基于 Metix AI 8.6 亿+ 全球人才库，对美英法 13 家前沿 AI 实验室做全量画像：谁在为谁输送与流失人才、各 Lab 的人员构成与留存、点名级人才互流网络、工程师与研究员层的组织拼图，以及基于全量档案的华人维度分析。Built on Metix AI's global talent database of 860M+ profiles, a full-population profile of 13 frontier AI labs across the US, UK and France: who feeds and loses talent to whom, each lab's staff composition and retention, a named mutual-poaching flow network, an org reconstruction of the engineer and researcher tiers, and a full-profile analysis of the Chinese-talent dimension.
以下数字为 Metix AI 数据库口径（数据时点约 2026 年上半年），统计对象 = 当前在职于 13 家前沿 AI Lab、坐标美国/英国/法国的人才。The figures below follow Metix AI database methodology (data as of roughly the first half of 2026); the population = talent currently employed at the 13 frontier AI labs and based in the US, UK or France.
13 家 Lab 可见技术池中华人占 34.6%，其中 Meta 研究线华人浓度达 49.1%。这与 MacroPolo「美国机构顶级 AI 研究者本科 38% 来自中国」、黄仁勋「全球 50% AI 研究者是华人」的宏观判断在个体档案层面互证，是市场上首次用全量档案做的华人分章。Chinese talent makes up 34.6% of the visible technical pool across the 13 labs, with concentration reaching 49.1% on Meta's research track. At the individual-profile level this corroborates the macro claims from MacroPolo ("38% of top AI researchers at US institutions did their undergrad in China") and Jensen Huang ("50% of the world's AI researchers are Chinese") — the market's first Chinese-talent chapter built on full-population profiles.
SignalFire 口径两年留存 Anthropic 80% 居首、OpenAI→Anthropic 净流约 8:1；xAI 12 位联创中 Musk 之外 11 位全部离场（含 4 位华人联创 2026 年初密集出走）。任期数据印证：xAI 现任技术人才任期中位仅 8 个月、Anthropic 9 个月，均处高速扩张/换血期。On SignalFire's methodology, Anthropic leads on two-year retention at 80%, with the OpenAI→Anthropic net flow running roughly 8:1; of xAI's 12 co-founders, all 11 besides Musk have departed (including 4 Chinese co-founders who left in quick succession in early 2026). Tenure data confirms it: median tenure for current technical talent is just 8 months at xAI and 9 months at Anthropic, both in a phase of rapid expansion and turnover.
OpenAI/Anthropic 大量用 Member of Technical Staff 扁平头衔（研究序列可见占比仅个位数），Meta 研究线与 DeepMind 则显性标注研究科学家（占比 73.3% / 50.2%）。这意味着对 OpenAI/Anthropic 的人才评估必须回到作品而非 title。OpenAI and Anthropic make heavy use of the flat "Member of Technical Staff" title (the visible research-track share is in single digits), while Meta's research track and DeepMind explicitly label Research Scientists (73.3% / 50.2%). The implication: evaluating OpenAI/Anthropic talent has to come back to the body of work, not the title.
Meta MSL 2025 年高价入职队列任期中位 23 个月、40.0% 已过 36 个月；叠加 xAI 中层流动与 OpenAI 留任金到期，Meta MSL's high-priced 2025 hiring cohort has a median tenure of 23 months, with 40.0% past the 36-month mark; layer on mid-level mobility at xAI and OpenAI's expiring retention grants, and1703 人「高流动性」名单a 1703-person "high-mobility" list，2026 年下半年是这一人才市场流动性的高位。 makes the second half of 2026 the high-water mark for mobility in this talent market.
以下事件全部经公开信源逐条核实（来源清单见研究底稿），只保留影响人才流动判断的事实。金额均为报道口径，公司多未确认。Every event below has been verified item by item against public sources (source list in the research backup); we keep only the facts that bear on talent-flow judgments. All dollar figures are as reported; most have not been confirmed by the companies.
2025-06 Altman 公开指控 Meta 开出「$100M 签字费」；两周内 8 人从 OpenAI 跳 Meta（含苏黎世三人组）；从 Apple 挖走基础模型负责人庞若鸣（报道口径 $200M+ 多年包）；MSL 流出 44 人名单中约 50% 为华人、75% 持 PhD。Meta 官方仅否认「单笔签字费」措辞。In 2025-06 Altman publicly accused Meta of offering "$100M signing bonuses"; within two weeks 8 people jumped from OpenAI to Meta (including the Zurich trio); Meta poached foundation-model lead Ruoming Pang from Apple (a reported $200M+ multi-year package); of the 44-person MSL roster, roughly 50% are Chinese and 75% hold a PhD. Meta officially denied only the "single signing-bonus" wording.
