I was looking through Mercor this week and there is something I found much more interesting than the valuation, the founders being incredibly young or even how fast the company has grown.

Look at how they described themselves in January 2024:

“Redefining Hiring With AI.”

Now look at the first sentence on their website:

“Mercor is organizing human intelligence to power the AI economy.”

That's quite a change 😂

And I think there is a really good founder lesson hidden inside it.

Because Mercor didn't start as an AI training data company.

It started much closer to what we would call a recruiting marketplace.

The original idea was to connect engineers, particularly in India, with US companies looking for talent.

Then they built technology around the problem.

Search across resumes, GitHub profiles, portfolios and interview transcripts.

The pitch was basically:

There are millions of people out there. Companies are terrible at figuring out who is actually good. Let's use AI to make that matching much better.

Which is already a very big idea.

But then something interesting happened.

Some of the most advanced AI companies in the world started needing people.

Just not in the way Mercor originally imagined.

There is a funny contradiction happening in AI right now.

The more capable the models become, the more valuable certain humans become.

A frontier model doesn't just need someone clicking:

good answer / bad answer.

If you want a model to understand investment banking, you need someone who understands investment banking.

If you want it doing sophisticated legal work, you need lawyers.

Medicine? Doctors.

Software engineering? Great engineers.

Consulting? People who actually know what a good consulting deliverable looks like.

The training problem was moving from basic data labeling toward something much more complicated:

human expertise.

Mercor already had infrastructure designed to find, interview, assess and match people.

And suddenly some extremely rich customers had an enormous need for exactly that capability.

Brendan Foody has said that after meeting with OpenAI early on, they realized the human data market itself was changing.

The old model was crowdsourcing huge numbers of people for relatively simple tasks.

The new model needed specialists.

People with real domain knowledge who could teach models how professionals actually think and work.

And Mercor was weirdly well positioned for that.

Not because this was necessarily what they originally set out to build.

Because of what they had already learned how to do.

Find the right human for a very specific problem.

That turned out to be much more valuable than simply filling another software engineering role.

If you told the Mercor founders in 2023:

"You are building infrastructure for sourcing doctors, lawyers, bankers, scientists and engineers so they can teach frontier AI models how to perform professional work"

I'm not sure that would have been the pitch.

They were matching engineers with companies.

Then they were building an AI recruiting platform.

Then AI labs started becoming increasingly important customers.

By February 2025, most of Mercor's reported revenue was already coming from AI labs.

Later that year it was increasingly being described not as a recruiting startup, but as the company connecting AI labs with domain experts.

And then the numbers became a bit crazy.

Mercor said it went from basically zero to a $500M annualized revenue run rate in 17 months.

Earlier this year, it said it crossed $1B annualized.

Today its website says it is paying more than $4M per day to people in its network.

There are a million things you can debate about those numbers, how this market develops, margins, revenue concentration, how long AI labs keep spending at this level etc.

But I think focusing only on the growth misses the more interesting thing.

The market told them what their valuable capability actually was.

And they listened.

This is something I see founders struggle with all the time.

We get attached to the noun.

Then customers start behaving in a way that doesn't quite fit the noun.

And instead of getting curious, we try to push them back toward the business we already decided we were building.

But customers don't care about your category.

They care about getting their problem solved.

Mercor thought they were building better recruiting infrastructure.

Their AI customers basically said:

And suddenly the thing underneath recruiting — identifying who actually knows what — starts looking more important than recruiting itself.

That distinction is really interesting to me.

Because the visible product was hiring.

The underlying asset was understanding human capability.

Those are not the same business.

This is where I think the story gets more interesting than a normal "startup pivot" story.

They didn't build a recruiting product, realize nobody wanted it and throw everything away.

Almost the opposite.

The original capability became more useful.

The AI interviewer mattered.

The candidate network mattered.

The matching mattered.

The vetting mattered.

The ability to find someone with a ridiculously specific combination of expertise mattered.

They just found a customer willing to pay much more for the output.

Mercor itself has described contracting experts for AI training as a kind of forcing function for its broader talent ambitions.

And this makes sense.

