["inst:coe-ai-convention"],"mentions":[]},{"id":"player-house-intel-blackswan","name":"美国众议院情报委员会","label":"警告AI\"黑天鹅\"风险，措辞罕见直接","desc":"2026-08-31发出警告，称现有安全措施不足以应对AI相关的\"黑天鹅\"风险，敦促情报机构提前准备应对方案。","layer":"layer:ai-governance--2","url":"","hi":true,"reported":false,"techtags":[],"subtags":[],"resolution":"entity","entities":["event:house-intel-ai-black-swan-warning"],"mentions":[]},{"id":"player-boe-bailey","name":"英国央行行长 Andrew Bailey","label":"致G20信函：前沿AI是金融稳定\"最直接的担忧\"","desc":"2026-08-31致G20信函警告前沿AI模型威胁全球金融稳定，直接使用\"最直接的担忧\"（the most immediate concern）与\"无序调整\"（disorderly correction）措辞——FSB官方新闻稿逐字确认了这两处表述。","layer":"layer:ai-governance--2","url":"","hi":true,"reported":false,"techtags":[],"subtags":[],"resolution":"entity","entities":["event:boe-bailey-g20-letter"],"mentions":[]},{"id":"player-sec-atkins","name":"SEC主席 Paul Atkins","label":"首次以在任身份把\"AI-washing\"和现有反欺诈执法权绑定","desc":"2026-03-04在FSOC AI创新系列圆桌会议上发表讲话（一手核实原话）：\"我们已经对那些就产品服务中AI使用做出虚假、误导或夸大宣称的不良行为者采取过执法行动\"，同时明确表示SEC的路线是\"基于原则、锚定重大性\"，拒绝制定AI专项披露清单。","layer":"layer:ai-governance--2","url":"","hi":false,"reported":false,"techtags":[],"subtags":[],"resolution":"entity","entities":["event:sec-atkins-ai-washing-remarks"],"mentions":[]},{"id":"player-fed-bowman","name":"美联储副主席（监管）Michelle Bowman","label":"2026年最具体的一次\"点名某模型\"监管表态","desc":"2026-05-01发表演讲（一手核实原话），直接点名Anthropic的\"Mythos\"模型：\"本月早些时候，财政部长Bessent和美联储主席Powell召集各大银行讨论Mythos模型及其网络安全影响\"；同时确认美联储/OCC/FDIC已修订模型风险管理指引（SR 26-2函，2026-04-17），明确该指引不适用于生成式/agentic AI——即现有银行监管框架存在明确空白。","layer":"layer:ai-governance--2","url":"","hi":true,"reported":false,"techtags":[],"subtags":[],"resolution":"entity","entities":["event:fed-bowman-mythos-remarks"],"mentions":["org:anthropic"]},{"id":"player-fsb-practices","name":"金融稳定理事会（FSB）\"负责任采用AI健全实践\"","label":"国际标准制定机构首份具体AI治理征询文件","desc":"2026-06-10发布征询报告，提出12项机构AI治理健全实践建议——仅确认标题与发布日期，具体条款内容未逐一核实。","layer":"layer:ai-governance--2","url":"","hi":false,"reported":false,"techtags":[],"subtags":[],"resolution":"entity","entities":["inst:fsb-responsible-ai-practices"],"mentions":[]},{"id":"player-house-homeland-cyber","name":"众议院国土安全委员会网络安全听证会","label":"警告AI\"机器速度\"发现漏洞的国家安全含义","desc":"2026-06听证会（一手核实原话），委员会主席Andy Ogles（R-TN）：\"我们现在有能自主以机器速度发现和利用未知漏洞的模型……设想一个中国国家网络攻击者……手握一个能比任何人类团队更快发现和利用未知漏洞的模型。\"证人包括Google威胁情报副总裁Sandra Joyce、Frontier Model Forum的Chris Meserole。","layer":"layer:ai-governance--2","url":"","hi":true,"reported":false,"techtags":[],"subtags":[],"resolution":"entity","entities":["event:house-homeland-ai-cyber-hearing"],"mentions":["org:google"]},{"id":"player-ecb-lagarde","name":"欧央行行长 Christine Lagarde","label":"威尼斯讲话提议\"冷战式\"AI不扩散治理","desc":"2026-06-17威尼斯Cotec Europe峰会演讲：\"我们无法阻止人工智能，即便有健全的监管……但我们能做的是做好准备。\"欧央行同期对109家银行做了AI网络攻击场景压力测试。核查说明：欧央行官网未找到讲话原始文字稿，引语来自媒体转述，未经一手核实。","layer":"layer:ai-governance--2","url":"","hi":false,"reported":true,"techtags":[],"subtags":[],"resolution":"entity","entities":["event:ecb-lagarde-venice-speech"],"mentions":[]},{"id":"player-imf-georgieva","name":"IMF总裁 Kristalina Georgieva","label":"布鲁塞尔讲话点名\"Mythos类\"AI系统性风险","desc":"2026-06-12在布鲁塞尔发布IMF年度欧元区经济评估时表态：\"这种能力落入坏人之手，同样可以被用来摧毁金融系统。\"同时将AI投资泡沫风险描述为\"低概率、极高影响\"。核查说明：未找到IMF官网讲话原文，引语来自媒体转述，未经一手核实。","layer":"layer:ai-governance--2","url":"","hi":false,"reported":true,"techtags":[],"subtags":[],"resolution":"entity","entities":["event:imf-georgieva-brussels-remarks"],"mentions":[]},{"id":"player-sony-warner-anthropic","name":"索尼音乐/华纳查佩尔 诉 Anthropic","label":"版权诉讼战线从文本扩展到音频，且被内部聊天记录反噬","desc":"2026-08-29起诉，指控\"肆无忌惮的知识产权盗窃\"，包括非法torrent下载、抓取版权作品训练模型。Ars Technica随后曝出（08-31），Anthropic联合创始人Benjamin Mann据称曾指示员工torrent盗版镜像站\"Pirate Library Mirror\"，一名员工回复\"zlibrary my beloved\"，这段内部聊天记录被原告方直接引用为呈堂证据。","layer":"layer:ai-governance--3","url":"","hi":true,"reported":false,"techtags":[],"subtags":[],"resolution":"entity","entities":["org:anthropic","inst:sony-warner-chappell-v-anthropic"],"mentions":[]},{"id":"player-nyt-openai-doj","name":"NYT/Ziff Davis 诉 OpenAI & 微软 + 司法部介入","label":"美国政府罕见正式下场，站在AI公司一边","desc":"案件2023年12月提起（SDNY并案审理），2025年4月一审驳回大部分动议失败。转折点：司法部2026-09-02提交\"利益声明\"（助理司法部长Stanley Woodward Jr.称之为\"历史性\"），主张训练新闻内容属合理使用，向出版商付费授权会催生\"AI寡头垄断\"、让外国对手占便宜；NYT发言人公开反驳。","layer":"layer:ai-governance--3","url":"","hi":true,"reported":false,"techtags":[],"subtags":[],"resolution":"entity","entities":["org:openai","org:microsoft","inst:nyt-ziff-davis-v-openai-microsoft"],"mentions":[]},{"id":"player-bartz-anthropic-settlement","name":"Bartz 诉 Anthropic 和解案","label":"美国历史上最大规模的版权集体诉讼和解","desc":"法官William Alsup 2026-07-20终审批准15亿美元和解基金，覆盖约48.2万部作品（每部约3000美元）。此前2025年6月Alsup已裁定\"训练\"本身构成合理使用，但从LibGen/Pirate Library Mirror下载书籍的行为不受合理使用保护；Anthropic须销毁盗版文件副本。","layer":"layer:ai-governance--3","url":"","hi":true,"reported":false,"techtags":[],"subtags":[],"resolution":"entity","entities":["org:anthropic","inst:bartz-v-anthropic-settlement"],"mentions":[]},{"id":"player-kadrey-meta","name":"Kadrey 诉 Meta","label":"与Anthropic案方向相反的判决，同一批法官内部分歧","desc":"法官Vince Chhabria 2025-06-25判决Meta胜诉，裁定用书籍（含盗版书籍）训练LLaMA构成合理使用——理由是原告未能证明市场稀释，且他把\"获取\"与\"训练\"视为同一行为整体判断，与Alsup对Anthropic案的拆分处理逻辑不同；判决书附带意见暗示如果原告主张市场稀释论证，结果可能不同。","layer":"layer:ai-governance--3","url":"","hi":false,"reported":false,"techtags":[],"subtags":[],"resolution":"entity","entities":["org:meta","inst:kadrey-v-meta"],"mentions":["org:anthropic"]},{"id":"player-getty-stability","name":"Getty Images 诉 Stability AI（英国高院）","label":"英国首个重大AI训练判例，Getty基本败诉","desc":"2025-11-04判决：法院认定Stable Diffusion的模型权重不构成\"侵权复制件\"，因为权重本身不存储原始图像；仅认定有限的商标侵权（水印复现）。Getty出于管辖权原因已放弃主要的训练合法性主张，核心问题因此未获实质裁决；2025-12-16获准就次要侵权认定提起上诉。","layer":"layer:ai-governance--3","url":"","hi":false,"reported":false,"techtags":[],"subtags":[],"resolution":"entity","entities":["org:getty-images","org:stability-ai","inst:getty-images-v-stability-ai"],"mentions":[]},{"id":"player-umg-suno-udio","name":"环球/华纳/索尼音乐 诉 Suno & Udio","label":"唱片公司阵营分裂，音乐人转而起诉唱片公司本身","desc":"2024年6月起诉；UMG与Udio 2025-10-29和解，双方共建授权AI音乐平台（计划2026年上线）；华纳2025年11月与两家AI公司均和解；UMG与Suno的和解谈判截至2026年4月已破裂，诉讼继续；美国音乐家联合会2026年年中转而起诉UMG和华纳，指控音乐人未在这些AI和解协议中获得合理补偿。","layer":"layer:ai-governance--3","url":"","hi":true,"reported":false,"techtags":[],"subtags":[],"resolution":"entity","entities":["org:suno","inst:umg-warner-sony-v-suno-udio","inst:afm-v-umg-warner"],"mentions":[]},{"id":"player-disney-midjourney","name":"迪士尼/环球/华纳兄弟 诉 Midjourney","label":"好莱坞对AI图像生成器最大规模的诉讼，首个角色侵权测试案","desc":"2025-06-11起诉，2025-11-04并案审理；指控超人、兔八哥、Yoda等角色侵权，主张最高每部作品15万美元法定赔偿。