OpenAI 首席研究官 Mark Chen 内部信「有人闯进我们家偷走了东西」，随即 recalibrating comp；2025-08 报道口径向约 1,000 名研究/工程员工发放每人约 $150 万、两年归属的留任奖金。收购未遂转挖人成为 Meta 标准打法（SSI、Thinking Machines 均被试探后遭引才，Tulloch 案报道口径最高 $1.5B/6 年，Meta 否认数字）。OpenAI Chief Research Officer Mark Chen's internal note — "someone has broken into our home and stolen something" — was followed by recalibrating comp; as reported in 2025-08, the company handed roughly 1,000 research/engineering staff a retention bonus of about $1.5M each, vesting over two years. Failed acquisition turning into talent raids became Meta's standard play (SSI and Thinking Machines were both probed and then poached; the Tulloch case was reported at up to $1.5B/6 years, a figure Meta denied).
SignalFire 口径：OpenAI→Anthropic 净流向约 8:1，DeepMind→Anthropic 约 11:1；两年留存 Anthropic 80% > DeepMind 78% > OpenAI 67% > Meta 64%。2026-05 Karpathy 加入 Anthropic 预训练团队（公司确认）。xAI 方向：12 位联创中 Musk 之外 11 位全部离场（截至 2026-03），其中华人联创 4 人 2026 年初密集出走。On SignalFire's methodology: the OpenAI→Anthropic net flow runs about 8:1 and DeepMind→Anthropic about 11:1; two-year retention is Anthropic 80% > DeepMind 78% > OpenAI 67% > Meta 64%. In 2026-05 Karpathy joined Anthropic's pretraining team (company-confirmed). On the xAI side: of the 12 co-founders, all 11 besides Musk have departed (as of 2026-03), including 4 Chinese co-founders who left in quick succession in early 2026.
OpenAI 2026-02 一周内从 Meta 回挖两名华人（庞若鸣、Pengchuan Zhang），同批华人在 Apple/Meta/OpenAI/Google 间循环流动；人才战从研究员外扩到商业化高管（Salesforce/Snowflake/Datadog 高管、Palantir FDE）与量化人才（Jane Street 实名案例跳 Anthropic）；中国大厂反向回挖坐实（吴永辉入字节 Seed、姚顺雨任腾讯首席 AI 科学家）。In 2026-02 OpenAI counter-poached two Chinese researchers from Meta within a single week (Ruoming Pang and Pengchuan Zhang), the same cohort of Chinese talent circulating among Apple, Meta, OpenAI and Google; the talent war widened from researchers to commercialization executives (Salesforce/Snowflake/Datadog leaders, Palantir FDEs) and quant talent (a named Jane Street case moving to Anthropic); and Chinese tech giants' reverse counter-poaching was confirmed (Yonghui Wu joining ByteDance Seed, Shunyu Yao becoming Tencent's chief AI scientist).
统计对象 = 11,914 名技术人才（研究科学家 / 研究工程师 / MTS / 工程 / 安全对齐），美英法三国。Population = 11,914 technical staff (Research Scientists / Research Engineers / MTS / Engineering / Safety & Alignment) across the US, UK and France.
读数：Meta (AI 研究线) 与 Google DeepMind 两家合计占可见技术池的 58%。注意各家档案覆盖率不同：DeepMind 以独立实体计、Google 主体下的研究员会漏计；SSI 团队极小且保密（公开口径约 20 人），可见档案天然偏少。Read: Meta (AI research track) and Google DeepMind together account for 58% of the visible technical pool. Note that profile coverage differs across labs: DeepMind is counted as a standalone entity, so researchers sitting under the Google parent are undercounted; SSI's team is tiny and secretive (publicly reported at about 20 people), so visible profiles are naturally sparse.
读数：研究浓度呈两极。Meta 研究线（73.3%）、DeepMind（50.2%）、Mistral、Kyutai、H 显性标注研究序列且占比过半；OpenAI（9.6%）、Anthropic（6.4%）、xAI 用扁平 MTS 头衔，研究占比被低估。气泡大小显示华人集中在量大的几家。Read: research density is bimodal. Meta's research track (73.3%), DeepMind (50.2%), Mistral, Kyutai and H explicitly label the research track and clear the 50% mark; OpenAI (9.6%), Anthropic (6.4%) and xAI use the flat MTS title, so their research share is understated. Bubble sizes show Chinese talent concentrated at the larger labs.
读数：OpenAI/Anthropic 的 MTS 列最厚（扁平头衔，不分研究/工程），Meta 研究线与 DeepMind 的研究科学家列最厚；安全与对齐是 Anthropic 的相对特色列。对买方的含义：要「明确研究科学家」去 Meta/DeepMind 挖，要「全栈型 MTS」去 OpenAI/Anthropic 挖。Read: OpenAI and Anthropic have the thickest MTS column (a flat title that doesn't split research from engineering), while Meta's research track and DeepMind have the thickest Research Scientist column; Safety & Alignment is Anthropic's relative signature column. For buyers: if you want "clearly designated Research Scientists," poach from Meta/DeepMind; if you want "full-stack MTS," poach from OpenAI/Anthropic.
