01核心结论Key Takeaways
以下数字为 Metix AI 数据库口径（数据时点约 2026 H1），统计对象 = 15 家美国 AI 算力 / 芯片公司的在册工程与研究人才（按权威 company_id 取在职、再按"硬件设计 + ML 系统"岗位口径清洗）。所有数字均为聚合统计，报告不展示任何个人信息。The figures below follow Metix AI database methodology (data as of roughly 2026 H1). The population = current engineering and research professionals at 15 U.S. AI-compute / chip companies (current employees pulled by authoritative company_id, then cleaned to a "hardware design + ML systems" role scope). All figures are aggregate statistics; the report displays no personal information.
同时懂物理硅（RTL / 验证 / 物理设计）与 ML 系统（CUDA / 编译器 / 内核）的人，是这场竞赛的硬通货。全图 56,861 人里仅People who understand both physical silicon (RTL / verification / physical design) and ML systems (CUDA / compilers / kernels) are the hard currency of this race. Of the 56,861 across the full map, only 8.4% 可识别为双栈，NVIDIA 也只有 are identifiable as dual-stack, and even NVIDIA has only 14.6%（约 1/7），传统芯片大厂近乎绝迹（博通 2.2%、英特尔 5.0%）。双栈密度，是一家公司"有多 AI 原生"的体温计。 (~1/7); at legacy chip giants they're nearly extinct (Broadcom 2.2%, Intel 5.0%). Dual-stack density is the thermometer of how AI-native a company really is.
它是 14 家公司里It is, among the 14 other companies, the #1 talent feeder for 9 家9 of them的头号人才来源——NVIDIA 的 13%、AMD 的 21%、Rivos 的 32%、Tenstorrent 的 31% 都来自英特尔，且其中约 — 13% of NVIDIA, 21% of AMD, 32% of Rivos and 31% of Tenstorrent come from Intel, and of those roughly 89% 在英特尔做的就是技术岗。算力竞赛的地基，很大程度上是英特尔流出的资深工程师在搭。 held technical roles at Intel. The foundation of the compute race is, to a large degree, being laid by senior engineers who flowed out of Intel.
大量进人、极少流出。在别家的人才来源里 NVIDIA 普遍只占 5–9%（唯一例外是 Groq，34%）。Massive inflow, minimal outflow. As a talent source for other companies, NVIDIA generally accounts for only 5–9% (the lone exception is Groq, at 34%).40 个月40 months中位任期 + NVDA 约十倍的股价 = 行业最硬的金手铐，把Median tenure + a roughly 10x move in NVDA stock = the hardest golden handcuffs in the industry, locking the 44% 的"4 年以上老兵"锁在原地。 of "4-years-plus veterans" in place.
可直接拿来源人：SambaNova = 甲骨文 / Sun（41%）、Rivos = 苹果芯片帮（被苹果起诉的那批）、Groq = 前 NVIDIA（34%）、AWS 自研芯片 Annapurna = 亚马逊内部转岗（54%）。读懂血统，就知道该去哪家挖哪种人。You can source straight from it: SambaNova = Oracle / Sun (41%), Rivos = the Apple-chip crew (the ones Apple sued), Groq = ex-NVIDIA (34%), AWS's in-house chip team Annapurna = internal Amazon transfers (54%). Read the lineage and you know which company to raid for which kind of engineer.
02存量地图：体量在大厂，密度在初创Stock Map: scale lives at the giants, density at the startups
先看盘子。这 15 家公司在美国的在册工程与研究人才共 56,861 人，但分布极不均：6 家芯片 / GPU 大厂占了 55,478 人（Start with the pie. These 15 companies hold 56,861 current engineering and research professionals in the U.S., but the distribution is wildly uneven: the 6 chip / GPU giants account for 55,478 (97.6%），8 家挑战者合计只有 1,290 人。最刺眼的反差是——人最多的英特尔（21,549），AI 双栈密度却几乎垫底（5.0%）。), while the 8 challengers combined hold just 1,290. The starkest contrast: Intel, with the most people (21,549), sits near the bottom on AI dual-stack density (5.0%).