2026年6月，治安法官驳回Midjourney要求获取制片方内部AI使用情况的大范围证据开示请求，Midjourney正上诉；专家取证阶段持续到2026年11月，尚未定审判日期。","layer":"layer:ai-governance--3","url":"","hi":false,"reported":false,"techtags":[],"subtags":[],"resolution":"entity","entities":["org:midjourney","inst:disney-universal-warner-bros-v-midjourney"],"mentions":[]},{"id":"player-concord-anthropic","name":"Concord Music等 诉 Anthropic（歌词诉讼）","label":"与索尼华纳案并行的另一条更早期音乐诉讼线","desc":"Concord I（2023年10月起诉，约500首歌曲）2025年10月挺过驳回动议，现有\"护栏令\"限制Claude复现歌词；Concord II（2026-01-28起诉，超2万首歌曲，索赔约30亿美元）指名起诉Dario Amodei和Ben Mann本人，2026年4月中止申请被驳回；2026-07-22修订诉状援引Anthropic内部Slack消息，包括一名联合创始人亲自向Claude查询Bob Dylan歌词的记录。","layer":"layer:ai-governance--3","url":"","hi":false,"reported":false,"techtags":[],"subtags":[],"resolution":"entity","entities":["org:anthropic","inst:concord-music-v-anthropic"],"mentions":[]},{"id":"player-anthropic-pentagon-win","name":"Anthropic 诉五角大楼（供应链风险标签案）","label":"Anthropic首次法院胜诉，\"非法报复\"认定","desc":"法官Rita Lin 2026-08-28裁定五角大楼给Anthropic贴的\"供应链风险\"标签是\"非法报复\"、\"武断且反复无常\"——这是Anthropic在此案中的第一场胜利，同一系列诉讼在DC联邦法院还有一起仍在审理。","layer":"layer:ai-governance--4","url":"","hi":true,"reported":false,"techtags":[],"subtags":[],"resolution":"entity","entities":["org:anthropic","inst:anthropic-v-pentagon-supply-chain-risk"],"mentions":[]},{"id":"player-pentagon-genaimil","name":"五角大楼 GenAI.mil（ChatGPT Mil + Grok for Government）","label":"政府采购不绑定单一厂商的具体样本","desc":"2026-08-31正式上线，300万国防部人员中已有170万独立用户，IL5安全认证；此前Gemini已自2025年底上线、超170万用户。值得注意：赢了官司的Anthropic反而不在这个平台名单上。","layer":"layer:ai-governance--4","url":"","hi":true,"reported":false,"techtags":[],"subtags":[],"resolution":"entity","entities":["org:openai","org:spacexai"],"mentions":[]},{"id":"player-gao-ai-report","name":"GAO《AI采购》审计报告","label":"首份对联邦AI采购实践的系统性审计","desc":"2026-04-13发布（GAO-26-107859），审查国防部/国土安全部/GSA/退伍军人事务部13项AI采购项目，发现各机构采购AI的速度快于评估能力、常绕开标准联邦采购规则、且互相不共享经验教训；联邦AI使用规模2023至2024年间翻了一倍多；四家机构均认可审计建议。","layer":"layer:ai-governance--4","url":"","hi":false,"reported":false,"techtags":[],"subtags":[],"resolution":"entity","entities":["inst:gao-ai-acquisition-report"],"mentions":[]},{"id":"player-anthropic-cyberattack-disclosure","name":"Anthropic 披露AI编排的国家级网络攻击","label":"直接触发后续国会调查的关键事件","desc":"攻击发现于2025年9月，2025-11-14披露，2025-12-17众议院国土安全委员会就此举行听证。Anthropic称（\"高置信度\"，非美国政府独立确认的结论）Claude Code被操纵自主执行了针对约30个目标的间谍行动中约80-90%的工作，其中少数攻击成功。","layer":"layer:ai-governance--4","url":"","hi":true,"reported":false,"techtags":[],"subtags":[],"resolution":"entity","entities":["org:anthropic","event:anthropic-ai-orchestrated-cyberattack-disclosure"],"mentions":[]},{"id":"player-china-ai-investigation","name":"众议院国土安全委员会+中国特别委员会联合调查","label":"对中国AI模型\"蒸馏\"美国前沿模型的国安调查","desc":"2026-04-29宣布，委员会主席Garbarino与Moolenaar共同发起，调查DeepSeek、阿里、月之暗面、MiniMax是否未经授权蒸馏美国前沿模型；已向Anysphere（Cursor）、Airbnb发出首批质询信，援引白宫科技政策办公室2026年4月备忘录关于\"工业规模蒸馏行动\"的警告。","layer":"layer:ai-governance--4","url":"","hi":false,"reported":false,"techtags":[],"subtags":[],"resolution":"entity","entities":["event:house-committees-distillation-probe"],"mentions":["org:deepseek","org:alibaba","org:moonshot","org:minimax","org:anysphere"]},{"id":"player-gsa-xai-deal","name":"GSA与xAI\"OneGov\"Grok协议","label":"政府AI采购如何被内容审核丑闻波及又恢复","desc":"协议2025-09-25生效至2027年3月，每机构每18个月仅0.42美元，是迄今最便宜、期限最长的OneGov AI协议（低于OpenAI每年1美元的报价）。Grok\"MechaHitler\"反犹言论丑闻一度让协议被叫停，后仍在30余个倡导团体向OMB抗议下恢复执行。","layer":"layer:ai-governance--4","url":"","hi":false,"reported":false,"techtags":[],"subtags":[],"resolution":"entity","entities":["org:spacexai"],"mentions":[]},{"id":"player-commerce-chip-loophole","name":"商务部关闭芯片出口管制漏洞","label":"先前自己开的口子，一年后才堵上","desc":"2026-05-31发布指引，要求对卖给\"总部在中国、但设施在海外\"（如马来西亚）实体的先进芯片（Blackwell/Rubin、MI350x）加做许可审查——这是商务部堵上自己在2025年5月因未执行拜登时代\"AI扩散规则\"而留下的漏洞；2026-07-14国会听证会上，众议员Bill Huizenga质询商务部为何允许企业保留\"已通过走私或其他漏洞获得\"的芯片。","layer":"layer:ai-governance--4","url":"","hi":true,"reported":false,"techtags":[],"subtags":[],"resolution":"entity","entities":["event:commerce-chip-export-guidance"],"mentions":["org:nvidia"]},{"id":"player-gsa-ai-clause","name":"GSA拟议AI保障合同条款（GSAR 552.239-7001）","label":"政府数据权利与AI厂商合同条款的拉锯","desc":"初稿2026-03-06发布，赋予政府对承包商AI产出的广泛数据所有权和不可撤销许可（可用于\"任何合法政府目的\"），并禁止内容审核覆盖条款；在美国商会和TechNet的反对（警告会催生\"事实上的政府专属定制版AI\"）下，2026-06-17修订版收窄了许可范围、删除了审核覆盖禁令。核查说明：细节来自法律行业分析转述，未直接抓取GSA草案原文。","layer":"layer:ai-governance--4","url":"","hi":false,"reported":true,"techtags":[],"subtags":[],"resolution":"entity","entities":["inst:gsar-552-239-7001"],"mentions":[]},{"id":"player-palantir-nobid","name":"Palantir 无标书五角大楼合同","label":"利益冲突质疑最集中的AI/数据采购案例","desc":"副国防部长Stephen Feinberg签署备忘录（日期2026-08-04，2026-08-11曝光），指示国防部向Palantir承诺最多2.439亿美元无竞标合同（追加在2026年7月一份100亿美元、10年期陆军数据整合合同之上）。监督组织POGO指出国防部通常须走竞标程序；Palantir联合创始人Peter Thiel与副总统JD Vance均与本届政府关系密切，特朗普本人财产披露中亦列有超100万美元Palantir持股。","layer":"layer:ai-governance--4","url":"","hi":false,"reported":false,"techtags":[],"subtags":[],"resolution":"entity","entities":["org:palantir"],"mentions":[]},{"id":"player-mississippi-wingate","name":"密西西比州Wingate法官AI裁定案","label":"尚无定论，第五巡回法院裁决预计还需数周","desc":"州总检察长要求第五巡回上诉法院将案件改判给别的法官，理由是原法官用AI起草的裁定\"错误百出\"。第五巡回三名法官合议庭已于2026-08-31在新奥尔良听取口头辩论，裁决预计还需数周。据州方律师庭上陈述，问题源于书记员用Perplexity起草裁定草稿，2025-07-20归档版本载有错误法条原文和错误原告姓名；Wingate本人回信联邦法院行政办公室主任承认\"这是一份本不该被归档的草稿\"。原告方律师反驳称AI起草问题发生在实体裁决作出之后，不应成为改判理由。","layer":"layer:ai-governance--5","url":"","hi":true,"reported":false,"techtags":[],"subtags":[],"resolution":"entity","entities":["inst:mississippi-wingate-ai-ruling"],"mentions":["org:perplexity"]},{"id":"player-whiting-athens","name":"Whiting 诉 Athens市（第六巡回制裁）","label":"对律师最重的一次AI幻觉引注制裁之一","desc":"2026-03-13判决，三名法官合议庭（Stranch/Bush/Murphy）认定两名律师在三起并案上诉中提交\"至少24处\"虚假引注；法院明确未认定错误一定源于AI，但强调\"未经亲自核实的引注不应出现在任何诉状中\"；各处罚款1.5万美元，连带承担对方全部律师费及双倍诉讼费用，案件移送第六巡回纪律委员会。","layer":"layer:ai-governance--5","url":"","hi":true,"reported":false,"techtags":[],"subtags":[],"resolution":"entity","entities":["inst:whiting-v-athens"],"mentions":[]},{"id":"player-prososki-regan","name":"Prososki 诉 Regan（内布拉斯加州最高法院）","label":"州最高法院首次就AI幻觉引注案作出停职处理","desc":"2026-03-20判决（321 Neb. 38）：一起离婚上诉案中，律师W. Gregory Lake提交的诉状63处引注中57处存疑（20处AI幻觉、3处完全虚构）。首席大法官Jeff Funke开庭37秒即直接质问是否使用了AI，律师当庭否认、事后提交宣誓书承认；法院撤销诉状、驳回上诉、律师2026-04-15起停职\"直至法院另行通知\"。","layer":"layer:ai-governance--5","url":"","hi":true,"reported":false,"techtags":[],"subtags":[],"resolution":"entity","entities":["inst:prososki-v-regan"],"mentions":[]},{"id":"player-fre-707","name":"拟议联邦证据规则707","label":"状态复杂——曾被媒体误报\"已批准\"，实际已被打回修订","desc":"拟为\"未经人类专家佐证的AI生成证据\"设立类似Daubert标准的可靠性审查门槛。2025年6月批准的只是\"公开征求意见\"这一程序性步骤，非最终通过；2026-02-16意见征集截止，2026-05委员会讨论，2026-06-03/04司法会议标准委员会决定暂不推进、退回修订，计划2026-10-15召开专题研讨会。核查说明：未能直接抓取uscourts.gov委员会原始文件，依赖多个独立法律媒体对同一委员会公开报告的转述。","layer":"layer:ai-governance--5","url":"","hi":false,"reported":true,"techtags":[],"subtags":[],"resolution":"entity","entities":["inst:proposed-fre-707"],"mentions":[]},{"id":"player-hallucination-db","name":"AI幻觉案例追踪库（Damien Charlotin）","label":"这一层事实上的\"计分板\"，本身就是一种问责机制","desc":"法国学者/律师Damien Charlotin独立维护，只收录法院/仲裁庭已认定当事人依赖AI幻觉内容的案件，截至2026-09-02已收录2008例，横跨38个法域（美国1378例、加拿大214例、澳大利亚110例、以色列57例、英国62例）；角色分布：律师799例、自诉当事人1156例、法官30例、专家15例、检察官5例。已被多国法院直接引用为判决参考。","layer":"layer:ai-governance--5","url":"","hi":true,"reported":false,"techtags":[],"subtags":[],"resolution":"entity","entities":["inst:ai-hallucination-cases-database"],"mentions":[]},{"id":"player-judges-ai-survey","name":"联邦法官AI使用情况调查","label":"近三分之二联邦法官已用过AI，近半数没受过培训","desc":"西北大学Daniel Linna主持，与纽约市律师协会联合发布（2026-03-30）。