18个月，70倍。
这不是加密货币的涨幅，是一家AI数据标注公司的收入增速。8月20日TechCrunch援引知情人士消息：AI训练数据公司Micro1的毛年化收入（gross annual run rate）在过去8个月从1亿美元飙升到5亿美元。把时钟再往前拨一点，2025年初这家公司的ARR才约700万美元。18个月70倍——这个增速在AI行业里也是罕见的。
大模型公司在前线烧钱打仗——OpenAI一季度亏几十亿，Anthropic虽然盈利但利润微薄——而给它们提供"弹药"的数据标注公司，正在闷声发大财。
从AI招聘到数据标注：一场完美的 pivot
Micro1的故事颇有戏剧性。创始人Ali Ansari四年前创办这家公司时，做的是AI驱动的技术招聘——跟HackerRank、Mercor类似，帮科技公司筛选工程师。但Ansari很快发现了一个奇怪的现象：客户买他的招聘平台，不是为了招人，而是为了把招来的工程师派去做数据标注——给AI模型输出做评估、打分、打标签。
2025年初，Ansari做了一个果断的决定：把公司从AI招聘彻底转型为AI数据标注。这个pivot踩中了时间点——2025-2026年正是RLHF（人类反馈强化学习）和模型评测需求爆发的时期，GPT-5、Claude Opus、Gemini等前沿模型对高质量人工标注数据的需求近乎无底洞。
Micro1的模式跟传统数据标注公司不同：它不是雇一群低收入工人在平台上做简单的图片分类，而是签约各领域的专家——医生、律师、科学家、工程师——按项目制为AI模型提供领域专业级别的标注和评估。这种"RLHF健身房"（reinforcement learning gyms）模式的客单价远高于普通标注，一个专家小时的费用可能是普通标注员的几十倍。
但真正让利润起飞的是合成数据。Micro1越来越多地用AI自动生成数据（比如视频内容的自动描述），这些数据不需要按小时付费给人类，毛利率高达80%-90%。更妙的是，同一份现成数据集可以卖给多个客户——这是"数据石油"的真正含义：挖一次，卖多次。
卖铲子的人永远不亏
Micro1不是孤例。竞争对手Mercor今年夏天毛收入达到了20亿美元run rate，Handshake年初达到10亿。数据标注赛道整体在爆发，有研究者甚至预测未来AI在数据上的支出可能匹敌在算力上的支出。
但在这一片繁荣背后，有两个值得注意的暗面。
第一，5亿是毛收入，不是净收入。Micro1要付给签约专家60%-70%的收入作为酬劳，净run rate在1.5-2亿美元之间。虽然这仍然是惊人的增速，但不像5亿这个数字看起来那么夸张。真正的高利润在合成数据和多售数据部分——这部分毛利80-90%，但目前占比还不明确。
第二，地缘政治阴影。Ansari上个月在X上公开宣称Micro1不向中国模型制造商出售数据，暗指竞争对手在做这件事。"有些人类数据公司在跟外国对手合作，结果你们今天在Kimi K3上看到了。"这番话把数据贸易跟地缘政治绑在了一起——当美国AI公司的训练数据被出售给中国模型公司，这算不算资敌？这个问题会越来越尖锐。
"大模型公司在前面烧钱拼模型能力，数据公司在后面卖铲子收钱——这像极了淘金热里的牛仔裤生意。"—— 一位AI投资人的观察
不管怎样，Micro1的增速是一个明确的信号：AI产业链上最先赚钱的，不一定是做模型的，往往是做基础设施的。从算力（Nvidia）到数据（Micro1、Mercor）到路由（Router.com、OpenRouter），AI淘金热里卖铲子的人，永远比淘金的人先赚到钱。
对AI创业者来说这也是启发：别老想着做下一个OpenAI，想想OpenAI和它的竞争对手们必须买什么——那才是稳赚不赔的生意。
明天见。
Seventy-fold in 18 months.
That's not a crypto pump — it's an AI data labeling company's revenue trajectory. On August 20, TechCrunch reported, citing a person familiar with the company, that AI training data firm Micro1 saw its gross annual run rate surge from $100 million to $500 million over the past eight months. Rewind to early 2025, and the company's ARR was roughly $7 million. Seventy-fold in 18 months — a growth rate nearly unheard of even in AI.
