世界模型与机器人仿真急需的图形、渲染、物理仿真人才，存量沉淀在游戏引擎公司。基于 Metix AI 8.6 亿+ 全球人才库，量化 Unity / Epic / Roblox / NVIDIA 图形线的这批人有多少、在哪、教育背景如何，以及他们正以什么速度流向 World Labs / Luma / Sora / Genie 等世界模型公司。The graphics, rendering, and physics-simulation talent that world models and robotics simulation desperately need is sitting inside game-engine companies. Drawing on Metix AI's 0.86B+ global talent base, this report quantifies how many of these people the Unity / Epic / Roblox / NVIDIA graphics lines hold, where they are, what their educational background looks like, and how fast they are flowing toward world-model companies such as World Labs / Luma / Sora / Genie.
以下数字为 Metix AI 数据库数据库口径（数据截至 2026 年上半年），地理范围美/英/加。「图形人才」= 图形与渲染 / 物理与仿真 / 引擎与工具链 / 3D 视觉与几何 四类可迁移职能。The figures below are based on the Metix AI database (data through H1 2026), covering the US / UK / Canada. "Graphics talent" = the four transferable functions of graphics & rendering / physics & simulation / engine & toolchain / 3D vision & geometry.
世界模型本质是「学出来的游戏引擎」：它要的实时渲染、物理仿真、3D 几何、神经渲染（NeRF / Gaussian Splatting）能力，正是引擎公司过去二十年积累的稀缺技能。本报告在 Unity / Epic / Roblox / NVIDIA 图形线中识别出A world model is, at its core, a "learned game engine": the real-time rendering, physics simulation, 3D geometry, and neural rendering (NeRF / Gaussian Splatting) it demands are exactly the scarce skills engine companies have accumulated over the past two decades. Across the Unity / Epic / Roblox / NVIDIA graphics lines, this report identifies 486 名图形/渲染/仿真人才486 graphics / rendering / simulation specialists，这是一座尚未被 AI 招聘充分触达的人才储备。 — a talent reserve that AI recruiting has barely begun to reach.
世界模型公司的现员工中Among current staff at world-model companies, 11.7% 来自游戏、视效或图形大厂；其中 3D / 视频生成型公司远高于纯语言模型团队（OpenAI (Sora 线) 达 18.0%）。直接从引擎三杰跳出的还少，但广义「图形 → 世界模型」的流动已成规模，是公开报道几乎空白、只能从档案挖出的人才信号。 come from gaming, VFX, or major graphics companies; 3D / video-generation companies run far higher than pure language-model teams (OpenAI (Sora line) reaches 18.0%). Direct hops out of the engine "big three" are still rare, but the broader "graphics → world model" flow has reached real scale — a talent signal almost absent from public coverage and visible only by mining the profiles.
World Labs 联创 Christoph Lassner（前 Epic Games + Meta Reality Labs，Gaussian Splatting 先驱 Pulsar 渲染器作者）、Luma CTO Alex Yu（Berkeley NeRF/Plenoxels 一作）、NVIDIA Cosmos 负责人 Ming-Yu Liu，都是「图形渲染 → 世界模型」的教科书路径。技能可迁移性有公开权威背书。World Labs co-founder Christoph Lassner (ex-Epic Games + Meta Reality Labs, Gaussian Splatting pioneer and author of the Pulsar renderer), Luma CTO Alex Yu (lead author of Berkeley's NeRF/Plenoxels), and NVIDIA Cosmos lead Ming-Yu Liu are all textbook "graphics rendering → world model" paths. The transferability of these skills carries authoritative public endorsement.
引擎侧图形池华人占 8.8%，PhD 率 25.6%。SIGGRAPH 约三成技术论文含中国机构作者，浙大 CAD&CG、清华是全球图形学华人两大「黄埔」；NVIDIA Cosmos 世界模型团队技术署名近半为华人。世界模型赛道的华人浓度天然高。In the engine-side graphics pool, Chinese talent makes up 8.8%, with a 25.6% PhD rate. Roughly 30% of SIGGRAPH technical papers list authors from Chinese institutions; Zhejiang University's CAD&CG lab and Tsinghua are the two great "academies" of Chinese graphics talent worldwide, and nearly half the technical bylines on NVIDIA's Cosmos world-model team are Chinese. The Chinese density of the world-model field is naturally high.