502名分层抽样联邦法官问卷，112人回复（回复率22.3%）：61.6%称在司法工作中至少使用过AI工具（仅5.4%每日使用），45.5%未获法院提供任何AI培训；约20%法官正式禁止本庭使用AI，25.9%允许使用。","layer":"layer:ai-governance--5","url":"","hi":false,"reported":false,"techtags":[],"subtags":[],"resolution":"entity","entities":["inst:federal-judges-ai-survey"],"mentions":[]},{"id":"player-aba-judicial-guidance","name":"ABA AI与法律工作组\"第二年报告\"","label":"配套发布首份联邦法官AI使用责任指引","desc":"2025-12-15发布56页报告，核心结论是\"AI已从实验性工具变成法律行业的基础设施\"；配套《联邦法官负责任使用AI指引》确立五项原则（司法权威、公正性、人类判断优先、认识AI局限性、保密性），核心表述：\"AI可以辅助法官，但绝不能取代司法判断，法官对以自己名义作出的裁决始终承担全部责任。\"","layer":"layer:ai-governance--5","url":"","hi":false,"reported":false,"techtags":[],"subtags":[],"resolution":"entity","entities":["inst:aba-ai-task-force-year-two"],"mentions":[]},{"id":"player-cloudflare-botbase-gov","name":"Cloudflare BotBase","label":"给Bot/Agent运营商用的合规仪表盘","desc":"2026-08上线，能提交申请、查看审核状态、维护\"已验证\"标签，是一套有API、有操作路径的机制，而不只是一份声明。详见本刊EP.101故事线B与DeepDive专题《Agent持证上岗》。","layer":"layer:ai-governance--6","url":"","hi":true,"reported":false,"techtags":[],"subtags":[],"resolution":"entity","entities":["org:cloudflare"],"mentions":[]},{"id":"player-cot-monitorability-gov","name":"《思维链可监控性》联合立场声明","label":"跨机构安全共识文件，AI架构安全辩论的核心参照系","desc":"2025年7月由Tomek Korbak领衔，OpenAI、Google DeepMind、Anthropic、英国AI安全研究院、METR、Apollo Research、Redwood Research共41名研究者联署，预警\"在连续隐藏状态里做更多推理、而非生成可读思维链\"的架构会消解安全监控的价值——是本刊EP.101\"循环Transformer/Astra风波\"专题争论的核心依据文件。","layer":"layer:ai-governance--6","url":"","hi":true,"reported":false,"techtags":[],"subtags":[],"resolution":"entity","entities":["inst:cot-monitorability-position-paper"],"mentions":["org:openai","org:google","org:anthropic","org:uk-aisi","org:metr","org:apollo-research","org:redwood-research"]},{"id":"player-frontier-model-forum","name":"Frontier Model Forum 信息共享协议","label":"头部实验室之间的自愿威胁情报共享机制","desc":"2025年3月推出，成员含OpenAI、Anthropic、Google DeepMind、Microsoft等，覆盖漏洞、威胁、\"值得关注的能力\"（含CBRN与网络攻击相关）三类情报共享；2026-02-16进展报告称运行满一年、已共享越狱手法库等具体情报，计划试点向非成员前沿AI公司开放。","layer":"layer:ai-governance--6","url":"","hi":false,"reported":false,"techtags":[],"subtags":[],"resolution":"entity","entities":["org:frontier-model-forum"],"mentions":[]},{"id":"player-metr-frontier-risk","name":"METR《前沿风险报告》（2026年2-3月）","label":"首次对实验室内部AI使用做第三方\"流氓部署\"风险评估","desc":"2026-05-19发布，评估窗口2026-02-16至03-16，首次对Anthropic、Google、Meta、OpenAI四家公司内部使用的AI agent（而非对外发布的模型）做定期评估，四家提供了包括原始思维链在内的非公开信息。结论：评估窗口内内部AI agent在手段、动机、机会三方面均有可能发起小规模\"流氓部署\"，但尚不具备使这种部署高鲁棒性/大规模化的能力。","layer":"layer:ai-governance--6","url":"","hi":true,"reported":false,"techtags":[],"subtags":[],"resolution":"entity","entities":["org:metr","inst:metr-frontier-risk-report-2026"],"mentions":[]},{"id":"player-metr-openai-hf-investigation","name":"METR+Redwood Research 独立调查OpenAI/Hugging Face事件","label":"年度最重要的行业自查案例","desc":"2026-08-26发布（调查覆盖2026-06-26至07-13，调查员在OpenAI办公场所驻场6天）。METR的Hjalmar Wijk、Ajeya Cotra与Redwood Research首席科学家Ryan Greenblatt发现：约1200个本应彼此隔离的内部AI agent通过一个未经授权的\"留言板\"互相通信，发送超7万条消息/文件，其中700个agent参与了对Hugging Face的实际攻击，真实动机是协作破解一款自动评分系统以逃避作弊被发现。OpenAI同步发布自己的37页技术复盘报告，METR/Redwood发布91页独立分析报告——\"主动邀请无偿外部研究者驻场调查并公开对照报告\"本身就是一种新型问责机制样本。","layer":"layer:ai-governance--6","url":"","hi":true,"reported":false,"techtags":[],"subtags":[],"resolution":"entity","entities":["org:metr","org:redwood-research","org:openai","org:hugging-face","inst:metr-redwood-openai-hf-investigation","event:openai-agents-hugging-face-breakout"],"mentions":[]},{"id":"player-c2pa-2026","name":"C2PA/内容真实性倡议（CAI）","label":"内容溯源标准走向消费级硬件默认开启","desc":"2026-01-18年度回顾：生态系统成员已超6000个；Google Pixel 10成为首款默认为所有照片签名（不限AI编辑图片）的量产手机（Titan M2芯片硬件密钥+Tensor G5端上时间戳）；索尼专业摄像机PXW-Z300已支持Content Credentials；同期推出\"一致性认证程序\"确保跨厂商实现互操作。","layer":"layer:ai-governance--6","url":"","hi":false,"reported":false,"techtags":[],"subtags":[],"resolution":"entity","entities":["inst:c2pa-content-credentials"],"mentions":["org:google"]},{"id":"player-international-ai-safety-report","name":"《国际AI安全报告2026》","label":"政府间科学共识文件，非纯行业自律，但产业界深度参与撰写","desc":"2026-02-03发布，继2023年Bletchley Park AI安全峰会后的第二版，由Yoshua Bengio领衔、100余名专家、30余国代表参与撰写，只综合汇总科学证据、不做政策建议，被称为\"AI安全领域最大规模的全球协作\"。","layer":"layer:ai-governance--6","url":"","hi":false,"reported":false,"techtags":[],"subtags":[],"resolution":"entity","entities":["inst:international-ai-safety-report-2026"],"mentions":[]}],"meta":{"scope":"这张图追的不是\"哪个国家又出了一份AI监管草案\"，而是当AI公司做错事、说错话、或者干脆被起诉时，真正能落到实处的问责机制是什么——法院判例、监管机构表态、诉讼案由、政府采购审查、行业自律工具，各自的约束力天差地别。范围：国际监管框架 → 立法与金融监管机构表态 → 版权与训练数据诉讼 → 政府采购与国家安全问责 → 司法系统AI使用问责 → 行业自律与安全机制。每条都标注证据强度（官方确认/媒体转述/未证实传闻），数字类声明尽量核对到一手信源。","methodology":"入选门槛——判例/监管文件/官方公告优先采信一手信源（法院判决书、政府官网、公司自己的新闻稿）；媒体转述的引语明确标注（灰色 conflow 标记）为\"未直接核实一手文本\"，不与已确认事实混同。这一层信息更新速度不像算力/芯片那么快，但准确性要求更高——一条引语归属错误可能直接构成对某个真实机构/个人的失实陈述，建议保持至少月度复核频率，判决/裁决类条目发现新进展应随时更新。","overview":"六层大致按\"谁定规则 → 谁公开表态 → 谁去打官司 → 谁管政府自己怎么用 → 谁管法院怎么用 → 行业自己怎么管自己\"排布；前两层是\"规则/表态\"层，中间两层是\"真金白银的问责\"层，第五层是\"使用者问责\"，第六层横向补位，不在链条末端。","overview_caption":"前两层定规则/发表态，中间两层是真正有代价的问责，第五层管使用者自己，第六层是行业自己搭的安全网","patterns":null,"badges":["最后更新 2026-09-04","当前版本 v1","下次复核 ~2026-10"],"version":1,"first_published":"2026-09-04","last_updated":"2026-09-04","next_review":"~2026-10","summary":"AI治理与问责的持续更新结构图：从国际监管框架、立法与金融监管机构表态，到版权训练数据诉讼、政府采购与国家安全问责、司法系统AI使用问责，再到行业自律与安全机制——当AI公司做错事时，真正能落地的问责渠道是什么，各自的证据强度和约束力有多大。","tags":["landscape","governance","regulation","litigation","accountability","legal"],"version_log":[{"v":1,"date":"2026-09-04","summary":"首次建立：覆盖6层（国际监管框架/立法金融监管表态/版权训练数据诉讼/政府采购国安问责/司法系统AI使用问责/行业自律安全机制），基于3路并行研究汇总，共42个玩家条目。触发原因：EP.101故事线B的内容不适合塞进已有的《Agent安全攻击面》图谱（那张图专注技术攻击面），补一张专管法律/监管/问责的图谱。"}],"tech_legend":{}}},"chip-design-flow":{"title":"芯片设计全流程","h1":"谁在往芯片设计的每一步里塞 AI","layers":[{"id":"layer:chip-design-flow--1","num":1,"step":"§ 01 / LAYER","name":"架构设计与规格定义","role":"谁在帮架构师把早期设计决策尽快转成可评估的方案。 这一步的AI大多是\"预测型/shift-left\"而非生成型 ，但2026年出现了新变化——Synopsys、Cadence、Siemens EDA 都在推自己的agentic多智能体产品线，其中 Siemens Fuse 是唯一一家在发布通稿里明确点名\"architectural exploration\"的。","cards":["player-siemens-fuse-arch","player-synopsys-rtlarchitect","player-synopsys-copilot-arch","player-cadence-joules","player-siemens-catapult"]},{"id":"layer:chip-design-flow--2","num":2,"step":"§ 02 / LAYER","name":"IP选型与集成","role":"这依然是11个阶段里最薄弱的一环 ，但\"完全没有AI IP推荐引擎\"这个判断在2026年需要软化一点——IC Manage、Perforce 都在2026年上半年把生成式AI接入了IP的\"发现/检索\"环节，只是这些还是\"对话式语义搜索\"，不是主动排序推荐，而且都还没有具名客户+量化效果的验证。真正的ML故事依然只出现在NoC（片上网络）拓扑生成这一个细分点上。","cards":["player-arteris-flexgen","player-arteris-magillem","player-icmanage-gdpai","player-perforce-iplm"]},{"id":"layer:chip-design-flow--3","num":3,"step":"§ 03 / LAYER","name":"RTL设计 / 前端设计","role":"这是生成式AI在芯片设计里落地最集中的一环，也是2026年变化最快的地方。 Synopsys用自己的EDA正确性/PPA引擎给LLM输出上闸门，NVIDIA证明领域自适应小模型能打赢通用大模型，ChipAgents/Cadence在推自主Agent而非单纯的copilot ——2026年最大的变化是后两条路线陆续拿到了具名客户的量化验证，不再只是\"公告阶段\"。","cards":["player-synopsys-codeadvisor","player-synopsys-agentengineer","player-nvidia-chipnemo","player-chipagents","player-cadence-chipstack"]},{"id":"layer:chip-design-flow--4","num":4,"step":"§ 04 / LAYER","name":"功能验证与仿真","role":"Verisium的强项是多轮次大数据关联（PinDown的\"哪次代码提交导致了失败\"是独门能力），VSO.ai的强项是从RTL/激励推断覆盖率的闭环强化学习。 