读数：三条主进水管清晰。① 大厂内部转岗是最粗的管子（Google→DeepMind 体量最大，本就是同一体系）；② 高校/科研直招（尤其 Meta 研究线与 DeepMind）是研究人才的第一入口，印证顶级 PhD 未毕业即被预订；③ 「初创/其他」的大体量反映前沿 Lab 之间与广义 AI 创业生态的高频流动。量化基金（Jane Street/Citadel 系）作为新管道已在 Anthropic/OpenAI 出现实名案例。Read: three main intake pipes stand out. ① Internal transfers within the big techs are the thickest pipe (Google→DeepMind is the largest, since it's the same system to begin with); ② direct hiring from academia/research (especially Meta's research track and DeepMind) is the primary entry point for research talent, confirming that top PhDs are spoken for before they graduate; ③ the large "startups/other" volume reflects the high-frequency flow among frontier labs and the broader AI startup ecosystem. Quant funds (Jane Street/Citadel alumni) have emerged as a new pipe, with named cases already appearing at Anthropic/OpenAI.
读数：现任技术人才中 2025 年入职 4763 人，是 2023 年低谷（895 人）的 5.3 倍，2024-2025 两年的爆发式入职就是这轮人才战的时间形状。注意 survivorship：早年队列已被离职稀释，曲线低估历史招聘量、越近越接近真实强度。Read: 4763 of the current technical staff joined in 2025, 5.3x the 2023 trough (895), and the explosive hiring of 2024-2025 is the temporal shape of this talent war. Mind survivorship: earlier cohorts are already diluted by attrition, so the curve understates historical hiring and gets closer to true intensity the more recent the year.
读数：右上「老兵区」= Meta 研究线（任期中位 23 个月、40.0% 过 36 个月，FAIR 老人沉淀）；左下「新军区」= xAI（8 个月）、Anthropic（9 个月）、Mistral（8 个月）正高速扩张，入职蜜月期人员相对稳定，12-24 个月后进入第一轮流动。OpenAI（13 个月）、DeepMind（15 个月）居中。Read: the upper-right "veterans zone" = Meta's research track (median tenure 23 months, 40.0% past 36 months, the sediment of FAIR old-timers); the lower-left "new recruits zone" = xAI (8 months), Anthropic (9 months) and Mistral (8 months), all expanding fast, with onboarding-honeymoon staff relatively stable before entering their first round of mobility at 12-24 months. OpenAI (13 months) and DeepMind (15 months) sit in the middle.
2024 年起的净流动（来源含 Character.AI/Inflection/Stability/Adept 四个解体扩散源）。目的地结构与比值是更稳健的读法。Net flow since 2024 (sources include the four dissolution/diffusion origins Character.AI/Inflection/Stability/Adept). Destination structure and ratios are the more robust read.
读数：人才互流网络是本报告对 VC/猎头最直接的资产。对角线外的格子就是已被验证的流动通道（候选人心理阻力最低）：从某 Lab 离职后落到另一家 Lab 的人数越多，说明这条引才路径越「通」。Character.AI/Inflection/Adept 三个被大厂 acqui-hire 解体的团队，其成员扩散去向尤其值得关注。Read: the mutual-poaching network is this report's most directly actionable asset for VCs and recruiters. The off-diagonal cells are the proven flow channels (lowest psychological resistance for candidates): the more people who leave one lab and land at another, the more "open" that hiring path is. The diffusion paths of the three teams dissolved by big-tech acqui-hires — Character.AI, Inflection and Adept — are especially worth watching.
读数：两套口径互相印证方向。Metix AI 可见留存（当前在职 / 当前 + 可见离职）受样本上限影响绝对值偏高，但 Lab 间排序与 SignalFire 的 2 年 cohort 口径（Anthropic 80% > DeepMind 78% > OpenAI 67% > Meta 64%）方向一致：Anthropic 留得住、Meta 留不住。Read: the two methodologies corroborate each other directionally. Metix AI visible retention (current / current + visible leavers) runs high in absolute terms because of the sample cap, but the ranking across labs matches SignalFire's 2-year cohort methodology (Anthropic 80% > DeepMind 78% > OpenAI 67% > Meta 64%): Anthropic holds onto people, Meta doesn't.
读数：流入流出均为样本，净值作方向参考。扩张期的 Lab（Anthropic/OpenAI）流入显著大于可见流出；xAI 在联创清零背景下流出方向明确。真实流出大于偏保守，但排序可信。Read: inflow and outflow are both samples, so net values are directional. Labs in expansion (Anthropic/OpenAI) show inflow well above visible outflow; against the backdrop of a co-founder exodus, xAI's outflow direction is clear. True outflow exceeds the conservative count, but the ranking is trustworthy.