Traditional recruiting has a terrible feedback loop.

You hire someone.

Then maybe three months later you get some vague idea whether they were actually good.

With these AI projects, Mercor can potentially see much faster whether the person it matched actually performed well on the work.

That creates data.

That data makes the matching better.

Better matching attracts more customers and experts.

And suddenly the "pivot" is actually making the original technology stronger.

I really like this because pivots are usually explained as:

We were doing A. It failed. So we started doing B.

Reality can be much messier.

Sometimes A works.

But doing A reveals B.

And B is 20x more valuable.

Obviously this one caught my attention because I spend a lot of my life thinking about recruiting.

At HiresLink we help companies hire talent across Latin America.

The obvious product is the placement.

Company needs someone → we find the right person → they hire them.

But after doing this enough times you realize every search also creates information.

What are companies suddenly hiring for?

Which skills are getting harder to find?

What are candidates actually asking to be paid?

Which countries have deeper supply for certain roles?

Where are budgets moving up?

How long does it actually take to fill something very specific?

We started turning more of that into our LATAM Talent Intelligence Report, which now covers 90,000+ pre-vetted candidates, 280+ placements, 77 roles and 8 countries.

And when you actually look inside the data, you start seeing things that the generic term "AI talent" completely hides.

For example, our current report shows roughly 9,500 specialized data annotators.

Machine learning engineers? Around 2,400.

AI product managers? Around 1,800.

NLP engineers? Around 900.

MLOps engineers? Around 600.

LLM / RLHF engineers? Around 350.

Same broad “AI talent” category.

Completely different supply dynamics.

And the salary spread tells the same story.

Our current senior LATAM benchmarks range from around $2,200/month for specialized data annotators to $8,500 for ML engineers, $8,800 for NLP engineers, $8,200 for MLOps and $9,200 for LLM / RLHF engineers. 

We publish the underlying LATAM salary and role benchmarks here.

That information exists because of the recruiting activity.

Which is what made the Mercor story click for me.

The placement might be the thing a customer pays you for today.

But the patterns created across hundreds or thousands of those transactions can become another asset entirely.

Mercor is obviously doing something at a completely different scale and in a different part of the market.

But the underlying question is similar:

What valuable thing is your company producing as a side effect of doing the thing customers already pay you for?

Sometimes that side effect might eventually be more interesting than the original product.

I'm also not completely convinced the lesson is simply:

"Follow whatever is growing fastest."

AI labs are spending enormous amounts of money right now.

That creates strange markets.

When a small number of extremely well-funded companies urgently need something, businesses can appear enormous very quickly.

And Mercor itself seems very aware that simply supplying experts forever isn't enough.

It's moving deeper into evaluations, benchmarks, reinforcement-learning environments and enterprise AI infrastructure.

Which to me is another signal.

The customer pull showed them where the money was.

Now they have to figure out which parts of that demand become durable infrastructure.

Those are two different problems.

You can follow the market into a huge opportunity and still eventually discover that the first version of that opportunity was temporary.

So I wouldn't take the Mercor lesson as:

Pivot every time somebody offers you more money 😂

I think it's more subtle.

Pay attention when customers repeatedly use your product in a way you didn't design it for.

Pay attention when one customer segment has a completely different willingness to pay.

Pay attention when the thing they value isn't the feature you keep talking about in your pitch deck.

And especially pay attention when solving their problem makes your underlying product stronger.

Those are signals.

Imagine removing the name of your company and everything you say you do from your website.

Then look only at customer behavior.

What are they actually paying you for?

What do they keep coming back for?

What do they ask you to do that isn't quite part of the original product?

What information, infrastructure, network or capability are you accumulating while serving them?

There might be another company hiding inside your company.

And your customers may already know it's there before you do.

Mercor started out trying to make recruiting better with AI.

A few years later, it's using recruiting infrastructure to organize the humans teaching AI how to work.

That's a much stranger company.

And maybe a much bigger one.

See you next week,

Julian

PS — I'm curious about this one. Have you ever had customers start pulling your company in a direction you didn't expect? Did you follow them or stick to the original plan? Reply and tell me — I read all of these.