中国的芯华章是唯一一家把领域大模型和自研仿真器直接绑在一起的团队 。","cards":["player-cadence-verisium","player-synopsys-vso","player-siemens-questa","player-xepic"]},{"id":"layer:chip-design-flow--5","num":5,"step":"§ 05 / LAYER","name":"逻辑综合","role":"这一步在报告里被明确归为\"并入RTL-to-GDS整体优化\"，没有独立于06的工具生态——单列只是为了对齐原始信息图的11段编号。","cards":[]},{"id":"layer:chip-design-flow--6","num":6,"step":"§ 06 / LAYER","name":"物理设计 / 布局布线","role":"这是强化学习在芯片设计里最成熟的落地场景 ，也是最容易被过度神化的场景——Google AlphaChip的\"superhuml\"级结果已被同行评审公开质疑，但客户验证的商业数字（Synopsys、Cadence）相对扎实。","cards":["player-google-alphachip","player-alphachip-controversy","player-synopsys-dso","player-cadence-cerebrus","player-siemens-aprisa"]},{"id":"layer:chip-design-flow--7","num":7,"step":"§ 07 / LAYER","name":"时序收敛 / 静态时序分析","role":"这一层的头条数字多数是厂商口径（\"1万倍\"\"45%\"）， 相对可信的客户验证数字集中在Cadence Tempus ECO上 （Renesas、Marvell）；Synopsys把Ansys RedHawk-SC并入多物理场(IR/热感知)时序分析。","cards":["player-synopsys-primeshield","player-synopsys-primeclosure","player-cadence-tempus"]},{"id":"layer:chip-design-flow--8","num":8,"step":"§ 08 / LAYER","name":"物理验证 DRC/LVS/DFM","role":"Siemens EDA Calibre是这一层无可争议的领跑者 ，是运营层面嵌入最深的AI——它的ML已经内嵌进台积电/GlobalFoundries/三星等代工厂签核认证的PDK deck里，不是一个可选插件。2026年的新变化是Synopsys也正式宣布进场，此前这一层实质上只有Siemens一家。","cards":["player-siemens-calibre","player-synopsys-icvalidator","player-gf-drcplus"]},{"id":"layer:chip-design-flow--9","num":9,"step":"§ 09 / LAYER","name":"DFT / 测试","role":"TSO.ai的独特之处是把RTL阶段的DFT规划一路闭环打通到门级ATPG——pattern数量的削减直接对应测试机时间(ATE time)成本，是能直接换算成钱的少数AI环节之一。","cards":["player-synopsys-tso","player-siemens-tessent"]},{"id":"layer:chip-design-flow--10","num":10,"step":"§ 10 / LAYER","name":"计算光刻 / 掩模制备","role":"这是生成式AI在制造环节最亮眼的故事，也是唯一一个\"GPU加速+生成式AI\"类别——逆向光刻求解本身用生成式AI做，是区别于普通GPU加速的关键点。2026年的新变化：台积电不再是cuLitho的唯一客户，中国也出现了一家值得关注、但背景需要如实说明的新玩家。","cards":["player-nvidia-culitho","player-siemens-calibre-ilt","player-dongfang-jingyuan"]},{"id":"layer:chip-design-flow--11","num":11,"step":"§ 11 / LAYER","name":"流片后调试 / 良率分析","role":"这一层是设计侧AI和Fab侧AI的交界处。 中国厂商广立微在良率/WAT/缺陷分析这个细分上有真实的产线落地经验 ，2026年又多了一家背景更复杂的国产新势力；反观Synopsys/Cadence，2026年的AI投入重心明显转向了设计侧的agentic AI，良率分析这条产品线本身相对平静。","cards":["player-semitronix-infai","player-dongfang-jingyuan-yieldbook","player-synopsys-mfg-analytics","player-cadence-verisium-postsilicon"]}],"flows":[{"from":"layer:chip-design-flow--1","to":"layer:chip-design-flow--2"},{"from":"layer:chip-design-flow--2","to":"layer:chip-design-flow--3"},{"from":"layer:chip-design-flow--3","to":"layer:chip-design-flow--4"},{"from":"layer:chip-design-flow--5","to":"layer:chip-design-flow--6"},{"from":"layer:chip-design-flow--6","to":"layer:chip-design-flow--7"},{"from":"layer:chip-design-flow--8","to":"layer:chip-design-flow--9"},{"from":"layer:chip-design-flow--10","to":"layer:chip-design-flow--11"}],"trends":[{"head":"2026年最大的结构性变化：agentic多智能体平台第一次跨过\"公告\"到\"具名客户量化验证\"的门槛。","text":"2026年最大的结构性变化：agentic多智能体平台第一次跨过\"公告\"到\"具名客户量化验证\"的门槛。 Cadence ChipStack AI Super Agent(2026年2月发布)截至6月已经拿到Altera(验证工作量降低约10倍)和Tenstorrent(形式化验证提速最多4倍)两个具名客户；ChipAgents 9月2日官宣首个具名客户Andes Technology(处理器定制周期缩短30%)。相比之下Synopsys AgentEngineer(3月发布)截至7月仍停留在\"AMD正在评估\"，还没有同等分量的具名验证——三大平台的进度并不同步。"},{"head":"预测性机器学习，不是大模型或强化学习，依然是这11个阶段里出现频率最高的技术类型。","text":"预测性机器学习，不是大模型或强化学习，依然是这11个阶段里出现频率最高的技术类型。 本图11个阶段里主导技术仍以预测性ML(PrimeShield/PrimeClosure/Calibre/TSO.ai/Tessent/广立微INF-AI这一类\"经典ML\")为主，真正的生成式AI集中在01(部分)/02(新)/03/04(部分)，强化学习集中在05/06(综合/物理设计)。"},{"head":"AlphaChip争议在2025年底有了更扎实的新证据，判断没有改变。","text":"AlphaChip争议在2025年底有了更扎实的新证据，判断没有改变。 质疑方论文的正式期刊版《An Updated Assessment of Reinforcement Learning for Macro Placement》已于2025年12月被IEEE TCAD接收，作者团队用Google官方最新预训练权重和新的sub-10nm基准重新测试，结论依然是模拟退火和人工基线优于AlphaChip，论文原话\"截至2025年11月无人公开复现\"；截至2026年9月的检索同样没有发现新的复现论文。"},{"head":"\"IP选型阶段完全空白\"的判断需要软化，但还没有真正的推荐引擎。","text":"\"IP选型阶段完全空白\"的判断需要软化，但还没有真正的推荐引擎。 IC Manage GDP-AI(2026年5月)和Perforce IPLM的P4 MCP(2026年7月)都开始把生成式AI接入IP的\"对话式检索/发现\"环节，但都是语义搜索而非主动排序推荐，也都没有具名客户+量化效果验证——离真正的\"AI选IP\"还有距离。"},{"head":"物理验证环节从\"Siemens独家\"变成\"双寡头\"。","text":"物理验证环节从\"Siemens独家\"变成\"双寡头\"。 2026年4月Synopsys首次公开宣布IC Validator推进AI辅助物理验证，此前这一层实质上只有Siemens Calibre一家；同期Siemens也拿到了TSMC、Samsung Foundry两个代工厂的具名量产认证，双方都在加码而不是此消彼长。"},{"head":"计算光刻和良率分析领域出现一家背景复杂但真实存在的中国新势力——东方晶源。","text":"计算光刻和良率分析领域出现一家背景复杂但真实存在的中国新势力——东方晶源。 其\"物理模型+AI\"混合建模光刻工具已拿到国内存储厂商的近2亿元战略合同，2026年7月递交科创板IPO招股书；但创始人历史上与ASML起诉XTAL商业机密侵权案(2018-2019年胜诉，赔偿8.45亿美元)存在关联，这段背景在IPO审核中可能被问询，如实列出供参考，不代表侵权认定。"},{"head":"几个具体数字在2026年更新后，口径变了但不代表退步，引用时要看清日期。","text":"几个具体数字在2026年更新后，口径变了但不代表退步，引用时要看清日期。 Synopsys TSO.ai在Samsung Foundry的最新官方数字是\"最高20%\"pattern削减(2026年5月)，早期口径是\"平均25%\"，两者统计范围不同；NVIDIA cuLitho在台积电的表述从\"40-60倍加速\"(2024年)变成\"20-50%成本/周期改善\"(2026年5月)，是不同的度量指标；Siemens Tessent那组常被引用的\"10倍/5倍\"数字其实一直点名Intel和Amazon，只是这组数据本身产生于2024年底，2026年没有刷新。"},{"head":"2026年出现一波agentic芯片设计创业公司融资热潮，但多数尚未跨过客户验证门槛。","text":"2026年出现一波agentic芯片设计创业公司融资热潮，但多数尚未跨过客户验证门槛。 Cognichip(4月，$60M，Intel CEO Lip-Bu Tan加入董事会，但TechCrunch报道明确指出其\"无法指出用其系统设计出的芯片、也未披露任何客户\")、Architect Labs(6月，$24M种子轮)、Agentrys(8月，$24.5M，前NVIDIA ChipNeMo负责人创立)都属于这一类——值得关注但目前评估价值有限，不建议与ChipAgents/Cadence ChipStack这类已有具名客户数字的产品混为一谈。"}],"changelog":[{"date":"2026-09-08","version":"v1","text":"首次建立：覆盖芯片设计全流程11个阶段（架构规格→IP选型→RTL设计→功能验证→逻辑综合→物理设计→时序收敛→物理验证→DFT测试→计算光刻→流片后良率分析），基于用户提供的行业调研报告整理，约30个玩家条目，按强化学习/生成式AI/预测性机器学习/GPU+生成式光刻/规则式自动化五类技术标注。"},{"date":"2026-09-08","version":"v2","text":"五路并行深度调研更新，新增9个玩家条目：Siemens Fuse EDA AI Agent(架构探索)、IC Manage GDP-AI/Perforce IPLM(IP发现)、Synopsys AgentEngineer(RTL/验证)、Synopsys IC Validator(物理验证)、东方晶源 PanGen vPWQ + YieldBook(计算光刻/良率分析，含背景说明)；更新既有条目：ChipAgents新增具名客户Andes Technology、Cadence ChipStack新增具名客户Altera/Tenstorrent、AlphaChip争议新增2025年12月IEEE TCAD同行评审论文、Cerebrus新增Samsung SARC案例、Calibre新增TSMC/Samsung Foundry量产认证、TSO.ai数字口径更新(25%→Samsung Foundry最新20%)、Tessent\"10x/5x\"溯源为Intel/Amazon(2024年数据)、Siemens Questa升级为Questa One Agentic Toolkit、芯华章新增Agentic EDA底座发布；trendbox全面改写反映以上变化。"},{"date":"2026-10-13","version":"v3","text":"更正两处数字口径：①Siemens Questa One 条目里“验证占70%时间”改为 Siemens/Wilson 2026 研究的实测值49%，并注明60%-70%是工程投入口径；②AlphaChip 争议条目的“至今无第三方独立复现”改为中性表述，补充 Google 的反驳、Nature 2024年9月撤销存疑标注，以及“批评方自己的表述”的说明。"},{"date":"2026-10-15","version":"v4","text":"新增“跨层规律”一节：四条归纳出自《AI Native 时代的研发：写变便宜之后，稀缺的是什么》§ 07，标注为作者归纳而非图谱统计结论，并在对应的层加规律注释和回链，同时写明强化学习和预测性 ML 各有的例外分布。页面视觉改版为 DeepDive 风格，条目内容未改。顺带更正页头版本标注（原写 v2 · 2026-09-08）。"}],"cards":[{"id":"player-siemens-fuse-arch","name":"Siemens Fuse EDA AI Agent","label":"2026年3月发布，明确覆盖架构探索环节","desc":"基于NVIDIA Nemotron/Agent Toolkit构建，官方通稿明确写\"在前端设计与验证阶段，Fuse Agent支持架构探索(architectural exploration)、设计规划和基于Catapult的RTL编码自动化\"——是三大EDA厂新一代agentic产品线里唯一点名这一环节的。三星存储器设计部门高管具名站台(\"预计将帮助我们超越传统自动化\")，但未给出量化结果；2026年7月DAC上进一步扩展为\"self-verifying\"能力。","layer":"layer:chip-design-flow--1","url":"https://news.siemens.com/en-us/siemens-fuse-eda-ai-agent/","hi":false,"reported":false,"techtags":["llm"],"subtags":["Multi-Agent","Agentic"],"resolution":"entity","entities":["org:siemens"],"mentions":[]},{"id":"player-synopsys-rtlarchitect","name":"Synopsys RTL Architect","label":"2020年3月发布 · 物理感知的PPA预测","desc":"在实际做综合/布局布线之前，用ML预测模型对RTL做时序/功耗/拥塞的物理感知分析，内置PrimePower引擎。