把每家公司拆成"偏硅（metal）/ 偏模型（model）/ 双栈"三类，画像立刻清晰：传统芯片厂是纯硬件军团，AI 初创则把天平往模型一侧拉。下面是完整花名册（中位任期 / 中位履历计算到报告日期）。Split each company into three buckets — "metal-leaning / model-leaning / dual-stack" — and the picture snaps into focus: legacy chip makers are pure-hardware battalions, while AI startups tilt the scales toward the model side. Below is the full roster (median tenure / median career computed to the report date).
| 公司Company | 阵营Camp | 工程研究人才Eng & research | 偏硅%Metal % | 偏模型%Model % | 双栈%Dual-stack % | 中位任期Median tenure | 中位履历Median career | 博士%PhD % |
|---|---|---|---|---|---|---|---|---|
| NVIDIA | GPU 双雄GPU duo | 12,134 | 51.5 | 31.5 | 14.6 | 40 月40 mo | 15.0 年15.0 yr | 14.1 |
| AMD | GPU 双雄GPU duo | 6,139 | 67.2 | 16.7 | 11.7 | 36 月36 mo | 15.1 年15.1 yr | 10.8 |
| 英特尔Intel | 传统芯片Legacy chip | 21,549 | 49.9 | 8.7 | 5.0 | 68 月68 mo | 16.5 年16.5 yr | 20.1 |
| 高通Qualcomm | 传统芯片Legacy chip | 8,196 | 48.1 | 15.0 | 8.1 | 58 月58 mo | 15.8 年15.8 yr | 12.1 |
| 博通Broadcom | 传统芯片Legacy chip | 5,313 | 42.0 | 4.8 | 2.2 | 90 月90 mo | 21.8 年21.8 yr | 9.9 |
| Marvell | 传统芯片Legacy chip | 2,147 | 65.2 | 6.2 | 5.2 | 51 月51 mo | 19.6 年19.6 yr | 11.1 |
| SambaNova | 挑战者Challenger | 179 | 61.5 | 45.3 | 26.8 | 48 月48 mo | 14.7 年14.7 yr | 6.7 |
| Etched | 挑战者Challenger | 141 | 77.3 | 27.0 | 21.3 | 12 月12 mo | 12.5 年12.5 yr | 8.5 |
| Tenstorrent | 挑战者Challenger | 304 | 84.5 | 26.3 | 21.1 | 16 月16 mo | 14.9 年14.9 yr | 12.5 |
| Cerebras | 挑战者Challenger | 222 | 49.1 | 39.6 | 19.8 | 24 月24 mo | 15.4 年15.4 yr | 15.8 |
| d-Matrix | 挑战者Challenger | 88 | 68.2 | 59.1 | 40.9 | 18 月18 mo | 15.5 年15.5 yr | 22.7 |
| Lightmatter | 挑战者Challenger | 157 | 78.3 | 17.8 | 12.7 | 17 月17 mo | 15.4 年15.4 yr | 30.6 |
| Rivos | 挑战者Challenger | 140 | 87.1 | 10.0 | 9.3 | 40 月40 mo | 13.8 年13.8 yr | 8.6 |
| Groq | 挑战者Challenger | 59 | 47.5 | 33.9 | 20.3 | 30 月30 mo | 16.7 年16.7 yr | 6.8 |
| AWS Annapurna | 云自研Cloud in-house | 93 | 71.0 | 59.1 | 41.9 | 19 月19 mo | 10.4 年10.4 yr | 15.1 |
03双栈稀缺性：一道单调的"AI 原生"梯度Dual-Stack Scarcity: a monotonic "AI-native" gradient
把"双栈密度"按公司排开，会看到一条几乎完美单调的梯度：从只做网络 / 模拟芯片的博通（2.2%），一路爬到天生为大模型造芯片的 d-Matrix、AWS Annapurna（约 41%）。一家公司离大模型有多近，它的简历里就有多少"既懂硅又懂模型"的人。Line up "dual-stack density" by company and you get an almost perfectly monotonic gradient: from Broadcom (2.2%), which only does networking / analog chips, climbing all the way to d-Matrix and AWS Annapurna (~41%), built from the ground up to make chips for large models. The closer a company sits to large models, the more "understands both silicon and models" people its résumés contain.
各公司"双栈桥梁人"占比（硬件 ∩ ML 系统）Share of "dual-stack bridge people" by company (hardware ∩ ML systems)
三个梯队Three tiers
传统芯片厂Legacy chip makers 博通 / 英特尔 / Marvell / 高通 ≈Broadcom / Intel / Marvell / Qualcomm ≈ 2–8%。它们是纯硅军团（偏硅 42–65%，偏模型常个位数），双栈几乎绝迹。. They are pure-silicon battalions (metal-leaning 42–65%, model-leaning often single digits), with dual-stack all but extinct.
GPU 双雄GPU duo AMD / NVIDIA ≈ 12–15%。常年活在"GPU × ML"的交叉口，是大厂里双栈最厚的。. Living at the "GPU × ML" intersection for years, they carry the thickest dual-stack layer among the giants.