以下论断均有公开信源支撑（完整清单见研究底稿），只保留影响人才判断的事实。Every claim below is backed by public sources (full list in the research backup); we keep only the facts that bear on talent judgment.
DeepMind GameNGen 论文标题即《Diffusion Models Are Real-Time Game Engines》，官方称「首个完全由神经模型驱动的游戏引擎」，单 TPU 实时跑 DOOM。Odyssey Agora-1 自述「a learned game engine，无手写玩法逻辑与传统渲染器」，并把仿真与渲染分开（与引擎架构同构）。World Labs 旗舰产品 Marble 直接输出 Gaussian splat 渲染 + 碰撞网格物理，喂进 NVIDIA Isaac Sim 训机器人。DeepMind's GameNGen paper is literally titled "Diffusion Models Are Real-Time Game Engines," billed officially as "the first game engine driven entirely by a neural model," running DOOM in real time on a single TPU. Odyssey's Agora-1 describes itself as "a learned game engine, with no hand-written gameplay logic or traditional renderer," and separates simulation from rendering (structurally isomorphic to engine architecture). World Labs' flagship product Marble outputs Gaussian-splat rendering plus collision-mesh physics directly, feeding into NVIDIA Isaac Sim to train robots.
公开口径：「实时物理、shader 编程、图形渲染与物理仿真，正是开发下一代世界模型系统所需」。神经渲染（NeRF / 3D Gaussian Splatting）与引擎渲染共用 GPU 光栅化/着色心智模型，3DGS 已可经插件进 Unreal Engine 5、OpenUSD（2026 官方支持）、Omniverse。会写引擎渲染管线的人，迁移成本最低。As stated publicly: "real-time physics, shader programming, graphics rendering, and physics simulation are exactly what's needed to build the next generation of world-model systems." Neural rendering (NeRF / 3D Gaussian Splatting) shares the same GPU rasterization/shading mental model as engine rendering, and 3DGS can already be plugged into Unreal Engine 5, OpenUSD (officially supported in 2026), and Omniverse. Those who can write an engine rendering pipeline face the lowest switching cost.
需求侧：World Labs（累计 $1.23B）、Decart（估值近 $4B）、Luma（估值 $4B）、Odyssey、DeepMind Genie、OpenAI Sora、NVIDIA Cosmos/Isaac/Omniverse 同时抢「会渲染会物理会 3D」的人。供给侧：游戏业 2022-2024 累计裁员 33,600+（Unity 累计约 3,200、Epic 830，含视效工作室 Wētā 256），把稀缺图形人才大批推向市场。Demand side: World Labs ($1.23B raised to date), Decart (valuation near $4B), Luma (valuation $4B), Odyssey, DeepMind Genie, OpenAI Sora, and NVIDIA Cosmos/Isaac/Omniverse are all competing for people who "can render, do physics, and handle 3D." Supply side: gaming laid off 33,600+ people cumulatively over 2022-2024 (Unity roughly 3,200, Epic 830, including 256 at VFX studio Wētā), pushing scarce graphics talent into the market en masse.
DeepMind CEO Hassabis 十六七岁即参与 Bullfrog《Theme Park》（1994 模拟经营游戏）的设计与主程，后做 Syndicate、Black & White，再创立 DeepMind。世界模型 Genie 系列被其称为通往 AGI 的踏脚石。游戏引擎与世界模型在同一批人的脑子里本就是一回事。DeepMind CEO Hassabis, at 16 or 17, worked on the design and lead programming of Bullfrog's Theme Park (a 1994 tycoon-simulation game), later contributed to Syndicate and Black & White, then founded DeepMind. He has called the world-model Genie series a stepping stone toward AGI. In the minds of the same group of people, game engines and world models were always the same thing.