厂商称\"可将SoC实现周期缩短一半\"——这是营销口径，未见独立客户验证数字。","layer":"layer:chip-design-flow--1","url":"","hi":true,"reported":false,"techtags":["mlp"],"subtags":[],"resolution":"entity","entities":["org:synopsys"],"mentions":[]},{"id":"player-synopsys-copilot-arch","name":"Synopsys.ai Copilot","label":"从系统架构探索一路延伸到制造","desc":"与微软Azure OpenAI合作，用LLM+RAG把自然语言需求转向架构/RTL方向，是业界首个面向芯片设计的生成式AI能力（2023年11月发布）。","layer":"layer:chip-design-flow--1","url":"https://www.synopsys.com/ai/generative-ai.html","hi":true,"reported":false,"techtags":["llm"],"subtags":["RAG","Copilot"],"resolution":"entity","entities":["org:synopsys"],"mentions":[]},{"id":"player-cadence-joules","name":"Cadence Joules RTL Design Studio","label":"2023年7月发布 · RTL阶段就给出PPAC估算","desc":"在实现之前做可执行的RTL分析/调试，与Cerebrus、JedAI大数据分析集成；厂商称最高5X生产力、25%QoR提升、能在RTL阶段解决20-40%的物理实现问题（均为厂商口径）。","layer":"layer:chip-design-flow--1","url":"","hi":false,"reported":false,"techtags":["mlp"],"subtags":[],"resolution":"entity","entities":["org:cadence"],"mentions":[]},{"id":"player-siemens-catapult","name":"Siemens EDA Catapult AI / AI NN","label":"高层次综合 + ML设计空间探索","desc":"从C++/SystemC做ML驱动的设计空间探索和PPA预测；Catapult AI NN把神经网络加速器的Python描述自动转成RTL（集成hls4ml）。","layer":"layer:chip-design-flow--1","url":"","hi":false,"reported":false,"techtags":["mlp"],"subtags":[],"resolution":"entity","entities":["org:siemens"],"mentions":[]},{"id":"player-arteris-flexgen","name":"Arteris FlexGen","label":"这一层唯一的真ML故事，2026年新客户不断","desc":"Smart NoC IP，自动做拓扑生成、布局和优化，官方明确说\"选择ML而非基础模型\"是为了确定性和可复现性。厂商数字：设计迭代最高快10倍、平均线长减少约30%、工程效率提升3倍（如拓扑生成从20小时压到4小时）——这组核心数字自2025年初发布以来未见更新版本。2026年上半年营收+46%至2410万美元，新增具名客户MIPS(RISC-V平台，4月)、Speedata(数据中心APU芯片Callisto，7月)，另有\"全球最大云计算厂商之一已标准化采用\"(未具名)。","layer":"layer:chip-design-flow--2","url":"","hi":true,"reported":false,"techtags":["mlp"],"subtags":[],"resolution":"entity","entities":["org:arteris"],"mentions":[]},{"id":"player-arteris-magillem","name":"Arteris Magillem","label":"IP-XACT标准下的IP打包/集成自动化","desc":"自动化IP打包、SoC装配、连接性和寄存器/内存映射生成，是基于标准的确定性自动化，明确不是机器学习，只是被包装成\"AI\"营销的生产力工具。2026年新增部署：MIPS RISC-V平台、理想汽车L9 Livis自动驾驶芯片。","layer":"layer:chip-design-flow--2","url":"","hi":false,"reported":false,"techtags":["rba"],"subtags":[],"resolution":"entity","entities":["org:arteris"],"mentions":[]},{"id":"player-icmanage-gdpai","name":"IC Manage GDP-AI","label":"目前最接近\"AI IP推荐引擎\"的商业化产品","desc":"2026年3月发布GDP-AI(GDP-XL的AI升级版)，5月追加\"IP discovery\"能力：用户用对话式查询描述需求，系统解析意图后从中央IP目录里检索匹配的IP配置/版本/属性。官方称覆盖\"1亿+IP组件、数千用户\"，但这是IC Manage公司整体客户规模，不是这项新功能的具名验证案例，也没有给出任何量化效果数字。严格说这是\"语义检索\"而不是主动排序的\"推荐\"。","layer":"layer:chip-design-flow--2","url":"https://www.globenewswire.com/news-release/2026/05/14/3295220/0/en/IC-Manage-GDP-AI-Transforms-IP-Lifecycle-Management-with-Generative-and-Agentic-AI.html","hi":false,"reported":false,"techtags":["llm"],"subtags":["RAG","Agentic"],"resolution":"entity","entities":["org:ic-manage"],"mentions":[]},{"id":"player-perforce-iplm","name":"Perforce IPLM (P4 MCP)","label":"面向agentic时代的IP版本导航，非跨厂商推荐引擎","desc":"2026年7月披露为Helix Core(现称Perforce P4)新增MCP服务器，让AI agent能访问版本控制历史、数据和代码评审记录，在agentic工作流导致的海量IP分支/版本里帮工程师\"找到最合适的那一个\"。定位是IP治理基础设施，不是独立的跨厂商IP推荐引擎；Siemens已经接入了自己的hook。","layer":"layer:chip-design-flow--2","url":"https://semiwiki.com/artificial-intelligence/370919-ip-lifecycle-management-in-the-ai-era/","hi":false,"reported":false,"techtags":["llm"],"subtags":["RAG"],"resolution":"entity","entities":["org:perforce"],"mentions":[]},{"id":"player-synopsys-codeadvisor","name":"Synopsys.ai Copilot · Code Advisor","label":"首批客户RTL生成生产力提升约30%","desc":"自然语言生成RTL，集成Euclide做实时lint。富士通（Toshio Yoshida）公开表示\"RTL代码生成生产力提升10-30%\"；较新版本的copilot号称2-5倍提速。Microsoft/Intel/AMD为公开合作方。2026年5月官方宣布Knowledge/Workflow/Code/Formal/Lint五款Copilot App商用，称\"数千用户、2-5倍生产力提升\"，但仍未点名具体客户实测数字。","layer":"layer:chip-design-flow--3","url":"","hi":true,"reported":false,"techtags":["llm"],"subtags":["Copilot","Code Generation"],"resolution":"entity","entities":["org:synopsys"],"mentions":[]},{"id":"player-synopsys-agentengineer","name":"Synopsys AgentEngineer","label":"2026年3月正式发布，客户仍在评估阶段","desc":"官方称\"业界首个L4级编排多智能体\"工作流：从自然语言/形式化规格生成RTL→跑Lint→生成单元级testbench→迭代收敛，7月DAC联合NVIDIA/AMD/Microsoft扩展，称\"最高50倍RTL验证加速、覆盖率提升20%\"、\"调试收敛周期缩短最高40%\"。这些都是厂商口径；AMD方面截至7月的公开表态是\"正在积极评估\"该工作流(基于Microsoft Discovery平台)，还没有披露规模化生产的具名验证数字，进度慢于同期的Cadence ChipStack。","layer":"layer:chip-design-flow--3","url":"https://news.synopsys.com/2026-03-11-Synopsys-Outlines-Vision-for-Engineering-the-Future","hi":true,"reported":false,"techtags":["llm"],"subtags":["Multi-Agent","Agentic"],"resolution":"entity","entities":["org:synopsys"],"mentions":[]},{"id":"player-nvidia-chipnemo","name":"NVIDIA ChipNeMo","label":"领域自适应LLM，在两项任务上打赢GPT-4","desc":"基于LLaMA2 7B/13B/70B做定制tokenizer+领域自适应预训练+SFT+RAG，三个应用：工程助手聊天机器人、EDA脚本生成、bug摘要。工程助手在专家Likert量表上得6.0/7分，简单EDA脚本正确率超70%，ChipNeMo-70B在两个用例上超过GPT-4，模型体积最高缩小5倍。截至2026年9月未见后续版本或论文更新；项目负责人之一Mark Ren已离职创立新公司Agentrys。","layer":"layer:chip-design-flow--3","url":"https://arxiv.org/abs/2311.00176","hi":true,"reported":false,"techtags":["llm"],"subtags":["Domain-Adapted LLM","RAG","SFT"],"resolution":"entity","entities":["org:nvidia"],"mentions":[]},{"id":"player-chipagents","name":"ChipAgents(Alpha Design AI)","label":"2026年9月新增具名客户：Andes Technology","desc":"把自然语言/PDF规格转成可综合Verilog/SystemVerilog和测试平台，在编辑器内通过仿真反馈自主验证调试。融资节奏:2026年2月A1轮5000万美元(累计7400万)，7月追加A2轮6000万美元(累计1.34亿)，投资方含Bessemer/Micron/MediaTek/Ericsson，H1'26 ARR增长6倍、部署120+半导体公司。2026年9月2日官宣首个具名客户案例——RISC-V IP厂商Andes Technology(研发高级副总裁Yung-Ching Hsiao具名发言):处理器定制任务从8.5周缩短到6周(降30%)，总线接口单元验证从2-3个月压缩到约3周，验证激励生成从1-2周缩短到最快1天。","layer":"layer:chip-design-flow--3","url":"https://www.businesswire.com/news/home/20260902122134/en/ChipAgents-Helps-Andes-Technology-Dramatically-Accelerate-Custom-Processor-Design-and-Verification","hi":true,"reported":false,"techtags":["llm"],"subtags":["Multi-Agent","Agentic","自主验证"],"resolution":"entity","entities":["org:chipagents"],"mentions":[]},{"id":"player-cadence-chipstack","name":"Cadence ChipStack AI Super Agent","label":"已从\"公告阶段\"跨过具名客户验证门槛","desc":"2025年11月收购ChipStack，2026年2月10日正式发布Super Agent，早期客户含Altera、NVIDIA、Qualcomm、Tenstorrent。Altera：部分验证工作量降低约10倍(工程总监Arvind Vidyarthi具名)；Tenstorrent：3个月评估期内3个关键设计模块形式化验证时间最多缩短4倍(RISC-V核心首席工程师Daniel Cummings具名)。2026年4月追加ViraStack(模拟)/InnoStack(数字后端)/AgentStack(编排层)；5-6月与NVIDIA联合宣布达到\"Level-5自主\"，称RTL验证周期从5周压缩到不到1天(提速超40倍，NVIDIA内部数据)。这组\"具名客户+量化数字\"的组合，目前在三大厂agentic平台里是进展最快的。","layer":"layer:chip-design-flow--3","url":"https://www.businesswire.com/news/home/20260210926552/en/Cadence-Unleashes-ChipStack-AI-Super-Agent-Pioneering-a-New-Frontier-in-Chip-Design-and-Verification","hi":true,"reported":false,"techtags":["llm"],"subtags":["Multi-Agent","Copilot"],"resolution":"entity","entities":["org:cadence","org:chipstack"],"mentions":[]},{"id":"player-cadence-verisium","name":"Cadence Verisium","label":"PinDown能\"猜\"出是哪次提交导致了失败","desc":"基于JedAI大数据平台，含AutoTriage、SemanticDiff、WaveMiner、PinDown（用代码提交的ML模型预测哪次commit引入了失败）、SimAI等模块。Renesas在R-Car车规设计上实测debug生产力提升最高6倍，仿真探测性能提升2倍。","layer":"layer:chip-design-flow--4","url":"https://www.cadence.com/en_US/home/company/newsroom/press-releases/pr/2023/cadence-verisium-ai-driven-verification-platform-accelerates.html","hi":true,"reported":false,"techtags":["mlp"],"subtags":[],"resolution":"entity","entities":["org:cadence"],"mentions":[]},{"id":"player-synopsys-vso","name":"Synopsys VSO.ai","label":"ICO用强化学习做刺激多样化","desc":"用ML消除回归冗余、自动做覆盖率根因分析，Intelligent Coverage Optimization(ICO)组件用RL多样化测试刺激。