读数（VC 视角）：创业率最高的来源就是 spinout deal flow 的第一机会集中区。Character.AI/Inflection/Adept 这类被解体团队的离职者创业与加入新 Lab 的比例显著偏高；DeepMind/OpenAI 的离职者则更多流向其他前沿 Lab 与互联网大厂。「自主创业」按当前 title 含 founder/stealth 判定，是下限，真实创业人数更高。Read (VC view): the sources with the highest founding rate are the first concentration of spinout deal flow. Leavers from dissolved teams like Character.AI/Inflection/Adept found companies and join new labs at notably higher rates; DeepMind/OpenAI leavers flow more to other frontier labs and internet big techs. "Founded a company" is identified by current titles containing founder/stealth, a lower bound, so the true number of founders is higher.
市场上无人用全量档案做过的章节。数据库口径：13 家 Lab 技术池中华人 4,119 人（高置信 3,981），占 34.6%。外部基准：MacroPolo 口径，美国机构顶级 AI 研究者中本科来自中国者占 38%；黄仁勋称「全球 50% 的 AI 研究者是华人」。A chapter no one in the market has done on full-population profiles. Database methodology: 4,119 Chinese in the technical pool across the 13 labs (3,981 high-confidence), 34.6% of the total. External benchmarks: on MacroPolo's methodology, 38% of top AI researchers at US institutions did their undergrad in China; Jensen Huang says "50% of the world's AI researchers are Chinese."
读数：Meta 研究线华人浓度 49.1% 居首（与 2025 流出名单「约 50% 华人」吻合），Thinking Machines、Reflection、World Labs 等新锐 Lab 同样高度华人化。整体 34.6% 的占比意味着任何前沿 Lab 的人才策略都绕不开华人网络。Read: Meta's research track leads on Chinese concentration at 49.1% (consistent with the 2025 outflow roster's "about 50% Chinese"), and emerging labs like Thinking Machines, Reflection and World Labs are likewise heavily Chinese. The overall 34.6% share means no frontier lab's talent strategy can sidestep the Chinese network.
读数：华人在研究科学家与工程两个序列都是主力，MTS（OpenAI/Anthropic 扁平头衔）中占比同样可观。这说明华人不只在「执行层」，在最核心的研究序列同样占有重要比例。Read: the Chinese cohort is a mainstay of both the Research Scientist and Engineering tracks, and its share within MTS (the OpenAI/Anthropic flat title) is sizable too. This shows Chinese talent isn't confined to the "execution layer" — it holds a significant share of the most core research track as well.
读数：华人池 PhD 率 51.5%（全池 36.3%）。「清北+浙大/中科大本科 → 美国 Top CS 博士 → Lab」是最高频路径，与 MacroPolo 的宏观结论在个体档案层面互证。Read: the Chinese pool's PhD rate is 51.5% (vs 36.3% for the whole pool). "Tsinghua/Peking + Zhejiang/USTC undergrad → top US CS PhD → lab" is the highest-frequency path, corroborating MacroPolo's macro conclusion at the individual-profile level.
掌舵层Leadership tier：Mark Chen（OpenAI 首席研究官）、赵晟佳 Shengjia Zhao（Meta MSL 首席科学家，清华 + Stanford）、翁荔 Lilian Weng（Thinking Machines 联创）、纪怀新 Ed Chi（Google DeepMind 研究 VP）、李飞飞（World Labs CEO）。: Mark Chen (OpenAI Chief Research Officer), Shengjia Zhao (Meta MSL chief scientist, Tsinghua + Stanford), Lilian Weng (Thinking Machines co-founder), Ed Chi (Google DeepMind VP of Research), Fei-Fei Li (World Labs CEO).
2025 Meta 引才名单中的华人Chinese on Meta's 2025 recruitment roster（流出 44 人名单约 50% 为华人）：毕树超 Shuchao Bi、余家辉 Jiahui Yu、任泓宇 Hongyu Ren、常慧文 Huiwen Chang、Ji Lin、Pei Sun、Xinyun Chen、翟晓华 Xiaohua Zhai 等，多为清北/浙大/中科大本科赴美博士。 (about 50% of the 44-person outflow roster are Chinese): Shuchao Bi, Jiahui Yu, Hongyu Ren, Huiwen Chang, Ji Lin, Pei Sun, Xinyun Chen, Xiaohua Zhai and others — most did their undergrad at Tsinghua/Peking/Zhejiang/USTC before a US PhD.