AI 加速器挑战者AI-accelerator challengers ≈ 20–42%。为大模型而生，把模型一侧的人才比例拉到大厂的两三倍。. Built for large models, they pull the model-side talent share to two or three times that of the giants.
这是个"下限"，但梯度是真的This is a "floor," but the gradient is real
双栈用关键词从自报头衔 + 技能里识别，Dual-stack is identified by keywords from self-reported titles + skills, so it 倾向低估tends to undercount（很多人没把全部技能写满）。但即便换成更严格的"物理硅 ∩ 模型"口径（把 CUDA / 编译器从金属侧剔除），梯度依然单调成立：博通 1.7% → NVIDIA 8.4% → d-Matrix 34%。 (many people never fill in all their skills). But even under a stricter "physical silicon ∩ model" scope (dropping CUDA / compilers from the metal side), the gradient still holds monotonically: Broadcom 1.7% → NVIDIA 8.4% → d-Matrix 34%.
换句话说：In other words: 绝对数会更高，相对排序不会变the absolute numbers go higher, the relative ranking doesn't move。谁更稀缺这件事，结论稳健。. On the question of who is scarcer, the conclusion is robust.
04英特尔工厂，与每家公司的"血统"The Intel Factory, and every company's "lineage"
人从哪来？答案出乎意料地一致：Where do the people come from? The answer is surprisingly consistent: 英特尔Intel。它是 14 家公司里 9 家的头号人才来源——而且越是硬核的硅片初创，英特尔的占比越高。AI 也许是英特尔输掉的战争，但这场战争的兵，很多是英特尔送出去的。. It is the #1 talent feeder for 9 of the other 14 companies — and the more hardcore the silicon startup, the higher Intel's share. AI may be the war Intel lost, but many of the soldiers fighting it were sent out by Intel.
各公司中"前英特尔"员工占比（占该公司工程研究人才）Share of "ex-Intel" employees by company (of that company's eng & research talent)
英特尔之外，剩下 5 家"非英特尔血统"的公司各有出处——而且每一条都能在公开记录里印证。读懂血统，就知道该去哪挖哪种人：Beyond Intel, the remaining 5 "non-Intel-lineage" companies each have their own origin — and every one can be verified in the public record. Read the lineage and you know where to raid for which kind of engineer:
| 公司Company | 头号血统Top lineage | 占比Share | 公开印证Public verification |
|---|---|---|---|
| Groq | 前 NVIDIAEx-NVIDIA | 34% | 创始人来自 Google 初代 TPU，但工程班底是 NVIDIA 系——"创始人血统 ≠ 队伍血统"。The founder came from Google's first-gen TPU, but the engineering bench is NVIDIA-bred — "founder lineage ≠ team lineage." |
| SambaNova | 甲骨文 + SunOracle + Sun | 24% + 17% | 联合创始人 Rodrigo Liang 出身 Oracle / Sun 的 SPARC 处理器谱系，整支队伍带数据库硬件基因。Co-founder Rodrigo Liang comes from the SPARC-processor lineage at Oracle / Sun, and the whole team carries database-hardware DNA. |
| Rivos | 苹果Apple | 20% | 2022 年被苹果起诉挖角芯片团队、窃取 SoC 机密，2024 年初和解——数据精确印证了那批苹果硅片人。Sued by Apple in 2022 for poaching its chip team and stealing SoC secrets, settled in early 2024 — the data precisely confirms that cohort of Apple-silicon people. |
| Etched | 苹果 + 英特尔Apple + Intel | 16% + 13% | 哈佛辍学生 2022 年创办、造 Transformer 专用 ASIC（Sohu），靠挖苹果 / 英特尔硅片老兵补齐硬件。Founded in 2022 by Harvard dropouts to build a Transformer-specific ASIC (Sohu), filling out its hardware bench by poaching Apple / Intel silicon veterans. |
| AWS Annapurna | 亚马逊内部Internal Amazon | 54% | Trainium / Inferentia 团队以内部转岗为主——云厂自研芯片是"内部造血"，不是市场挖角。The Trainium / Inferentia team is driven mainly by internal transfers — a cloud provider's in-house chips are "grown from within," not poached from the market. |
05NVIDIA 的金手铐：进得多，出得少NVIDIA's Golden Handcuffs: lots in, little out
英特尔的人四散而出，NVIDIA 却几乎不漏人——它在别家的人才来源里普遍只占 5–9%（唯一例外是 Groq）。原因写在任期里：NVIDIA 美国核心的中位在职任期是Intel's people scatter outward; NVIDIA barely leaks anyone — as a talent source for other companies it generally accounts for only 5–9% (the lone exception being Groq). The reason is written in the tenure: the median current tenure of NVIDIA's U.S. core is 40 个月40 months，, and 44% 的人已经待满 4 年。叠加 NVDA 在 2023–2025 年约十倍的涨幅，这是当下行业最硬的一副金手铐。 have already passed 4 years. Layer on NVDA's roughly 10x run across 2023–2025 — and these are the hardest golden handcuffs in the industry today.