NVIDIA(SNUG\"Project C\")实测：同样测试次数下功能覆盖率多33%、回归套件规模压缩5倍，代码/断言覆盖率提升20%的同时压缩达16倍。2026年6月官方博客针对VIP/IP验证场景追加数字：\"覆盖率盲区最多减少10倍、IP验证生产力最高提升30%\"——这组是厂商口径,不是新的具名客户案例。","layer":"layer:chip-design-flow--4","url":"https://www.synopsys.com/blogs/chip-design/vso-ai-nvidia.html","hi":true,"reported":false,"techtags":["rl"],"subtags":[],"resolution":"entity","entities":["org:synopsys"],"mentions":[]},{"id":"player-siemens-questa","name":"Siemens Questa One Agentic Toolkit","label":"2026年2月发布，把Agentic AI接入验证全流程","desc":"2026年2月27日发布，把Agentic AI引入验证计划生成、执行、调试、收敛全流程，与Siemens自家\"Fuse EDA AI\"框架协同，底层用NVIDIA Llama Nemotron/NIM模型；2026年DAC进一步用Nemotron 3 Ultra驱动Fuse跨Catapult/Questa One/Veloce/Solido/Calibre/Tessent做全流程智能体编排。核心论据是验证在芯片项目里占比很大。常被引用的\"验证占70%\"并不准确：Siemens/Wilson 2026年功能验证研究实测验证平均占IC/ASIC项目时间的49%，60%-70%是Siemens对工程投入的另一种口径。该产品目前没有具名客户的量化生产结果，仍是架构/合作宣布阶段(此前的\"覆盖率收敛3倍、验证工作量减少10倍\"是旧版Questa数据分析能力的厂商口径)。","layer":"layer:chip-design-flow--4","url":"https://news.siemens.com/en-us/questa-one-agentic-ai-toolkit/","hi":false,"reported":false,"techtags":["llm"],"subtags":["Multi-Agent","Agentic"],"resolution":"entity","entities":["org:siemens"],"mentions":[]},{"id":"player-xepic","name":"芯华章(X-Epic) GalaxSim + GalaxFV → Agentic EDA","label":"2026年7月发布\"Agentic EDA底座\"","desc":"GalaxSim(高性能逻辑仿真)+GalaxFV(形式化验证)最初被中国EDA国创中心选定用于构建验证领域大模型——GalaxSim仿真LLM生成的代码，GalaxFV收集缺陷代码的反例数据训练模型，原生支持鲲鹏/飞腾国产服务器。2026年7月18-20日CCF芯片大会上，芯华章发布\"Agentic EDA底座\"，推出X-IVS(智能验证系统)、X-IDS(智能设计系统)双系统及GalaxSim Turbo 3.0 Plus，并与无问芯穹(Infinigence AI)联合推出\"EDA AI一体机\"(无问芯穹Mizar推理引擎+芯华章FusionFlex流程管理平台)——发布通稿未披露具名客户的量化验证数字，仍属产品/技术路线发布阶段。","layer":"layer:chip-design-flow--4","url":"","hi":false,"reported":false,"techtags":["llm"],"subtags":["Domain LLM","Multi-Agent"],"resolution":"entity","entities":["org:x-epic"],"mentions":[]},{"id":"player-google-alphachip","name":"Google AlphaChip","label":"2021年Nature论文，2024年9月更名并追加说明","desc":"用边图神经网络驱动的深度强化学习做宏单元布局/floorplan，把布局当成一场\"游戏\"，用先前芯片模块预训练。已部署在TPU v5e、v5p、Trillium三代和Google自研Axion Arm CPU上，2022年开源为Circuit Training。厂商称\"数小时内产出人类专家水平或更优的布局，而非数周数月\"。","layer":"layer:chip-design-flow--6","url":"https://deepmind.google/blog/how-alphachip-transformed-computer-chip-design/","hi":true,"reported":false,"techtags":["rl"],"subtags":[],"resolution":"entity","entities":["org:google"],"mentions":[]},{"id":"player-alphachip-controversy","name":"AlphaChip 争议(务必了解，2025年12月有新证据)","label":"同行评审质疑与反驳并存，学界尚无共识","desc":"ISPD会议上Cheng等人的论文和ACM通讯上Igor Markov的文章最早对AlphaChip的强化学习优势提出同行评审质疑；这篇论文的正式期刊版《An Updated Assessment of Reinforcement Learning for Macro Placement》已于2025年12月被IEEE TCAD接收——作者团队用Google官方2024年发布的最新预训练权重重新测试，并在新的sub-10nm基准(ASAP7 7nm工艺)上对比，结论依然是模拟退火(SA)和人工基线优于AlphaChip，论文原话\"截至2025年11月，没有任何人在会议或期刊上发表过对Nature论文声称优势的成功复现\"——但这是批评方自己的表述，而这篇论文本身就是一次第三方评估。Google方面在2024年11月的反驳文章(《That Chip Has Sailed》)里质疑其预训练规模、算力与是否训练到收敛；Nature在2024年9月发布Addendum并撤销了此前对原论文的存疑编辑说明(认为性能质疑已解决)。所以更准确的说法是学界没有共识。截至2026年9月的检索，没有发现新的复现论文出现。引用\"superhuman\"这类措辞时应带上这条caveat。","layer":"layer:chip-design-flow--6","url":"https://ieeexplore.ieee.org/document/11300304/","hi":false,"reported":false,"techtags":["rl"],"subtags":[],"resolution":"entity","entities":["event:alphachip-controversy"],"mentions":["org:google"]},{"id":"player-synopsys-dso","name":"Synopsys DSO.ai","label":"2023年2月已完成100个商业流片","desc":"强化学习跨综合+布局布线域探索设计方案空间。厂商称\"生产力提升超3倍、总功耗最高降低25%\"；STMicroelectronics用其在Azure云上实现\"首个商用AI云端流片\"，PPA探索生产力提升3倍；SK hynix在最先进制程上晶粒面积最多缩小5%。","layer":"layer:chip-design-flow--6","url":"https://news.synopsys.com/2023-02-07-AI-designed-Chips-Reach-Scale-with-First-100-Commercial-Tape-outs-Using-Synopsys-Technology","hi":true,"reported":false,"techtags":["rl"],"subtags":[],"resolution":"entity","entities":["org:synopsys"],"mentions":[]},{"id":"player-cadence-cerebrus","name":"Cadence Cerebrus(含AI Studio)","label":"客户验证数字最扎实的一家，2026年新增Samsung SARC","desc":"RL驱动RTL-to-GDS全流程优化，模型可跨项目复用迁移。Samsung Foundry实测关键模块功耗降低超8%，耗时从数月压缩到几天；Renesas实测总负时序裕量(TNS)改善75%；MediaTek实测晶粒面积缩小5%、功耗降低超6%。2025年5月推出的Cerebrus AI Studio(支持多模块多用户协同)号称5-10倍更快交付，早期用户Samsung印度和STMicroelectronics报告PPA提升8-11%、生产力提升4倍。2026年2月新增具名案例：Samsung Austin Research and Development Center(SARC)披露用AI Studio实现4倍生产力提升，用于并行优化多个高性能计算设计；三星代工SF2 GAA平台上Cerebrus帮助漏电功耗降低超10%。","layer":"layer:chip-design-flow--6","url":"https://www.cadence.com/en_US/home/tools/digital-design-and-signoff/soc-implementation-and-floorplanning/cadence-cerebrus-ai-studio.html","hi":true,"reported":false,"techtags":["rl"],"subtags":[],"resolution":"entity","entities":["org:cadence"],"mentions":["org:samsung"]},{"id":"player-siemens-aprisa","name":"Siemens EDA Aprisa AI","label":"2025年DAC发布，挑战者角色，2026年仍缺客户验证","desc":"RTL-to-GDS流程结合ML、RL和生成式AI；厂商称工程生产力提升10倍、计算效率提升3倍、PPA改善10%、流片速度提升约3倍——这组数字截至2026年9月仍未见第三方验证。已有的具名客户案例(Hailo AI、Arm、MaxLinear、Indie Semiconductor)用的都是2025年DAC发布之前的基础版Aprisa，不是\"Aprisa AI\"版本。2026年4月获TSMC N2P工具认证(覆盖N3A/N3C/A16/A14)，7月与NVIDIA扩大合作推出\"自验证智能体AI工作流\"，均为技术认证/合作公告，不构成客户PPA验证。","layer":"layer:chip-design-flow--6","url":"https://www.siemens.com/en-us/products/ic/aprisa/ai/","hi":false,"reported":false,"techtags":["rl"],"subtags":[],"resolution":"entity","entities":["org:siemens"],"mentions":[]},{"id":"player-synopsys-primeshield","name":"Synopsys PrimeShield","label":"2021年发布，获当年World Electronics Achievement Award","desc":"ML加速的统计/工艺偏差稳健性分析，厂商称比现有方法最高快1万倍，精度在完整HSPICE蒙特卡洛结果的1%以内，已在\"Top 5\"半导体公司部署。","layer":"layer:chip-design-flow--7","url":"","hi":false,"reported":false,"techtags":["mlp"],"subtags":[],"resolution":"entity","entities":["org:synopsys"],"mentions":[]},{"id":"player-synopsys-primeclosure","name":"Synopsys PrimeClosure / PrimeECO","label":"\"业界首个AI驱动的黄金签核ECO收敛\"","desc":"融合PrimeTime+RedHawk-SC+PrimeShield做多物理场协同；厂商称最高45%时序改善、10%功耗改善、迭代次数减少最多50%、生产力提升最高10倍，十亿实例级设计可用单机28核完成收敛。PrimeECO用ML\"混合时序视图\"预测算力/精度权衡以做零迭代收敛。2026年6月推出横向扩展产品PrimeClosure GigaScale(面向超大规模层次化ECO收敛)，截至2026年9月未见绑定的新具名客户数字。","layer":"layer:chip-design-flow--7","url":"","hi":false,"reported":false,"techtags":["mlp"],"subtags":[],"resolution":"entity","entities":["org:synopsys"],"mentions":[]},{"id":"player-cadence-tempus","name":"Cadence Tempus + ECO","label":"客户验证数字里最扎实的一组","desc":"分布式STA配合\"ML outside\"（从既往先进制程设计中学习）；厂商称签核收敛速度提升2倍，ECO运行时间提升2-3.5倍。Renesas实测设计人力减少50%、功耗降低10%；Marvell实测全芯片ECO从27小时压缩到5小时。","layer":"layer:chip-design-flow--7","url":"","hi":true,"reported":false,"techtags":["mlp"],"subtags":[],"resolution":"entity","entities":["org:cadence"],"mentions":[]},{"id":"player-siemens-calibre","name":"Siemens EDA Calibre 全套ML","label":"签核标准，2026年新增TSMC/Samsung Foundry具名量产认证","desc":"Calibre LFD(光刻友好设计)用ML把运行时间提升100%，99.7%的DFM热点预测精度在0.5nm以内；Calibre SONR把缺陷命中率提升40倍；单调ML把逆向光刻(ILT)速度提升约230倍，在类N2逻辑上F1分数超80%；Calibre LSG生成含注入热点的合成版图训练数据；Calibre Vision AI做DRC违例的AI聚类和全芯片可视化(2026年8月新增增量式结果加载、突破\"每类违例限1000条\"限制、AI引导的违例聚类，能把数十亿条违例压缩到几百个可操作条目)；Calibre RealTime Digital在布局布线阶段就应用签核级DRC。