2026 循环流动2026 circular mobility：庞若鸣 Ruoming Pang 完成 Apple → Meta（$200M 报道口径）→ OpenAI（2026-02）三连跳；Pengchuan Zhang 同期从 Meta 回到 OpenAI；xAI 四位华人联创（吴宇怀、张国栋、戴子航、杨格）2026 年初全部离场，去向是下一个创业故事池。: Ruoming Pang completed a three-hop run, Apple → Meta (a reported $200M) → OpenAI (2026-02); Pengchuan Zhang returned from Meta to OpenAI in the same window; xAI's four Chinese co-founders (Yuhuai Wu, Guodong Zhang, Zihang Dai, Greg Yang) all departed in early 2026, and where they land is the next pool of founding stories.
回流中国Return to China：吴永辉（Google Fellow/GDM VP → 字节 Seed 负责人）、姚顺雨（OpenAI → 腾讯首席 AI 科学家）为最重磅实名案例；CNBC 2026-06 报道腾讯/阿里/字节正系统性回挖。: Yonghui Wu (Google Fellow/GDM VP → head of ByteDance Seed) and Shunyu Yao (OpenAI → Tencent chief AI scientist) are the most prominent named cases; CNBC reported in 2026-06 that Tencent/Alibaba/ByteDance are systematically counter-poaching.
华人互挖的点名级证据（Metix AI 数据库口径，离职者当前去向）：DeepMind 华人离职者大量流向 Meta 研究线与 OpenAI；OpenAI 离职华人显著流向 Anthropic；被收购解体的 Inflection/Stability/Adept 释放的华人扩散进各前沿 Lab 与创业生态。这与「同一批华人在 OpenAI/Meta/Apple/Google 间循环流动」的舆论观察在数据上吻合。Named evidence of Chinese mutual poaching (Metix AI database methodology, leavers' current destinations): Chinese DeepMind leavers flow heavily to Meta's research track and OpenAI; Chinese OpenAI leavers flow markedly to Anthropic; Chinese talent released by the acquired-and-dissolved Inflection/Stability/Adept diffuses into the various frontier labs and the startup ecosystem. This matches, in the data, the public observation that "the same cohort of Chinese talent circulates among OpenAI/Meta/Apple/Google."
The Information 等付费 org chart 只覆盖高管层。本节用全量档案把还原推进到带队层与 IC 厚度。层级依据公开职位 title 归类，非官方组织架构，仅作团队梯队结构概览；OpenAI/Anthropic 大量 MTS 头衔无层级信息。公开版人名默认模糊。Paid org charts like The Information's only cover the executive layer. This section uses full-population profiles to push the reconstruction down to the team-lead tier and IC depth. Levels are classified from public job titles — not the official org structure, just an overview of team-tier structure; OpenAI/Anthropic's many MTS titles carry no level information. Names are masked by default in the public version.
双头研究领导（CRO Mark Chen + 首席科学家 Jakub Pachocki）之下，公开 org chart 止于 VP 层。2026 年 Tworek/Weil/Peebles 等离任后中层是理解 OpenAI 的关键。Below the dual research leadership (CRO Mark Chen + Chief Scientist Jakub Pachocki), the public org chart stops at the VP layer. After the 2026 departures of Tworek/Weil/Peebles and others, the mid-level is the key to understanding OpenAI.
一年规模翻倍至约 5,200 人（公开口径）。预训练团队负责人 Nick Joseph，2026-05 Karpathy 加入该团队（公开信源）。Doubled in a year to about 5,200 people (publicly reported). Pretraining-team head Nick Joseph; in 2026-05 Karpathy joined that team (public source).
伦敦 + 湾区双中心（CTO Kavukcuoglu 2025 年移驻 Mountain View 兼 Google 首席 AI 架构师）。带队层流失是其主要风险（Microsoft AI 挖走 20+）。Dual hubs in London + the Bay Area (CTO Kavukcuoglu relocated to Mountain View in 2025, doubling as Google's chief AI architect). Team-lead attrition is its main risk (Microsoft AI poached 20+).
从名单中按级别、方向与履历强度精选出三组代表性人物。档案事实来自 Metix AI 数据库；标注「公开核实」者已对照 2025-2026 公开信源确认现职（前沿 Lab 档案更新滞后，公开信源优先）。本节人物均来自公开职业档案。A 组为行业公众人物，B 组为资深技术骨干，C 组为跨界与高潜画像。公开版人名默认模糊。Three groups of representative profiles, selected from the list by level, direction and strength of track record. Profile facts come from the Metix AI database; those marked "publicly verified" have had their current role confirmed against 2025-2026 public sources (frontier-lab profiles update with a lag, so public sources take priority). All profiles in this section come from public professional records. Group A is industry public figures, Group B is senior technical backbone, Group C is crossover and high-potential profiles. Names are masked by default in the public version.