NVIDIA 工程研究人才的在职任期分布（n≈11,894）Current-tenure distribution of NVIDIA's eng & research talent (n≈11,894)
可挖的，是 <24 个月那 31%What's poachable is the 31% under 24 months
约Roughly 3,700 人3,700 people入职不到两年，初始 RSU 远未归属完——跳槽放弃的账面收益最小，是 NVIDIA 阵营里最现实的挖角窗口。越往 48 个月以上走，金手铐越紧：那 44%（约 5,200 人）手里是已大幅 in-the-money 的归属股票，几乎撬不动。 have been on board under two years, their initial RSUs far from fully vested — they forfeit the least paper gains by jumping, making them the most realistic poaching window in the NVIDIA camp. The further you go past 48 months, the tighter the handcuffs: that 44% (~5,200 people) hold deeply in-the-money vested stock and are nearly impossible to pry loose.
初创的打法正好相反Startups play it exactly the other way
Etched（中位任期 12 月）、Tenstorrent（16 月）、Lightmatter（17 月）都很"新"——它们用Etched (median tenure 12 mo), Tenstorrent (16 mo) and Lightmatter (17 mo) are all very "young" — they use 未上市股权pre-IPO equity反向操作：趁人还没在大厂 vesting 满之前撬出来，用上市前的期权对赌，去换大厂那份已经兑现的确定性。 to run the reverse play: pry people out before they fully vest at a giant, betting pre-IPO options against the giant's already-realized certainty.
06猎人画像与实操指南Hunter Profile & Field Playbook
把上面的结构翻译成可执行的源人动作。三件事先记住：这是一场用Translate the structure above into executable sourcing moves. Remember three things first: this is a war fought with 老兵veterans打的仗、要按, you source by 偏硅 / 偏模型metal / model标签分赛道找人、时机决定能不能撬动。 tags split into lanes, and timing decides whether you can pry someone loose.
都是老兵，不是应届All veterans, no new grads
全图中位履历The full-map median career is 13–22 年13–22 years；连最年轻的 Etched 也有 12.5 年，博通甚至 21.8 年。算力战是用资深工程师打的——别拿"招应届"的预算和话术来打这个市场。; even the youngest, Etched, sits at 12.5 years, and Broadcom is as high as 21.8. The compute war is fought with senior engineers — don't bring a "new-grad hiring" budget and pitch to this market.
按 metal / model 标签源人Source by metal / model tags
要Want 纯硅片设计pure silicon design（RTL / 物理设计 / 验证）→ 去 Rivos（偏硅 87%）、Tenstorrent（85%）、Lightmatter（78%）、Etched（77%）。 (RTL / physical design / verification) → go to Rivos (metal-leaning 87%), Tenstorrent (85%), Lightmatter (78%), Etched (77%).
要Want 编译器 / ML 系统Compilers / ML systems→ 去 AWS Annapurna、d-Matrix（偏模型 59%）、SambaNova（45%）、Cerebras（40%）。 → go to AWS Annapurna, d-Matrix (model-leaning 59%), SambaNova (45%), Cerebras (40%).
博士浓度决定话术PhD density shapes the pitch
要研究底子（光子 / 模拟 / 编译）→ Lightmatter（博士 30.6%）、d-Matrix（22.7%）、英特尔（20.1%）。Want a research foundation (photonics / analog / compilers) → Lightmatter (PhD 30.6%), d-Matrix (22.7%), Intel (20.1%).
要工程交付型、不必要博士 → SambaNova（6.7%）、Groq（6.8%）、Etched（8.5%）。Want delivery-focused engineers where a PhD isn't necessary → SambaNova (6.7%), Groq (6.8%), Etched (8.5%).
想要这 56,861 人里的某一类——的完整名单？Want the full list of a specific slice of these 56,861 people?
本报告是聚合透视；落到具体人，Metix AI 可按"双栈 / 偏硅 / 偏模型 + 公司 + 任期窗口"筛出可对接的候选人名单，也可为任意目标公司生成同款人才 X 光。This report is an aggregate x-ray; to get down to named individuals, Metix AI can filter a contactable candidate list by "dual-stack / metal-leaning / model-leaning + company + tenure window," and can generate the same talent x-ray for any target company.