Mercor announced $2 billion in gross annualized revenue as of June.
$500M. Then $1B. Now $2B. All within 9 months.
ALL NEWSAI & STARTUPS
The trajectory is real and it is astonishing. But $2 billion is not Mercor's revenue.
What Mercor actually does
It's a two-sided marketplace for human expertise.
On one side: 30,000+ vetted specialists across dozens of countries — physicists, attorneys, doctors, bankers, engineers, competitive programmers. On the other: AI labs that need those people to answer hard questions, grade model outputs, and produce domain data no scrape can generate. Customers include OpenAI, Anthropic, Google DeepMind and Meta, plus newer labs like Reflection AI and Thinking Machines.
Mercor routes the work, vets the talent, and takes roughly a 35% cut of the payment volume passing through.
Which means the $2 billion is gross payment volume. Apply the take rate and Mercor's own net revenue is somewhere around $700 million annualised.
Still extraordinary for a three-year-old company. Roughly a third of the headline. If you're an investor reading a pitch deck this quarter and someone quotes "gross annualised revenue," that word is now doing about $1.3 billion of work.
Why this business exists at all
The public internet has been scraped. What frontier labs are short of isn't text — it's judgment.
You cannot download a securities lawyer's reasoning about an ambiguous disclosure. You cannot scrape how a cardiologist weighs two bad options. As models move from answering questions to doing work, the bottleneck stops being data volume and becomes the quality of expert reasoning captured at scale.
Mercor's insight was that this is a labour market problem wearing a data problem's clothes. Everyone else built annotation tooling. Mercor built recruiting infrastructure, then pointed it at AI labs. That's why a company originally pitched as AI hiring ended up as the picks-and-shovels layer of frontier model training.
The vacuum they walked into
Here's the part the growth chart doesn't show you, and it matters enormously.
In June 2025, Meta took a 49% non-voting stake in Scale AI at roughly a $29 billion valuation. Scale's founder Alexandr Wang left to run Meta Superintelligence Labs.
Then the business broke. Google moved to cut ties. OpenAI began phasing out its work with Scale. Not because the labelling got worse overnight — because Scale's product was never labelling. It was neutrality. No lab hands its model-development pipeline to a vendor part-owned by a direct competitor.
Mercor's hypergrowth is not purely demand creation. A meaningful share of it is inherited — customers who needed a data partner that wasn't standing next to Meta.
Which is exactly where the risk lives
So Mercor's real moat isn't its technology, its 30,000 experts, or its take rate. All of that is replicable with enough capital.
Its moat is that nobody owns them. And that is the rarest, most fragile kind of moat there is — because it can be destroyed by a single wire transfer, and the transfer would look like good news on the day it's announced.
Mercor has been in talks to raise $500 million at a $20 billion valuation, roughly double its $10 billion mark from eight months earlier. At that size, the natural next investors are strategic ones. For most companies, a hyperscaler on the cap table is validation. For Mercor, it's the Scale AI playbook running again.
Two more things I'd watch:
Concentration. Around 90% of revenue comes from foundation model companies — the most sophisticated buyers on earth, all of whom are vertically integrating everything they can. Several hyperscalers are already reported to be evaluating in-house data operations. Mercor's customers are its most credible future competitors.
Take-rate compression. A 35% cut on a $2 billion flow is an enormous target. When your top three customers represent most of your volume and each is spending hundreds of millions, they will eventually ask why the middle layer costs a third.
My read
This is one of the great growth stories of the AI cycle and I don't want to undersell it — $100M to $2B gross in fifteen months is not a fluke of accounting.
But the honest framing is narrower than the headline. Mercor sells the one input labs can't synthesise, to nine customers who would all prefer to own it, protected mainly by the fact that none of them can buy it without destroying it for the others.
That's a genuinely clever position. It's also a position that depends on Mercor's discipline about whose money it takes — which is a strange thing for a $20 billion valuation to rest on.
Question for you: is "we are the neutral party" a durable moat, or just a temporary one that lasts until the cheque is big enough? Scale AI's answer took about a year to arrive.
© 2026 Tech Produit. All rights reserved.