前沿 Lab 薪酬已分裂为两个市场：levels.fyi 可查的「职级市场」与绕开职级体系的「名单市场」（$10M 到 $1.5B 报道口径）。数字为 2026-06 检索的自报样本中位，非公司官方。Frontier-lab compensation has split into two markets: the "level market" you can look up on levels.fyi, and the "list market" that bypasses the leveling system ($10M to $1.5B, as reported). Figures are the median of self-reported samples retrieved in 2026-06, not official company numbers.
| Lab | 普通职级中位 TCRegular-level median TC | 高阶参考Senior reference | 股权机制Equity mechanism | 备注Notes |
|---|---|---|---|---|
| OpenAI | L5 $819K / L6 $1.23M | MTS 样本 $300K base + ~$500K/年 PPUMTS sample: $300K base + ~$500K/yr PPU | PPU 利润分享单位，4 年线性归属，tender 提供流动性PPU profit-participation units, 4-year linear vesting, liquidity via tender | 2025-08 报道口径：约 1,000 人 × $1.5M 留任金（2 年归属）As reported in 2025-08: ~1,000 people × $1.5M retention grant (2-year vest) |
| Anthropic | SWE 中位 $665KSWE median $665K | Lead 中位 $785KLead median $785K | 常规私司股权 + tenderStandard private-company equity + tender | 现金中位低于 OpenAI 仍留存第一（SignalFire）Median cash below OpenAI, yet first on retention (SignalFire) |
| Google DeepMind | L6 RS $750K-1M | L7 $950K-1.4M | GSU 上市股票，流动性最好GSU public stock, best liquidity | RS 同级股权比 SWE 高 5-15%RS equity 5-15% higher than SWE at the same level |
| Meta (MSL) | E7 中位 $1.30ME7 median $1.30M | 名单市场 $10M-100M+/年List market $10M-100M+/yr | RSU + 名单制特殊包RSU + list-based special package | 庞若鸣 $200M+、Tulloch 最高 $1.5B/6 年均为报道口径Ruoming Pang $200M+ and Tulloch up to $1.5B/6 years are both as reported |
| xAI | SWE $205-640K | 样本少Small sample | 私司期权（SpaceX 收购后置换）Private-company options (swapped after the SpaceX acquisition) | 2026-02 起并入 SpaceX 体系Folded into the SpaceX system from 2026-02 |
| Mistral (巴黎)Mistral (Paris) | 巴黎中位 €89.5KParis median €89.5K | 研究员 $490-950K（估算口径）Researchers $490-950K (estimated) | BSPCE 期权BSPCE options | 欧洲工程岗与美国差一个数量级，研究岗溢价 3-5 倍European engineering roles trail the US by an order of magnitude; research roles carry a 3-5x premium |
研究序列与工程序列的分层固化：DeepMind 同级研究岗股权高 5-15%，Mistral 研究岗总包约为工程岗 3-5 倍，Meta 名单制包只发生在研究员/研究领导层。量化基金（Jane Street/Citadel 系）与 Lab 互相抬价，Anthropic/OpenAI 主动办 mixer 挖入门级 quant，实名案例已出现（Jane Street 两人 2025 年跳 Anthropic）。The split between the research and engineering tracks has hardened: at DeepMind, same-level research roles carry 5-15% more equity; at Mistral, research total comp is roughly 3-5x engineering; and Meta's list-based packages occur only at the researcher / research-leadership level. Quant funds (Jane Street/Citadel alumni) and the labs bid each other up, and Anthropic/OpenAI proactively host mixers to poach entry-level quants, with named cases already appearing (two Jane Street people moved to Anthropic in 2025).
① 职级市场可以对表谈判，名单市场只能用使命/股权上行/算力自由度竞争；② OpenAI PPU 与 DeepMind GSU 的流动性差异是引才话术的实操杠杆（PPU 依赖公司组织 tender，GSU 随时可卖）；③ 欧洲（Mistral/Kyutai/H）是同等人才密度下的薪酬洼地，适合预算有限的买方建研发点。① The level market can be negotiated against the table; the list market can only be competed for with mission, equity upside and compute freedom; ② the liquidity gap between OpenAI's PPU and DeepMind's GSU is a practical lever in recruiting pitches (PPU depends on the company organizing a tender, GSU can be sold anytime); ③ Europe (Mistral/Kyutai/H) is a compensation depression at equivalent talent density, well suited for budget-constrained buyers building an R&D site.
把图谱变成动作：VC 看 spinout 信号，猎头看窗口与通道，HR 看防守。Turn the map into action: VCs watch spinout signals, recruiters watch windows and channels, HR watches defense.
