Micro1, a Scale AI competitor, touts crossing $100M ARR

Micro1’s meteoric rise over the past two years has placed it in a cohort of AI companies that are scaling at breakneck speed. The three-year-old startup, which helps AI labs recruit and manage human experts for training data, started the year with about $7 million in annual recurring revenue (ARR).

Today, the company claims to have surpassed $100 million in ARR, founder and CEO Ali Ansari told TechCrunch. That figure is also more than double the revenue Micro1 reported in September when it announced its $35 million Series A at a $500 million valuation.

Ansari, 24, then said that Micro1 is working with leading AI labs, including Microsoft, as well as Fortune 100 companies that are racing to improve large language models through post-training and reinforcement learning. Their demand for high-quality human data has fueled a rapidly growing market that Ansari predicts will grow from $10 to $15 billion today to nearly $100 billion within two years.

The rise of Micro1, and that of larger competitors such as Mercor and Surge, accelerated after OpenAI and Google DeepMind reportedly cut ties with Scale AI following Meta’s $14 billion investment in the supplier and its decision to hire Scale’s CEO.

Although Micro1’s ARR is growing rapidly, according to its founder, it has not yet matched its rivals: Mercor’s more than $450 million, sources told TechCrunch, and Surge’s reported $1.2 billion by 2024.

Ansari attributes Micro1’s growth to its ability to quickly recruit and evaluate domain experts. Like Mercor, Micro1 started as an AI recruiter called Zara, matching tech talent with software roles before moving into the data training market. That tool now interviews and vets applicants looking for an expert role on the platform.

In addition to providing expert-level data to leading AI labs, Ansari says two new segments, barely visible today, are on track to reshape the economics of human data.

The first involves non-AI-native Fortune 1000 companies that will build AI agents for internal workflows, support operations, finance, and industry-specific tasks.

Developing these agents requires systematic evaluation: testing frontier models, assessing their output, choosing winners, refining them, and continuously validating their performance in production. Ansari argues that this cycle relies heavily on human experts evaluating AI behavior at scale.

The second is pre-training in robotics, which requires high-quality, human-generated demonstrations of everyday physical tasks. Micro1 is already building what Ansari calls the world’s largest robotics pre-training dataset, collecting demonstrations from hundreds of generalists recording object interactions in their homes. Robotics companies will need vast amounts of this data before their systems can function reliably in homes and offices, he said.

“We expect a large portion of product budgets at non-AI native companies will go toward evaluations and human data, from 0% to at least 25% of product budgets,” said the CEO, who founded Micro1 while at UC Berkeley. “We’re also helping robotics labs create robotics data; these two areas will account for a huge share of that $100 billion per year market.”

Even as new markets emerge, Micro1’s current growth still comes primarily from elite AI labs and AI-heavy enterprises. The startup is expanding its work with these labs on reinforcement learning, the feedback loop to test and improve model behavior.

Micro1 hopes its early move into robotics data and enterprise agent development, in addition to scaling its specialized RL environments, will help it capture additional market share as the data wars heat up.

For now, Ansari says the company is focused on scaling responsibly, paying experts well and keeping people at the center of an industry built on training machines.

The company currently manages thousands of experts in hundreds of domains, ranging from highly technical fields to surprisingly offline disciplines. Many earn nearly $100 an hour, according to Ansari.

“There are Harvard professors and Stanford PhDs who spend half their week training AI through Micro1,” says Ansari. “But the bigger shift is in the volume and scope of roles. It’s expanding into areas you wouldn’t expect to matter for language model training, including offline and less technical fields. We’re very optimistic about where this is going.”