Micro1, a four-year-old AI data startup, has reached a $500 million gross annual run rate, up from $100 million just eight months ago, as demand for AI training data continues to surge. The company works with AI labs and corporations to provide training data using domain experts including doctors, lawyers and scientists.
The $500 million figure is gross annualized revenue, not net revenue or profit. Micro1 reportedly retains about 60% to 70% of the amount, putting its implied net annual run rate at roughly $150 million to $200 million. The company is also expanding beyond human-led data work into synthetic data, model evaluation and other forms of AI training infrastructure.
Micro1’s growth comes as AI companies increasingly pay for more than basic data labeling. They need specialized human expertise, model evaluations, synthetic datasets, agent-training environments and real-world data to improve increasingly capable models.
The company’s rapid expansion offers a useful view of how the AI industry’s data bottleneck is changing, from finding large amounts of labeled data to finding the right data, expertise and environments to train and evaluate advanced AI systems.
Micro1’s numbers show how quickly the market has accelerated
Micro1’s growth has been unusually fast even by AI startup standards. The company started 2025 with roughly $7 million in ARR, reached about $50 million by September, and crossed $100 million by December. It has now gone from that $100 million level to a $500 million gross annualized run rate in just eight months, according to TechCrunch.
Micro1 is not alone in seeing this kind of acceleration. Mercor crossed $2 billion in gross annualized revenue in June 2026, while rival Handshake reached roughly $1 billion earlier this year. That suggests the growth is not just a Micro1 story.
The broader AI training data industry is expanding rapidly as AI labs and enterprises spend more on expert data, evaluations and other inputs needed to train and improve increasingly capable models. Our team’s research also points to the industry having enough demand to support multiple large players at the same time.
The important shift is that these companies are no longer competing only in traditional data labeling. They are increasingly supplying specialized human expertise and other forms of high-value training data, turning what was once a relatively narrow annotation market into a much larger infrastructure layer for AI development.
AI training is expanding beyond LLMs
Micro1 is moving into areas where AI systems need to learn through interaction, not just process text or images.
The company is working on reinforcement learning environments, where AI agents perform tasks, receive feedback and improve through repeated attempts. This is becoming more important as AI shifts toward agents that can complete multi-step tasks.
Robotics creates a different data requirement. Robots need examples of how people physically interact with objects and environments, including the actions they take and the sequence in which they take them. Micro1 has been collecting this type of data by having hundreds of generalists record themselves interacting with everyday objects in their homes.
Micro1 is also putting more focus on AI evaluation, which the company describes as an important layer of proprietary intellectual property as companies build systems to continuously measure and improve model performance. Its analysis of why the AI evaluation layer could become a company’s real IP explains this shift in more detail.
Closing Lines
AI training is moving beyond datasets that teach models what to say toward data that shows how AI should act, how well it performs and how it should improve. That gives companies like Micro1 a much broader role across AI agents, robotics and other real-world AI systems.