① 创业率排行（4.4 节）指向 spinout 的高发来源；② xAI 多位华人联合创始人 2026 年初离场、去向未公开，是值得关注的组队信号；③ LeCun（AMI Labs $1.03B 种子）与 Tworek（Core Automation）印证「高管离职到成轮约 6 个月」的节奏；④ 任期结构（3.7）可用于预判：高价引入 Meta MSL 的 2025 队列若在 2026 下半年松动，新一批组队窗口随之打开。① The founding-rate ranking (Section 4.4) points to the high-frequency sources of spinouts; ② several of xAI's Chinese co-founders departed in early 2026 with destinations undisclosed — a team-forming signal worth watching; ③ LeCun (AMI Labs $1.03B seed) and Tworek (Core Automation) confirm the "roughly 6 months from executive departure to a closed round" cadence; ④ tenure structure (3.7) can be used to anticipate: if Meta MSL's high-priced 2025 cohort loosens in H2 2026, a new wave of team-forming windows opens with it.
① 人才互流矩阵（4.1）显示哪些流动通道在历史上已被验证；② 任期窗口与组织调整叠加，可定位流动性较高的群体；③ MTS 头衔不分级，能力评估需回到公开作品（模型贡献名单/论文/系统）；④ 华人技术群体有清晰的校友与社区脉络，可对照华人维度的院校分布理解其分布。① The mutual-flow matrix (4.1) shows which flow channels have been historically proven; ② tenure windows layered with organizational reshuffles help locate the more mobile groups; ③ the MTS title carries no level, so capability assessment has to come back to public work (model-contribution rosters/papers/systems); ④ the Chinese technical cohort has clear alumni and community threads — read its distribution against the institution breakdown in the Chinese-talent dimension.
① 对照 4.2 留存基准定位自身；② 业内已成型的留才工具：股权 refresh、留任金与使命叙事，OpenAI 案例给出价格参考（$1.5M × 1,000 人）；③ 关注「任职 36 个月以上且晋升放缓」的群体（3.7 流动窗口的镜像）；④ 竞业与 garden leave 在美国市场基本无效，留人靠给方向而非锁人。① Locate yourself against the retention benchmarks in 4.2; ② the industry's established retention tools — equity refreshes, retention grants and mission narrative — with OpenAI's case giving a price reference ($1.5M × 1,000 people); ③ watch the group with "36+ months of tenure and slowing promotion" (the mirror image of the 3.7 mobility window); ④ non-competes and garden leave are largely ineffective in the US market, so you keep people by giving them direction, not by locking them in.
本报告的检索、画像、流动分析全部由 Metix AI 完成。可按同样口径为任意公司生成定制图谱：全量长名单导出、组织拼图、互挖监测、邮箱解锁与多渠道触达，并按「只为合格面试付费」计费。No interview, no charge.The retrieval, profiling and flow analysis in this report were all done by Metix AI. We can generate a custom map for any company on the same methodology: full long-list export, org reconstruction, mutual-poaching monitoring, email unlock and multi-channel outreach — billed on a "pay only for qualified interviews" basis. No interview, no charge.
8.6 亿+ 全球人才画像860M+ global talent profiles11,914 人 Lab 技术池长名单11,914-person lab technical-pool long list人才互流网络季度监测Quarterly monitoring of the mutual-poaching network只为合格面试付费Pay only for qualified interviewsOpenAI、Anthropic、Google DeepMind、Meta(AI 研究线)、xAI、SSI、Thinking Machines、Mistral AI、Reflection AI、World Labs、Physical Intelligence、H Company、Kyutai。地理范围 = 档案常驻地在美国/英国/法国。Google DeepMind 以独立实体口径统计，在 Google 主体下挂 DeepMind 头衔者会少量漏计。OpenAI, Anthropic, Google DeepMind, Meta (AI research track), xAI, SSI, Thinking Machines, Mistral AI, Reflection AI, World Labs, Physical Intelligence, H Company, Kyutai. Geographic scope = profiles based in the US/UK/France. Google DeepMind is counted as a standalone entity, so a small number of people holding a DeepMind title under the Google parent are undercounted.
Meta 体量巨大且 MSL 无独立实体，本报告的「Meta(AI 研究线)」= 当前在 Meta 且 title/headline 可识别为 AI 研究/工程相关者（含明确标注 Superintelligence Labs/FAIR/GenAI 的档案），是可识别子集而非 Meta AI 全量，绝对数偏保守。Meta is enormous and MSL is not a standalone entity, so this report's "Meta (AI research track)" = people currently at Meta whose title/headline is identifiable as AI research/engineering related (including profiles explicitly tagged Superintelligence Labs/FAIR/GenAI). It is an identifiable subset rather than all of Meta AI, so the absolute numbers run conservative.
技术人才池 = 研究科学家 + 研究工程师 + MTS(未细分) + 工程 + 安全与对齐。OpenAI/Anthropic/SSI 大量使用 Member of Technical Staff 头衔不分研究/工程，单列为 MTS，研究序列占比因此是下限口径。解决方案/产品/GTM 不计入技术池。Technical talent pool = Research Scientists + Research Engineers + MTS (unspecified) + Engineering + Safety & Alignment. OpenAI/Anthropic/SSI make heavy use of the Member of Technical Staff title without splitting research from engineering, broken out separately as MTS, so the research-track share is a lower-bound methodology. Solutions/product/GTM are not counted in the technical pool.