统计对象 = Unity / Epic / Roblox / NVIDIA 图形线中 486 名图形人才（图形渲染 / 物理仿真 / 引擎工具链 / 3D 视觉四类可迁移职能），美/英/加。Population = 486 graphics specialists across the Unity / Epic / Roblox / NVIDIA graphics lines (the four transferable functions of graphics rendering / physics simulation / engine toolchain / 3D vision), in the US / UK / Canada.
读数：图形与渲染、物理与仿真、引擎与工具链、3D 视觉与几何这四类「可迁移到世界模型」的职能，构成引擎侧技术池的核心。游戏玩法/客户端、基础设施等职能可迁移性较低，不计入本报告的图形人才池。Reading: the four functions transferable to world models — graphics & rendering, physics & simulation, engine & toolchain, 3D vision & geometry — form the core of the engine-side technical pool. Functions such as gameplay/client and infrastructure have lower transferability and are excluded from this report's graphics-talent pool.
统计对象 = 618 名世界模型/视频/机器人公司（World Labs / Luma / Runway / Sora 线 / Genie 线 / Physical Intelligence）的现任技术人才，看其中多少来自游戏、视效或图形大厂。Population = 618 current technical staff at world-model / video / robotics companies (World Labs / Luma / Runway / Sora line / Genie line / Physical Intelligence), measuring how many came from gaming, VFX, or major graphics companies.
读数：走廊不是均匀的。3D 重建/视频生成型公司（如 Luma、World Labs）的图形来源占比显著高于纯语言模型团队，因为它们的产品就是渲染与 3D 一致性。这印证「图形深的世界公司最先吃到引擎人才」。Reading: the corridor is not uniform. 3D-reconstruction / video-generation companies (e.g., Luma, World Labs) show a markedly higher graphics-pedigree share than pure language-model teams, because their products are precisely about rendering and 3D consistency. This confirms that "the most graphics-heavy world-model companies absorb engine talent first."
读数：流入世界模型公司的图形人才里，直接来自引擎三杰（Unity/Epic/Roblox）的还是少数，更多来自图形/视效大厂（NVIDIA、Pixar、Weta、Autodesk 等）与游戏工作室（Naughty Dog、Insomniac、EA 等）。这说明「引擎 → AI」的窄通道之外，存在一条更宽的「图形/视效 → 世界模型」走廊。Reading: among the graphics talent flowing into world-model companies, those coming directly from the engine big three (Unity/Epic/Roblox) remain a minority; more come from major graphics/VFX companies (NVIDIA, Pixar, Weta, Autodesk, etc.) and game studios (Naughty Dog, Insomniac, EA, etc.). This shows that beyond the narrow "engine → AI" channel lies a wider "graphics/VFX → world model" corridor.
创始人级Founder-level：Christoph Lassner（前 Epic Games + Meta Reality Labs，Gaussian Splatting 先驱 Pulsar 渲染器作者）→ World Labs 联合创始人；Ben Mildenhall（NeRF 一作）→ World Labs 联创；Amit Jain（前 Apple Vision Pro passthrough）→ Luma AI 联合创始人；Alex Yu（Berkeley NeRF/Plenoxels 一作）→ Luma AI CTO。: Christoph Lassner (ex-Epic Games + Meta Reality Labs, Gaussian Splatting pioneer and author of the Pulsar renderer) → World Labs co-founder; Ben Mildenhall (lead author of NeRF) → World Labs co-founder; Amit Jain (ex-Apple Vision Pro passthrough) → Luma AI co-founder; Alex Yu (lead author of Berkeley NeRF/Plenoxels) → Luma AI CTO.
研究/工程级（数据库样本，详见第 6 节）Research/engineering level (database sample, see Section 6)：清华 + Stanford 出身、履历含 NVIDIA 的资深研究科学家 → DeepMind 世界模型方向；前 NVIDIA 工程师 → DeepMind；前 Naughty Dog → Luma 工程负责人。这一层「图形/游戏 → 世界模型」的 IC 级流动公开几乎无报道，是档案数据库的独家增量。: a senior research scientist from Tsinghua + Stanford with NVIDIA on the résumé → DeepMind's world-model track; an ex-NVIDIA engineer → DeepMind; an ex-Naughty Dog engineer → Luma's head of engineering. This layer of IC-level "graphics/gaming → world model" movement is almost entirely unreported publicly, and is the exclusive added signal of a profile database.