Frontier labs are burning cash on the front lines — OpenAI losing billions per quarter, Anthropic profitable but with thin margins — while the data labeling companies supplying their "ammunition" quietly print money.
From AI Recruiting to Data Labeling: A Perfect Pivot
Micro1's origin story is pure startup drama. When founder Ali Ansari launched the company four years ago, it was an AI-powered technical recruiting platform — similar to HackerRank or Mercor, helping tech companies screen engineers. But Ansari quickly noticed something odd: customers were buying his recruiting platform not to hire people, but to put the recruited engineers to work labeling data — evaluating, scoring, and tagging AI model outputs.
In early 2025, Ansari made the decisive call: pivot the entire company from AI recruiting to AI data labeling. The pivot timed perfectly — 2025-2026 saw explosive demand for RLHF (Reinforcement Learning from Human Feedback) and model evaluation, as GPT-5, Claude Opus, Gemini, and other frontier models developed a nearly bottomless appetite for high-quality human-labeled data.
Micro1's model differs from traditional labeling shops. Instead of hiring low-wage workers for simple image classification, it contracts domain experts — doctors, lawyers, scientists, engineers — to provide domain-professional-level labeling and evaluation on a project basis. These "RLHF gyms" command much higher ticket sizes than basic labeling; an expert hour can cost dozens of times a standard labeler's rate.
But the real profit accelerator is synthetic data. Micro1 increasingly uses AI to auto-generate data (such as automated video content descriptions) with zero per-hour human costs, delivering gross margins of 80-90%. Even better, the same off-the-shelf dataset can be sold to multiple customers — that's the real meaning of "data oil": drill once, sell many times.
The Shovel Sellers Never Lose
Micro1 isn't alone. Competitor Mercor hit $2 billion in gross run rate this summer; Handshake reached $1 billion earlier this year. The entire data labeling sector is exploding, with some researchers hypothesizing that future AI spending on data could rival spending on compute.
But beneath the boom, two shadows loom.
First, $500M is gross, not net. Micro1 pays out 60-70% of revenue to contracted experts, putting net run rate at $150-200 million. That's still staggering growth, but less eye-popping than the $500M headline. The real margins are in synthetic data and multi-sold datasets, where gross margins hit 80-90% — though the current revenue mix is unclear.
Second, geopolitical shadows. Ansari publicly declared on X last month that Micro1 doesn't sell data to Chinese model makers, implying competitors do. "Some human data companies work with foreign adversaries. And the results show today in Kimi K3." The comment ties data trade to geopolitics — when American AI companies' training data gets sold to Chinese model firms, is that aiding a competitor? This question will only sharpen.
"Frontier labs burn cash competing on model capability while data companies collect on the sidelines selling shovels — it's the jeans-and-pickaxes business of the AI gold rush."— An AI investor's observation
Either way, Micro1's growth is a clear signal: the first companies to profit in the AI value chain aren't necessarily the model builders — they're often the infrastructure plays. From compute (Nvidia) to data (Micro1, Mercor) to routing (Router.com, OpenRouter), in every gold rush the shovel sellers get paid before the miners.
There's a lesson for AI entrepreneurs here: don't fixate on being the next OpenAI. Think about what OpenAI and its competitors absolutely must buy — that's the business with guaranteed demand.
See you tomorrow.
AI淘金热里最先赚钱的不是做模型的，是卖铲子的——数据标注正在复刻Nvidia的故事。
—— Dawn Vision编辑部
In the AI gold rush, the first money isn't made by model builders — it's made by shovel sellers. Data labeling is Nvidia's story, repeated.
— The Dawn Vision Editorial Desk
Micro1 · 5亿美元run rate · 18个月70倍 · 数据标注 · 合成数据 · 80%毛利 · Ali Ansari · Mercor · AI卖铲人 · RLHF
Micro1 · $500M run rate · 70x in 18 months · data labeling · synthetic data · 80% margins · Ali Ansari · Mercor · AI shovel sellers · RLHF