采用五信号交叉验证（姓名族裔模型 / 汉字 / 中文 / 中国院校 / 多拼写姓氏库），分高/中置信，主口径 = 高 + 中。人才库数据时点约 2026 年上半年；前沿 Lab 档案更新存在滞后，重点人选已逐人对照公开信息复核，2025-2026 的最新职位变动以公开信源为准标注。We use five-signal cross-validation (name-ethnicity model / Chinese characters / Chinese language / Chinese institutions / multi-spelling surname library), split into high/medium confidence, with the primary methodology = high + medium. The talent database is as of roughly the first half of 2026; frontier-lab profiles update with a lag, so the key profiles have been re-checked one by one against public information, and the latest 2025-2026 role changes are annotated according to public sources.
| Lab | 在职画像current-employee profiles | 技术池Technical pool | 研究序列占比Research-track share | 华人技术Chinese technical | 华人占比Chinese share | PhD 率PhD rate |
|---|---|---|---|---|---|---|
| Meta (AI 研究线)Meta (AI research track) | 4,228 | 4,092 | 73.3% | 2,011 | 49.1% | 59.8% |
| Google DeepMind | 4,123 | 2,837 | 50.2% | 837 | 29.5% | 36.9% |
| OpenAI | 5,495 | 2,628 | 9.6% | 746 | 28.4% | 16.0% |
| Anthropic | 2,740 | 1,251 | 6.4% | 246 | 19.7% | 16.0% |
| xAI | 1,717 | 710 | 5.1% | 217 | 30.6% | 16.1% |
| Mistral AI | 528 | 263 | 51.7% | 21 | 8.0% | 17.5% |
| Thinking Machines | 88 | 64 | 1.6% | 24 | 37.5% | 43.8% |
| Reflection AI | 41 | 21 | 0.0% | 7 | 33.3% | 42.9% |
| H Company | 32 | 19 | 63.2% | 0 | 0.0% | 10.5% |
| World Labs | 24 | 12 | 8.3% | 6 | 50.0% | 33.3% |
| Physical Intelligence | 25 | 9 | 11.1% | 3 | 33.3% | 33.3% |
| Kyutai | 13 | 8 | 75.0% | 1 | 12.5% | 37.5% |
| SSI | 3 | 0 | 0% | 0 | 0% | 0% |
① 数据时效Data recency：人才库为静态数据（约 2026 年上半年），前沿 Lab 人员变动极快、档案更新滞后明显（高频跳槽人群常隔数月才改档案），所有数字为方向性而非实时；重点人选已逐人对照公开信源复核并标注。: the talent database is a static snapshot (roughly the first half of 2026); frontier labs see extremely rapid staff turnover and a pronounced profile-update lag (high-frequency job-hoppers often don't update their profiles for months), so all figures are directional rather than real-time; the key profiles have been re-checked and annotated one by one against public sources.
② 覆盖率不均Uneven coverage：本报告基于公开职业档案聚合：SSI 等保密团队、DeepMind（以独立实体计）与 Meta（取可识别 AI 研究线子集）的覆盖度受限。各 Lab 之间比较以占比与结构为主、绝对数为辅。: this report aggregates public professional records, so coverage is limited for secretive teams like SSI, for DeepMind (counted as a standalone entity) and for Meta (taken as the identifiable AI-research-track subset). Comparisons across labs lean on shares and structure, with absolute numbers secondary.
③ 职能与层级推断Function and level inference：基于 title/headline 关键词；MTS 等扁平头衔使研究序列占比与带队层人数均为下限。: based on title/headline keywords; flat titles like MTS make both the research-track share and the team-lead-tier headcount lower bounds.
④ 华人识别为概率判定Chinese identification is probabilistic：主口径 4,119 人 = 高置信 3,981 + 中置信 138；使用西文名且无中国信号的华裔会漏检。: the primary methodology gives 4,119 = 3,981 high-confidence + 138 medium-confidence; people of Chinese descent who use a Western name and carry no Chinese signal are missed.
⑤ 流出与留存Outflow and retention：离职者检索有样本上限，流出为偏保守；本报告「可见留存率」与 SignalFire 的 2 年 cohort 口径不同，对照阅读而非互相替代。: leaver retrieval has a sample cap, so outflow runs conservative; this report's "visible retention rate" differs from SignalFire's 2-year cohort methodology, so read them side by side rather than as substitutes for each other.
⑥ 金额口径Dollar-figure methodology：薪酬包均为媒体报道口径，公司多未确认；引用时请保留「报道口径」字样。: compensation packages are all as reported by the media and mostly unconfirmed by the companies; please keep the "as reported" qualifier when citing them.