Gross revenue, however, does not translate directly into revenue retained by Micro1. Contractors typically keep about 60% to 70% of revenue in this type of labeling business, putting Micro1's net annual run rate between $150 million and $200 million.
Micro1 remains smaller than some of its competitors. Mercor reached $2 billion in gross annualized revenue this summer, while Handshake crossed $1 billion earlier this year. Their growth alongside Micro1 points to substantial spending by AI labs seeking data created or evaluated by people with expertise in specific fields.
Micro1 is also trying to change the economics of its business through synthetic and reusable datasets. The startup can automatically generate descriptions of video content without relying on human workers for the process, while some prebuilt datasets can be licensed to more than one customer. Those reusable datasets can produce gross margins of 80% to 90%, substantially different from the contractor-heavy portion of the operation.
The company is extending that work into robotics as well. Micro1 is developing a pre-training dataset by paying hundreds of people to record interactions with everyday objects in their homes, creating material that can be used for robotics training.
Micro1 emerged from a different business model. Founder Ali Ansari initially built the company as an AI recruiting platform before seeing data-labeling customers use its system to identify and hire annotation engineers. The company subsequently shifted its focus toward the training-data market. That market has also created questions about who can buy reusable datasets. Because the same dataset can potentially be licensed to several customers, critics have raised concerns about American data providers supplying material that could also help foreign AI developers. Chinese model Kimi K3 has become part of that debate.
Ansari publicly criticized companies that sell training data to foreign adversaries, writing on X last month: “Some human data companies work with foreign adversaries. nd the results show today in Kimi K3. We believe it's shameful to claim American AI dominance desires while selling millions worth of data to countries that we are in adversarial competition with.”
Micro1 is positioning itself against that practice as American AI labs seek greater control over where the data they purchase can ultimately be used. That stance could differentiate the company with U.S. customers, though limiting buyers could also constrain the potential market for reusable datasets, which are among its higher-margin products.
Contract values are rising as well, according to a person familiar with Micro1's finances. The increase comes as researchers debate whether future spending on data could eventually approach the amount devoted to computing resources, potentially expanding the opportunity for companies supplying specialized training material.
Micro1's financing has begun to reflect its rapid revenue growth. The startup raised a Series A at a $500 million valuation last September and may have recently completed another funding round at a higher valuation. Micro1 did not respond to a request for comment about the potential financing.
The company's next challenge is turning its rapid gross revenue expansion into a more durable business. Its contractor model leaves Micro1 with a considerably smaller net run rate, while greater use of synthetic and reusable data could improve margins. At the same time, those higher-margin datasets introduce questions about how broadly the company is willing to sell them.
For now, the jump from $100 million to $500 million in eight months shows how quickly spending on training data has expanded. Micro1, Mercor and Handshake are each building sizable businesses around the human expertise, model evaluation and specialized datasets required to develop and improve AI systems, making data an increasingly significant component of the infrastructure behind model development.
This analysis is based on reporting from the tech buzz.
Image courtesy of Micro1.
This article was generated with AI assistance and reviewed for accuracy and quality.