引擎侧图形人才中华人 43 人，占 8.8%。外部基准：SIGGRAPH 约三成技术论文含中国机构作者；MacroPolo 口径顶级 AI 研究者本科 47% 来自中国。Engine-side graphics talent includes 43 Chinese individuals, or 8.8%. External benchmarks: roughly 30% of SIGGRAPH technical papers list authors from Chinese institutions; on MacroPolo's measure, 47% of top AI researchers did their undergraduate degree in China.
华人图形池 PhD 率 25.6%（全图形池 11.1%）。主流路径 = 中国图形强校（浙大 CAD&CG、清华、中科大、上科大、港科大）本硕 → 北美图形/3D 视觉 PhD（常见导师 Kanazawa、Jiajun Wu、Hao Su、Sanja Fidler）→ NVIDIA / World Labs / Luma 或自创世界模型公司。中国院校 Top 见 3.5 节右图。The Chinese graphics pool has a 25.6% PhD rate (vs. 11.1% for the full graphics pool). The mainstream path = a bachelor's/master's at a top Chinese graphics school (Zhejiang University CAD&CG, Tsinghua, USTC, ShanghaiTech, HKUST) → a graphics / 3D-vision PhD in North America (common advisors: Kanazawa, Jiajun Wu, Hao Su, Sanja Fidler) → NVIDIA / World Labs / Luma, or founding a world-model company. See the right-hand chart in Section 3.5 for the top Chinese institutions.
世界模型核心岗World-model core roles：Ming-Yu Liu 劉洺堉（NVIDIA Research VP，领 Cosmos 世界基础模型）、Alex Yu（Luma AI 联创/CTO，NeRF/Plenoxels 一作）、Jun Gao 高俊（NVIDIA Toronto，GET3D 一作）、Yongxin Chen、Kaichun Mo 牟凯淳、Tsung-Yi Lin 林宗毅（NVIDIA Cosmos 团队）。: Ming-Yu Liu 劉洺堉 (NVIDIA Research VP, leading the Cosmos world foundation model), Alex Yu (Luma AI co-founder/CTO, lead author of NeRF/Plenoxels), Jun Gao 高俊 (NVIDIA Toronto, lead author of GET3D), Yongxin Chen, Kaichun Mo 牟凯淳, and Tsung-Yi Lin 林宗毅 (NVIDIA Cosmos team).
学界上游（世界模型人才输送）Academic upstream (the world-model talent feeder)：Hao Su 苏昊（UCSD + Hillbot CTO）、Xiaolong Wang 王小龙（UCSD）、Jiajun Wu（Stanford）、Jun-Yan Zhu 朱俊彦（CMU）、Sida Peng（浙大）。NVIDIA Cosmos 世界基础模型论文技术署名近半为华人。: Hao Su 苏昊 (UCSD + Hillbot CTO), Xiaolong Wang 王小龙 (UCSD), Jiajun Wu (Stanford), Jun-Yan Zhu 朱俊彦 (CMU), and Sida Peng (Zhejiang University). Nearly half the technical bylines on NVIDIA's Cosmos world-foundation-model paper are Chinese.
避免误判Avoiding misattribution：Sanja Fidler（克罗地亚裔，领 NVIDIA Spatial Intelligence Lab）、Angjoo Kanazawa（日裔，Berkeley）、Ben Mildenhall / Justin Johnson / Christoph Lassner（World Labs 联创，非华人）均为非华人。: Sanja Fidler (Croatian, leading NVIDIA's Spatial Intelligence Lab), Angjoo Kanazawa (Japanese, Berkeley), and Ben Mildenhall / Justin Johnson / Christoph Lassner (World Labs co-founders, non-Chinese) are all non-Chinese.