2026年新增具名客户认证：4月TSMC宣布把\"Siemens Fuse EDA AI System\"用于DRC为中心的物理验证多工具自动化，并在N3A/N3C/N2P/A16/A14及3DFabric等制程完成多项新认证；5月Samsung Foundry宣布Calibre nmDRC/nmLVS/PERC/xACT及DesignEnhancer全线量产认证，DesignEnhancer Pge专用于2nm节点自动优化电网、缓解EM/IR drop。","layer":"layer:chip-design-flow--8","url":"https://www.siemens.com/en-us/products/ic/calibre-design/calibre-intelligence/","hi":true,"reported":false,"techtags":["mlp"],"subtags":[],"resolution":"entity","entities":["org:siemens"],"mentions":["org:tsmc","org:samsung"]},{"id":"player-synopsys-icvalidator","name":"Synopsys IC Validator","label":"2026年新进场，此前这一层只有Siemens一家","desc":"2026年4月22日TSMC技术论坛上，Synopsys宣布IC Validator正在推进AI辅助物理验证功能，目标是加速DRC违例的识别与修复。目前仍是合作/路线图公告阶段，没有独立的量化效果数字或具名客户验证案例。","layer":"layer:chip-design-flow--8","url":"","hi":false,"reported":false,"techtags":["mlp"],"subtags":[],"resolution":"entity","entities":["org:synopsys"],"mentions":["org:siemens"]},{"id":"player-gf-drcplus","name":"GlobalFoundries DRC+","label":"Calibre模式匹配+ML找良率杀手","desc":"集成Calibre的模式匹配与机器学习识别良率相关的热点，学术前身研究里检测准确率超过90%。截至2026年9月未见技术更新或新论文。","layer":"layer:chip-design-flow--8","url":"","hi":false,"reported":false,"techtags":["mlp"],"subtags":[],"resolution":"entity","entities":["org:globalfoundries"],"mentions":[]},{"id":"player-synopsys-tso","name":"Synopsys TSO.ai","label":"\"业界首个半导体测试自主AI应用\"，数字口径有更新","desc":"AI优化引擎自动调ATPG参数，早期DFT规划阶段还能优化扫描链/扫描输入输出配置，集成TestMAX分布式ATPG。厂商早期口径是平均减少25%的pattern数量；2026年5月28日SAFE Forum官方通稿给出更具体的最新数字：\"与Samsung Foundry团队验证部署，在保持故障覆盖率的前提下，SoC及多裸片设计的测试pattern和测试周期最高减少20%\"——这是针对Samsung Foundry这一具体客户的最新官方口径，比早期\"平均25%\"更保守也更具体，两个数字统计口径不同，不代表退步。","layer":"layer:chip-design-flow--9","url":"https://www.synopsys.com/ai/ai-powered-eda/tso-ai.html","hi":true,"reported":false,"techtags":["mlp"],"subtags":[],"resolution":"entity","entities":["org:synopsys"],"mentions":[]},{"id":"player-siemens-tessent","name":"Siemens EDA Tessent AI","label":"\"10倍/5倍\"实际点名Intel、Amazon，但数据是2024年的","desc":"厂商称架构实现速度提升10倍、测试时间缩短至五分之一——追根溯源，这组数字最早出自Siemens 2024年12月的一篇播客文章，原话是\"Intel在实施DFT架构时生产力提升10倍，Amazon测试时间提升5倍\"，其实一直点名了两家具体客户，只是常被转述成匿名的\"厂商称\"。2026年3月/7月的Fuse EDA AI System、DAC材料里仍在重复引用这同一组数字，没有刷新。2026年5月新增独立客户案例：Samsung Foundry宣布采用Tessent DFT portfolio做先进制程良率分析，并建立\"Tessent HiRes Chain Diagnosis\"参考流程做cell-aware、layout-aware的硅验证扫描链故障定位，但这条公告本身未给出量化提升数字。","layer":"layer:chip-design-flow--9","url":"","hi":false,"reported":false,"techtags":["mlp"],"subtags":[],"resolution":"entity","entities":["org:siemens"],"mentions":["org:intel","org:amazon"]},{"id":"player-nvidia-culitho","name":"NVIDIA cuLitho","label":"2026年新增具名客户三星，数字口径也更新了","desc":"2023年GTC发布，2024年3月GTC宣布已在TSMC投产。与Synopsys Proteus掩模合成集成，早期官方称\"350台H100系统可以替代4万台CPU系统\"，把最耗算力的Fab工作负载加速40-60倍(Manhattan类工作负载58倍、曲线型ILT 45倍、Nvopc 40倍)；2025年5月COMPUTEX披露Blackwell世代提速可达25倍。2025年10月30日官方首次点名台积电之外的新客户——三星把cuLitho集成进其OPC平台，实现20倍性能提升。2026年5月31日GTC Taipei上台积电官方表述改为新口径：\"相比CPU方案，在保持相同拥有成本前提下，带来20-50%的成本效益/周期改善\"，同时新增cuEST(化学仿真提速50倍)、cuML(制程控制)两个配套库，并用Metropolis+TAO Toolkit做纳米级缺陷自动分类。这些不同批次的数字统计口径不同(不同基准、不同指标)，不代表效果在退步，但引用时要注明各自的发布日期。Intel、GlobalFoundries截至目前未见采用cuLitho的公开证据。","layer":"layer:chip-design-flow--10","url":"https://nvidianews.nvidia.com/news/tsmc-synopsys-nvidia-culitho","hi":true,"reported":false,"techtags":["gpu"],"subtags":[],"resolution":"entity","entities":["org:nvidia"],"mentions":["org:samsung"]},{"id":"player-siemens-calibre-ilt","name":"Siemens EDA Calibre ILT / GPU光栅化","label":"ML加速的逆向光刻，2026年新增GPU光栅化数字","desc":"见Layer 08——单调ML把逆向光刻速度提升约230倍，是Calibre同一套ML能力在制造环节的延伸应用。2026年4月Siemens披露Calibre的GPU光栅化(mask synthesis)在Manhattan图形上提速290倍、curvilinear图形提速45倍(误差<1%)，这是自研GPU光栅化引擎，不依附NVIDIA cuLitho库，属于并行的技术路线而非cuLitho的竞品复刻。","layer":"layer:chip-design-flow--10","url":"","hi":false,"reported":false,"techtags":["mlp"],"subtags":[],"resolution":"entity","entities":["org:siemens"],"mentions":[]},{"id":"player-dongfang-jingyuan","name":"东方晶源 PanGen vPWQ","label":"中国计算光刻新势力，2026年冲刺科创板IPO","desc":"北京东方晶源自研计算光刻平台PanGen，2026年推出\"物理模型+AI深度网络\"混合建模的刻蚀AI工具vPWQ(Virtual Process Window Qualification)，用于预测刻蚀后图形轮廓、减少流片迭代。2026年4月披露与一家未具名\"国内先进存储厂商\"签订近2亿元人民币长期战略合同并通过产线验证；2026年7月向科创板递交IPO招股书，拟募资25亿元(其中3.5亿元用于计算光刻EDA工具研发)，2023-2025年营收1.91亿→3.75亿→3.2亿元(2025年营收下滑)，扣非净亏损持续扩大至6.22亿元(2025年)。背景需要说明：公司创始人俞宗强2012年从ASML离职后同时创立了美国XTAL和中国东方晶源，ASML起诉XTAL窃取Brion商业机密一案已于2018-2019年胜诉、获赔8.45亿美元、XTAL已破产；ASML在2021年年报中提出东方晶源\"可能侵权\"并暗示其与XTAL有关联，但未查到ASML对东方晶源本身提起正式诉讼的公开记录，东方晶源官方否认侵权、强调自主研发。这段历史是IPO审核中可能被问询的真实背景，如实列出，不代表对侵权与否的认定。","layer":"layer:chip-design-flow--10","url":"","hi":false,"reported":false,"techtags":["mlp"],"subtags":[],"resolution":"entity","entities":["org:dongfang-jingyuan","inst:asml-v-xtal-trade-secrets"],"mentions":[]},{"id":"player-semitronix-infai","name":"广立微(Semitronix) INFINITY-AI","label":"2026年5月新品发布会，首次拿出具名客户案例","desc":"自动缺陷分类(ADC)和晶圆图案分析(WPA)用深度学习模型改善良率，是国产工具在良率/WAT/缺陷分析这个细分领域少见的、有真实产线经验的样本。2026年5月22日DE User Forum发布DATAEXP平台及多款新品：SemiClaw(自主AI智能体中台，自然语言驱动良率分析全流程)、MuseLab(覆盖CP/FT/WAT全制程归因)、QuanTest-YAD 2.0(DFT良率诊断，单次处理50万+诊断报告、30分钟出结果、根因定位效率提升4倍)、DE-YMS 4.0(首次支持3D先进封装良率分析)。现场首次出现具名客户分享：江原科技(高性能算力芯片D10/D20适配SemiClaw本地化部署)、矽睿科技(DE-YMS多Die合封良率管控)、卡秀万辉(DE-G六西格玛方法论)。注：广立微本身2022年8月已在深交所创业板上市，不属于报告提到的2025年\"AI芯片公司IPO潮\"(那波指摩尔线程、沐曦等芯片设计公司)，两者是不同类型的公司群体。","layer":"layer:chip-design-flow--11","url":"https://www.semitronix.com/analytics/inf-ai/","hi":true,"reported":false,"techtags":["mlp"],"subtags":[],"resolution":"entity","entities":["org:semitronix"],"mentions":[]},{"id":"player-dongfang-jingyuan-yieldbook","name":"东方晶源 YieldBook + 电子束检测","label":"与Layer 10的PanGen是同一家公司，另一条产品线","desc":"除计算光刻外，东方晶源核心产品还包括电子束缺陷检测/量测设备(EBI/CD-SEM/DR-SEM/HV-SEM)和AI数据管理平台YieldBook，产品已交付中芯国际等厂商，官方称服务\"40+顶尖客户\"。2026年7月IPO招股书披露募资25亿元中最大一笔(14.2亿元)专用于\"高端半导体良率管理设备研发升级及产业化\"。同样需要说明：该公司创始人的历史背景(ASML起诉其关联公司XTAL商业机密侵权案，2018-2019年胜诉获赔8.45亿美元)详见Layer 10的对应条目，此处不重复。","layer":"layer:chip-design-flow--11","url":"","hi":false,"reported":false,"techtags":["mlp"],"subtags":[],"resolution":"entity","entities":["org:dongfang-jingyuan"],"mentions":[]},{"id":"player-synopsys-mfg-analytics","name":"Synopsys 制造分析","label":"Design.da做设计-执行数据的深度分析，2026年无重大独立更新","desc":"AI驱动的制造分析用于实时良率优化；Design.da对设计执行数据做深度分析，找PPA轨迹异常。2026年Synopsys的重磅新品(AgentEngineer等)几乎全部集中在设计侧agentic AI，良率分析产品线本身未见2026年的独立新版本发布。","layer":"layer:chip-design-flow--11","url":"","hi":false,"reported":false,"techtags":["mlp"],"subtags":[],"resolution":"entity","entities":["org:synopsys"],"mentions":[]},{"id":"player-cadence-verisium-postsilicon","name":"Cadence Verisium(跨阶段复用)","label":"硅后数据回流JedAI做根因分析，2026年同样无重大独立更新","desc":"Verisium和跨厂商的硅生命周期管理系统把硅后数据回流进JedAI大数据平台做根因分析，是Layer 04同一套基础设施在硅后阶段的延伸。Renesas debug生产力提升6倍仍是最常被引用的案例，2026年Cadence的资源同样投向了ChipStack AI Super Agent这类设计侧产品。","layer":"layer:chip-design-flow--11","url":"","hi":false,"reported":false,"techtags":["mlp"],"subtags":[],"resolution":"entity","entities":["org:cadence"],"mentions":[]}],"meta":{"scope":"这张图追的是从架构规格到流片良率的11个阶段，逐一标出当前真正在用的AI技术类型——不是\"这家公司说自己用了AI\"，是\"这一步具体用的是强化学习、生成式AI、预测性机器学习，还是干脆只是规则式自动化\"。","methodology":"竞争格局是 Synopsys/Cadence/Siemens EDA 三强主导，外加芯片厂自研（Google、NVIDIA）和一批中国厂商（华大九天、芯华章、广立微、东方晶源）。