micro1 Pledges $1 Billion to Buy Real Company Data for AI Training, Financed by Citi and Hercules Capital

micro1, a data lab that supplies training data and evaluations to AI companies, announced on Thursday a commitment to spend $1 billion over the next 12 months acquiring and licensing enterprise operational data, financed by capital provided by Citi and Hercules Capital. The buying spree runs through its Company Data Partnerships program, which the company positions as a new revenue stream for businesses sitting on years of accumulated operational records.

The point of the purchases, per the announcement, is not to hoard data but to build reinforcement-learning environments that mirror how real companies actually run: incomplete information, competing priorities, edge cases that demand judgment. Models and agents trained inside those environments, the company argues, learn to navigate workflows, make decisions and complete complex tasks in a way that generic internet text cannot teach.

The announcement is heavy on ambition and light on mechanics. micro1 did not disclose how many partnerships it has signed, what the data costs per deal, or how the de-identification process is audited. The $1 billion figure is a commitment financed by lenders rather than revenue, which makes it a bet on demand: the program only pays off if AI labs keep paying for licensed operational data at scale.

micro1 has quietly become one of the busier suppliers in that market. A Forbes profile in September reported the company had raised more than $100 million at an estimated $4 billion valuation, with annualized revenue that had grown from roughly $7 million to more than $500 million and customers including frontier labs, Microsoft, Amazon and the robotics firm 1X — all figures from that report and largely company-sourced, not independently audited.

The move sharpens a split in the data economy. One camp pays experts to demonstrate knowledge — the math proofs, the code reviews, the medical reasoning. micro1 is betting the scarcer asset is process: how a logistics firm resolves a disputed shipment, how a finance team closes a quarter, how a support organization escalates a failure. If agentic AI is the product, its training set is arguably the enterprise itself.

For the enterprises, the offer is monetization of exhaust they already generate. For the industry, the program is another sign that the raw-web era of training data is ending, and the bidding is moving to proprietary, workflow-level data that no crawler can reach. Whether $1 billion a year is enough to matter — and whether the RL environments built from licensed data prove out — will show up in benchmarks and agent evaluations over the next year.