The Uber and Lyft Moment for AI Training
After $37k on Mercor in four months, my projects paused. Here is why I installed micro1, and the transferable skills to jump platforms without starting from zero.
Two weeks ago I opened the Mercor dashboard and stared at a screen that said, essentially, nothing. My active projects had paused. This was not a surprise, exactly. The AI training data market is more volatile than anyone talks about publicly. Projects get funded, data buckets fill up, evaluation cycles close, and suddenly the pipeline that was paying $100 an hour last week goes quiet. Then it reopens. Then it pauses again. If you are new to this world, the instinct is to assume something is wrong with you, with the platform, or with the entire industry. None of those conclusions are usually correct.
What actually happens is that the AI labs behind the scenes run internal experiments on a cadence that is completely disconnected from your personal income needs.
So I did what any rational operator does when one of their income streams temporarily dries up. I installed the second app.
A few years ago, every serious gig driver I knew had both Uber and Lyft open on their dashboards at the same time. Not because they disliked either platform, but because running on a single app is how you get quietly starved out. The AI training economy is reaching the same moment. If Mercor is the Uber of this space, with the broadest marketplace and the loudest IPO chatter, then micro1 is the Lyft. A credible second platform with its own philosophy, its own pay cycles, and its own approach to how it matches experts to AI labs.
Here is what I learned in the last two weeks.
A brief history of micro1
The company you know as micro1 was not originally called micro1. It was called Moontek, incorporated on September 1, 2019, in Los Angeles, by a founder named Ali Javid. For its first three years, Moontek was a boutique talent arbitrage firm. The business model was essentially to find elite technical talent in emerging markets who were priced below their US peers, and connect them to American startups. It was a service business, not a platform.
The pivot started quietly in June 2022, when the company released its first AI-powered technical assessment tool. By 2023, Moontek had been folded into a new identity called micro1, headquartered in Palo Alto. This is a pattern I find fascinating from a product perspective. The original talent network was the moat. The AI vetting tool was the scaling mechanism that turned a services business into software.
Today micro1 is run by Ali Ansari, a 24-year-old serial entrepreneur who is completing a part-time master’s at Stanford under the mentorship of Stefano Ermon, one of the most respected names in reinforcement learning. This detail matters more than it looks. Ansari’s academic connection to Ermon tells you exactly where the platform is going: RL environments, agentic workflows, verifiable reasoning tasks. The technical roadmap is not accidental.
The financials back this up. micro1 entered 2025 with roughly $7 million in ARR. By December it had crossed $100 million. The most recent funding round valued the company around $500 million, with chatter about the next round landing well above $2.5 billion. The backers include Founders Fund, 01 Advisors, and Jason Calacanis. The board includes Adam Bain, the former COO of Twitter.
None of this guarantees a great experience for contributors. But it does tell you that micro1 is not a weekend project. It has the capital, the customers, and the academic lineage to be a meaningful second platform for the foreseeable future.
Why I am not writing Mercor's obituary
I want to be clear about something before I go further. I am not writing Mercor’s obituary, and I do not think the playbook I shared in my previous article has expired. If anything, the thesis behind that playbook is getting stronger.
In March, Ansari and his chief economist Mark Esposito published a paper I would recommend every aspiring AI trainer read once. It is called “No Last Mile: A Theory of the Human Data Market.” The argument is simple. Human data is not temporary scaffolding that gets removed once the building is finished. It is a permanent maintenance layer. Models depreciate. Task contexts drift. Standards evolve. New edge cases emerge constantly. The paper’s conservative calibration predicts a steady-state labor share of five to seven percent for human-data work in the AI economy. Small as a percentage. Enormous as a dollar figure.
If you internalize this thesis, the Mercor pause stops feeling like an existential threat and starts feeling like a scheduling issue. Projects will resume. New ones will start. In the meantime, you do what any rational operator does and diversify your distribution.