按级别、方向稀缺度与履历强度筛出三组共 12 人。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（游戏/AI 人才流动快、档案更新有滞后，公开信源优先）。人才匹配度评级与全量名单见 Metix AI 平台。公开版人名默认模糊。Three groups totaling 12 people, screened by seniority, scarcity of specialization, and résumé strength. Profile facts come from the Metix AI database; those marked "publicly verified" have had their current role confirmed against 2025-2026 public sources (gaming/AI talent moves fast and profile updates lag, so public sources take precedence). Talent match-fit ratings and the full list are available on the Metix AI platform. Names are masked by default in the public version.
总包 TC = base + 股票年化 + 奖金。levels.fyi 为 2026-06 检索值。引擎侧与世界模型侧的薪酬落差是这条走廊的主要拉力之一。Total comp (TC) = base + annualized equity + bonus. levels.fyi figures retrieved 2026-06. The pay gap between the engine side and the world-model side is one of the corridor's main pulls.
| 公司 / 类别Company / category | 图形/ML 工程师中位 TCMedian TC, graphics/ML engineer | 资深参考Senior reference | 备注Notes |
|---|---|---|---|
| Unity | $220-300K | Staff $350K+ | 2024 大裁员后股票承压Equity under pressure after the 2024 layoffs |
| Epic Games | $250-340K | Principal $450K+ | 私司，Unreal 引擎团队Private; Unreal Engine team |
| Roblox | $280-380K | Principal $500K+ | 上市，工程占比 75%Public; engineers make up 75% |
| NVIDIA（图形/研究）NVIDIA (graphics/research) | $350-500K | Senior/Principal $700K-1M+ | 股票增值使老员工纸面包极高Stock appreciation makes long-tenured staff's paper comp extremely high |
| World Labs / Luma / Decart | $300-450K（私司）$300-450K (private) | 研究骨干含早期股权Research core includes early-stage equity | 现金 + 高弹性早期股权，World Labs SF 岗约 $200-350K baseCash + high-upside early equity; World Labs SF roles around $200-350K base |
| OpenAI(Sora) / DeepMind(Genie) | $500K-1.2M+ | 研究序列更高Research track runs higher | 一线 Lab 薪酬，对图形研究人才形成强拉力Top-tier lab pay, a strong pull on graphics-research talent |
引擎侧图形工程师中位 TC（$220-380K）显著低于一线 AI Lab 的世界模型/视频研究岗（$500K-1.2M+），叠加 2024 年游戏业裁员与股票承压，构成强烈的「向上跳槽」拉力。早期世界模型公司（World Labs/Luma/Decart）现金不一定更高，但早期股权 + 做前沿方向 + 把图形技能直接用在 AI 上，是对引擎人才的核心吸引力。Engine-side graphics engineers' median TC ($220-380K) sits well below world-model / video research roles at top AI labs ($500K-1.2M+); combined with the 2024 gaming layoffs and equity pressure, this creates a strong "trade-up" pull. Early-stage world-model companies (World Labs/Luma/Decart) don't necessarily pay more cash, but early equity + working at the frontier + applying graphics skills directly to AI is the core draw for engine talent.
世界模型/机器人公司若以「图形/渲染岗」对标游戏业薪酬开价，再叠加 AI 前沿叙事与早期股权，对引擎侧资深图形人才有明显竞争力；一线 Lab 则可用纯现金优势直接虹吸 NVIDIA/Pixar 级的图形研究人才。If world-model / robotics companies price "graphics/rendering roles" against gaming-industry comp and layer on a frontier-AI narrative plus early equity, they are clearly competitive for senior engine-side graphics talent; top labs, meanwhile, can use a pure cash advantage to siphon off NVIDIA/Pixar-caliber graphics researchers directly.
把走廊变成名单：HR 看去哪找人，猎头看可迁移画像，VC 看团队图形基因。Turn the corridor into a list: HR for where to find people, recruiters for transferable profiles, VCs for a team's graphics DNA.