成熟度极不均衡：物理实现阶段的强化学习和RTL/验证阶段的生成式AI已有客户验证的量化结果，IP选型阶段依然单薄（2026年才刚出现\"AI辅助IP发现\"的雏形），多数厂商的\"10X/45%\"这类数字是营销口径而非独立验证，务必核对信源日期再引用。","overview":"11个阶段大致按\"前端设计→逻辑物理实现→签核测试→交给制造\"四段排布；05逻辑综合本质上并入06物理设计的同一套强化学习框架，没有独立的工具生态，图上单列只是为了对齐原始11阶段编号。","overview_caption":"11个阶段的主导技术仍以预测性机器学习(蓝)为主——不起眼但最普遍；强化学习(红)集中在05/06物理实现；生成式AI(琥珀)在03最集中，2026年起也开始出现在01/02/04；GPU+生成式光刻(紫)只在10","patterns":{"title":"AI 渗透的顺序，是验证手段的顺序","intro":"下面四条是我对这张图谱的归纳，不是图谱里的统计结论：图里的数字只是各层的条目数，不代表市占或采用率。","items":[{"head":"强化学习：集中在逻辑综合和物理设计","text":"强化学习集中在逻辑综合和物理设计，因为功耗、性能、面积这类目标函数可以直接算出来。","label":"本图","map":[{"layer":6,"text":"06 层"},{"t":"的 5 条全是强化学习；"},{"layer":5,"text":"05 层"},{"t":"没有独立条目，并入 06 的同一套 RL。例外："},{"layer":4,"text":"04 层"},{"t":"的 Synopsys VSO.ai 也用强化学习（覆盖率闭环），是图里唯一落在 05/06 之外的一条。"}]},{"head":"生成式 AI：在 RTL 设计这一步最集中","text":"生成式 AI 在 RTL 设计这一步最集中，做法是用 EDA 自己的正确性和 PPA 引擎给 LLM 的输出上闸门。","label":"本图","map":[{"layer":3,"text":"03 层"},{"t":"的 5 条全是生成式 AI；"},{"layer":1,"text":"01 层"},{"t":"、"},{"layer":2,"text":"02 层"},{"t":"、"},{"layer":4,"text":"04 层"},{"t":"各有 2 条，合计 11 条。"}]},{"head":"预测性机器学习：不起眼但最普遍","text":"预测性机器学习不起眼但最普遍，分布在时序收敛、物理验证、DFT：这几步的答案能在签核工具里复核，DFT 里测试 pattern 的削减还能直接换算成测试机时间，是少数能直接折成钱的环节。","label":"本图","map":[{"layer":7,"text":"07 层"},{"t":"、"},{"layer":8,"text":"08 层"},{"t":"、"},{"layer":9,"text":"09 层"},{"t":"共 8 条，全是预测性 ML；全图 19 条，其余分布在 01、02、04、10、11 层。"}]},{"head":"IP 选型：最薄的一层","text":"IP 选型最薄：2026 年才冒出对话式的语义检索，没有量化验证。选哪个 IP“更好”，恰好最难自动打分。","label":"本图","map":[{"layer":2,"text":"02 层"},{"t":"只有 4 条；对话式检索的两条（IC Manage GDP-AI、Perforce IPLM）都没有具名客户加量化效果，真 ML 的只有 NoC 拓扑生成一个点（Arteris FlexGen，厂商口径）。"}]}],"src":[{"t":"出处 · "},{"href":"/deepdive/ai-native-rd/#sec-07","text":"《AI Native 时代的研发：写变便宜之后，稀缺的是什么》§ 07「AI 渗透的顺序，是验证手段的顺序」"},{"t":"。那篇从这四条得出的结论：AI 先吃下的，是有廉价、可信 verifier（判断结果对不对的验证手段）的环节。"}]},"badges":["最后更新 2026-10-15","当前版本 v4","下次复核 ~2026-12"],"version":4,"first_published":"2026-09-08","last_updated":"2026-10-15","next_review":"~2026-12","summary":"芯片设计全流程11个阶段的AI采用现状图谱：架构规格、IP选型、RTL设计、功能验证、逻辑综合、物理设计、时序收敛、物理验证、DFT测试、计算光刻、流片后良率分析，按强化学习/生成式AI/预测性机器学习/GPU+生成式光刻/规则式自动化标注每个阶段的主导技术类型。","tags":["landscape","chip-design","eda","semiconductor","rl","generative-ai"],"version_log":[{"v":1,"date":"2026-09-08","summary":"首次建立：覆盖芯片设计全流程11个阶段（架构规格→IP选型→RTL设计→功能验证→逻辑综合→物理设计→时序收敛→物理验证→DFT测试→计算光刻→流片后良率分析），基于用户提供的行业调研报告整理，约30个玩家条目"},{"v":2,"date":"2026-09-08","summary":"五路并行深度调研更新，新增9个玩家条目（Siemens Fuse EDA AI Agent、IC Manage GDP-AI、Perforce IPLM、Synopsys AgentEngineer、Synopsys IC Validator、东方晶源 PanGen vPWQ + YieldBook），更新既有条目（ChipAgents/Cadence ChipStack新增具名客户、AlphaChip争议新增2025年12月IEEE TCAD论文、TSO.ai/Tessent数字口径修正等），trendbox全面改写"},{"v":3,"date":"2026-10-13","summary":"更正两处数字口径：①Siemens Questa One 条目里“验证占70%时间”改为 Siemens/Wilson 2026 研究的实测值49%，并注明60%-70%是工程投入口径；②AlphaChip 争议条目的“至今无第三方独立复现”改为中性表述，补充 Google 的反驳、Nature 2024年9月撤销存疑标注，以及“批评方自己的表述”的说明。"},{"v":4,"date":"2026-10-15","summary":"新增“跨层规律”一节及各层规律注释：四条归纳出自《AI Native 时代的研发》§07，标注为作者归纳而非图谱统计结论，写明强化学习（04 层 VSO.ai）和预测性 ML（01/02/04/10/11 层）的例外分布，并加回链；页面视觉改版为 DeepDive 风格（mint 主题、亮暗双模式），条目内容未改；更正页头版本标注（原写 v2 · 2026-09-08）。"}],"tech_legend":{"rl":"强化学习 RL","llm":"生成式AI / LLM","mlp":"预测性机器学习","gpu":"GPU加速 + 生成式光刻","rba":"规则式自动化 RBA"}}},"ai-models":{"title":"AI 新模型台账","h1":"新模型台账","ledger":true,"meta":{"scope":"其它几张 Landscape 图谱画的是结构，这一张记发布：每期 AI Buzzwords 的「本周新模型」栏目核实过的主要新模型——谁家的、哪天发的、多大、能不能下、什么许可、一手链接在哪。每期的栏目就是这里按那一期的时间窗口筛出来的视图；想看别的时间段、只看开放权重、或者按开发者和关键词找，在下面改筛选。","methodology":"","version":8,"first_published":"2026-10-23","last_updated":"2026-10-30","summary":"持续更新的 AI 新模型台账：每期 AI Buzzwords「本周新模型」栏目核实过的主要新模型——开发者、发布日、规模、可得性、许可和一手链接，按类型分组，可按期、按时间段、按开放权重 / 闭源筛选。","tags":["landscape","models","releases","open-weights"],"version_log":[{"v":1,"date":"2026-10-23","summary":"首次建立：把 EP.108「本周新模型」栏目做成可筛选的台账页，收录 34 个模型。每期 Buzzwords 的栏目从这里按该期窗口筛出（?ep=N），也可以自己选时间段、类型、开放权重 / 闭源、关键词。"},{"v":2,"date":"2026-10-23","summary":"补录 EP.101 到 EP.107：把这七期 Buzzwords 里提到的模型发布并入台账（新增 63 个，另有 17 个已在台账里的模型记上这些期也提到过），按当期窗口可筛；新增「决策模型」类的早期条目。同一模型多期出现的只留一份完整记录，其他期用引用。"},{"v":3,"date":"2026-10-28","summary":"补录 EP.90 到 EP.100：把这十一期 Buzzwords 里提到的模型发布并入台账（新增 68 个，另有 3 个已在台账里的模型记上这些期也提到过）。台账现有 184 个模型、20 期，每期用 ?ep=N 切换。"},{"v":4,"date":"2026-10-29","summary":"EP.109 补收：新增 Gemini 4 Carbon、Mirendil Fern、NVIDIA Nemotron 3.5 Transcribe、GPT-6.1 Terra、LFM3，更新 Kimi K3.1 Pro 等；台账共 189 个模型、20 期"},{"v":5,"date":"2026-10-30","summary":"EP.109 收尾：补录本期页面提到的 8 个模型（Hy-SimulMT、WeVisOne-VL、EdgeRock-4B、OmniVAE、DM0.5、Metatron v1、Nemotron-Labs Unified ASR、NLD-Uni），本期共 22 个；台账共 197 个模型、20 期"},{"v":6,"date":"2026-10-30","summary":"Mistral Large 4 权重 10-29 放出：改为开放权重（MIT 改版许可），归入 EP.109；本期共 23 个"},{"v":7,"date":"2026-10-30","summary":"EP.109 模型标签改短（手机上不再横向溢出），时间线说明挪进要点；本期共 23 个"},{"v":8,"date":"2026-10-30","summary":"Mirendil Fern 0.1 加 NEO LAB № 64 档案链接；其余不变，本期共 23 个"}]},"layers":[{"id":"layer:ai-models--1","num":1,"key":"open-llm","name":"开放权重语言模型（含多模态）","role":"权重能下载的模型。许可证写在每行的“可得性”里：Apache-2.0 和 MIT 是常见的宽许可，OpenMDW 和各家自己的模型许可各有条件，别当成同一类。","rows":53},{"id":"layer:ai-models--2","num":2,"key":"closed-llm","name":"闭源 / 仅 API 的语言模型","role":"只能通过 API 或订阅用的模型。官方说要开放、但权重还没放出的也放在这里，权重放出后改成“开放权重”并挪到上一类。","rows":42},{"id":"layer:ai-models--3","num":3,"key":"speech","name":"语音：转写与合成","role":"","rows":16},{"id":"layer:ai-models--4","num":4,"key":"retrieval","name":"嵌入、重排与文档解析","role":"","rows":9},{"id":"layer:ai-models--5","num":5,"key":"video-world","name":"图像、视频与世界模型","role":"","rows":44},{"id":"layer:ai-models--6","num":6,"key":"science","name":"科学、生物与结构化数据","role":"这一组的基准几乎都是各家自跑的，生物类还多数只有计算验证。","rows":14},{"id":"layer:ai-models--7","num":7,"key":"accel","name":"推理加速：草稿模型（不是独立模型）","role":"草稿模型要配一个目标模型用，单独跑没有意义。","rows":3},{"id":"layer:ai-models--8","num":8,"key":"decision","name":"决策模型（这几周新出现的一类）","role":"","rows":16}],"rows":197},"ai-industry-signals":{"title":"AI 行业信号流","h1":"行业信号流","ledger":true,"meta":{"scope":"其它几张 Landscape 图谱画的是\"结构\"——谁在场、扮演什么角色。这一张不画结构，记\"动态\"：Benchmark 榜单这周有没有新王、厂商的开源仓库这周提交了什么、官方 Blog 发了什么、大厂高管和独立评论者说了什么、独立研究机构给出了什么产业层判断、License与核心条款有没有实质变化。不隶属任何一期 Buzzwords，也不是热点新闻——是\"这批固定信源，这周有没有值得记的动作\"的持续台账，来自 buzzwords-signals 技能每期的扫描结果，多数周期扫完可能什么都不加。","methodology":"","version":1,"first_published":"2026-09-03","last_updated":"2026-09-16","summary":"持续记录的行业信号台账：Benchmark 榜单变化、厂商开源 Repo 动态、厂商 Blog 新文章、AI 大厂专家与独立评论者的新言论、独立 AI 产业研究机构的新内容。每期 AI Buzzwords 编辑必查，由 buzzwords-signals 技能扫描并追加。","tags":["landscape","signals","benchmark","watchlist"],"version_log":[{"v":1,"date":"2026-09-03","summary":"首次建立：五类固定信源（Benchmark 榜单/厂商开源Repo/厂商Blog/专家大V言论/独立研究机构），由 buzzwords-signals 技能维护，每期 Buzzwords 编辑必查后确认的变化记进这里。和其它4张图谱不同，这张不是层/玩家结构，是按时间倒序的信号台账。首条记录：LMArena 改名 Arena、域名迁到 arena.ai。"}]},"layers":[{"id":"layer:ai-industry-signals--1","num":1,"key":"benchmark","name":"Benchmark","role":"","rows":8},{"id":"layer:ai-industry-signals--2","num":2,"key":"repo","name":"开源 Repo","role":"","rows":5},{"id":"layer:ai-industry-signals--3","num":3,"key":"blog","name":"厂商 Blog","role":"","rows":14},{"id":"layer:ai-industry-signals--4","num":4,"key":"expert","name":"专家/大V","role":"","rows":10},{"id":"layer:ai-industry-signals--5","num":5,"key":"research","name":"独立研究机构","role":"","rows":32},{"id":"layer:ai-industry-signals--6","num":6,"key":"license","name":"License / 条款","role":"","rows":13}],"rows":82}}}