① 别只在 AI 圈招人，最稀缺的渲染/物理/3D 人才存量在 Unity/Epic/Roblox/NVIDIA 图形线与视效工作室（Pixar/Weta/Naughty Dog）；② 优先锁定有 NeRF/Gaussian Splatting/可微渲染/物理仿真背景的人，迁移成本最低；③ 2024 游戏业裁员潮释放的资深图形人才仍在消化，是窗口；④ 本报告可识别出 197 人高匹配名单。① Don't recruit only within the AI world — the scarcest rendering/physics/3D talent sits in the Unity/Epic/Roblox/NVIDIA graphics lines and in VFX studios (Pixar/Weta/Naughty Dog); ② prioritize people with NeRF / Gaussian Splatting / differentiable-rendering / physics-simulation backgrounds, who have the lowest switching cost; ③ the senior graphics talent released by the 2024 gaming layoffs is still being absorbed, which is a window; ④ this report can identify a high-fit shortlist of 197 people.
① 走廊（4 节）显示哪些来源已被验证流通（候选人接受度高）；② 最值钱画像 = 图形渲染 + 3D 几何 + 物理仿真三选二，且有 GPU/shader/光栅化实操；③ 引擎三杰直跳世界模型仍少，更宽的「图形/视效大厂 → 世界模型」通道更现实；④ 华人候选人用 SIGGRAPH 作者网络 + 浙大/清华图形系校友链精准触达。① The corridor (Section 4) shows which sources are already proven to flow (high candidate receptivity); ② the most valuable profile = two of the three (graphics rendering + 3D geometry + physics simulation), with hands-on GPU/shader/rasterization experience; ③ direct hops from the engine big three to world models remain rare, so the wider "major graphics/VFX company → world model" channel is more realistic; ④ reach Chinese candidates precisely via the SIGGRAPH author network plus the Zhejiang University / Tsinghua graphics-department alumni chain.
① 世界模型公司的「图形深度」可量化（现员工图形来源占比，4 节）：3D/视频型公司应高、纯 LLM 型偏低；② 创始团队是否有 NeRF/GS/渲染器/引擎背景，是世界模型团队成色的硬信号（World Labs Lassner、Luma Alex Yu 是模板）；③ 用引擎→世界模型走廊密度判断一家公司是否真在做「世界」而非贴标签。① A world-model company's "graphics depth" is quantifiable (the graphics-pedigree share of current staff, Section 4): 3D/video companies should run high, pure-LLM ones lower; ② whether the founding team has a NeRF/GS/renderer/engine background is a hard signal of a world-model team's quality (World Labs' Lassner and Luma's Alex Yu are the templates); ③ use the density of the engine → world-model corridor to judge whether a company is genuinely building "worlds" rather than just slapping on the label.
本报告的检索、画像、走廊分析全部由 Metix AI 完成。可按同样口径为任意公司生成定制图谱：全量图形人才长名单导出、引擎→AI 流动监测、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.All the search, profiling, and corridor analysis in this report was done by Metix AI. We can generate a custom map for any company on the same methodology: export the full graphics-talent long list, monitor engine → AI flows, unlock emails and reach out across channels — billed on a "pay only for qualified interviews" basis. No interview, no charge.
8.6 亿+ 全球人才画像0.86B+ global talent profiles486 人引擎侧图形人才长名单486-person engine-side graphics long list引擎→世界模型走廊监测Engine → world-model corridor monitoring只为合格面试付费Pay only for qualified interviews供给侧（引擎/图形）= Unity、Epic Games、Roblox、NVIDIA 图形/Omniverse/Isaac 线；需求侧（世界模型/视频/机器人）= World Labs、Luma AI、Runway、OpenAI Sora 线、DeepMind Genie 线、Physical Intelligence。地理 = 美国/英国/加拿大。NVIDIA 与 OpenAI/DeepMind 取与图形/世界模型相关的可识别子集，绝对数偏保守。Supply side (engine/graphics) = Unity, Epic Games, Roblox, and the NVIDIA graphics/Omniverse/Isaac lines; demand side (world model/video/robotics) = World Labs, Luma AI, Runway, the OpenAI Sora line, the DeepMind Genie line, and Physical Intelligence. Geography = US / UK / Canada. For NVIDIA and OpenAI/DeepMind we take the identifiable subset related to graphics / world models, so the absolute numbers run conservative.
图形人才 = 图形与渲染 / 物理与仿真 / 引擎与工具链 / 3D 视觉与几何 四类可迁移职能（基于 title/headline 关键词判定）。走廊 = 需求侧现任技术人才中，历史雇主含游戏工作室、引擎三杰或图形/视效大厂者。走廊为可见样本，方向性指标。Graphics talent = the four transferable functions of graphics & rendering / physics & simulation / engine & toolchain / 3D vision & geometry (classified from title/headline keywords). Corridor = demand-side current technical staff whose past employers include a game studio, an engine big-three company, or a major graphics/VFX company. The corridor is a visible sample and a directional indicator.
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。重点人选已逐人对照公开信息复核。Five-signal cross-validation (name-ethnicity model / Chinese characters / Chinese text / Chinese institution / multi-spelling surname library), graded high/medium confidence, with the main scope = high + medium. Key candidates have been re-checked one by one against public information.
数据截至 2026 年上半年；游戏与 AI 人才流动快、档案更新有滞后；重点人选与点名案例以 2025-2026 公开信源为准标注。研究底稿 research_gamevideo.md / research_cn.md 同目录交付。Data through H1 2026; gaming and AI talent moves fast and profile updates lag; key candidates and named cases are annotated against 2025-2026 public sources. The research backups research_gamevideo.md / research_cn.md are delivered in the same directory.
| 公司Company | 侧Side | 在职画像Current profiles | 图形人才Graphics talent | 图形占比Graphics share | 华人Chinese | 华人占比Chinese share |
|---|---|---|---|---|---|---|
| Roblox | 供给Supply | 3,158 | 66 | 2.1% | 723 | 22.9% |
| Unity | 供给Supply | 682 | 54 | 7.9% | 49 | 7.2% |
| Epic Games | 供给Supply | 623 | 137 | 22.0% | 20 | 3.2% |
| NVIDIA (图形/仿真)NVIDIA (graphics/simulation) | 供给Supply | 262 | 229 | 87.4% | 35 | 13.4% |
| Luma AI | 需求Demand | 209 | 6 | 2.9% | 0 | 0.0% |
| DeepMind (Genie 线)DeepMind (Genie line) | 需求Demand | 159 | 14 | 8.8% | 26 | 16.4% |
| OpenAI (Sora 线)OpenAI (Sora line) | 需求Demand | 111 | 15 | 13.5% | 16 | 14.4% |
| Runway | 需求Demand | 91 | 1 | 1.1% | 0 | 0.0% |
| Physical Intelligence | 需求Demand | 26 | 0 | 0.0% | 8 | 30.8% |
| World Labs | 需求Demand | 22 | 3 | 13.6% | 8 | 36.4% |
① 覆盖率Coverage：本报告基于公开职业档案聚合，规模型公司覆盖度更高；NVIDIA/OpenAI/DeepMind 取相关子集，绝对数偏保守，各项为数据库口径。: this report aggregates public career profiles, with higher coverage of larger companies; for NVIDIA/OpenAI/DeepMind we take the relevant subset, so absolute numbers run conservative, and every figure is on a database basis.
② 职能判定Function classification：基于 title/headline 关键词，「图形人才」为可迁移职能的近似，个别 title 模糊者可能误分。: based on title/headline keywords, "graphics talent" is an approximation of transferable functions, and the occasional ambiguous title may be misclassified.
③ 走廊为可见样本The corridor is a visible sample：需求侧公司体量小（多为数十至数百人），走廊绝对数不大、为方向性指标；直接引擎三杰→世界模型流动尤其稀疏，广义图形/视效→世界模型走廊更稳健。: demand-side companies are small (mostly tens to hundreds of people), so the corridor's absolute numbers are modest and serve as a directional indicator; direct engine-big-three → world-model movement is especially sparse, while the broader graphics/VFX → world-model corridor is more robust.
④ 华人识别为概率判定Chinese identification is probabilistic，使用西文名且无中国信号的华裔会漏检。, so people of Chinese descent who use a Western name and show no Chinese signals will be missed.