Mercor’s $20bn Price Tag on Human Judgement

Posted by & filed under , .

A balance scale weighing a laptop against a human brain, with a hazy city skyline behind it.Mercor raised $750 million last week at a $20 billion valuation, double what it was valued at in October 2025. If you haven’t come across them: Mercor staffs domain experts (doctors, lawyers, engineers, bankers and the likes) to build evals and training data for the frontier AI labs. Its CEO says it went from a $1 million to a $3 billion revenue run rate in 27 months. Twenty-seven months! Most side projects take longer than that to get past the landing page 🙂

Now, the funding round itself isn’t what interests me. What interests me is what the number says about where the value sits in AI right now, and what that means for the way we lead engineering teams.

The Scarce Thing Isn’t Compute

Here’s the thing: everybody talks about GPUs and data centres, and yes, those cost a fortune. But the labs, which have more compute than anybody, are paying billions a year for something else: people who know what a good answer looks like in their field, and can write it down in a way a machine can be tested against. Knowledge you can buy by the hour, apparently, is cheap. Judgement, written down as an eval, is what’s expensive.

And if that’s true for the labs, I’d argue it’s even more true for your engineering organisation.

Your Senior Engineers Are Your Eval Set

I’m convinced the future of software engineering is a “lights out” setup: we “code” through specs, and agents write, test and ship the code. We’re not fully there yet, but every month we get a bit closer. In that world, what’s the most valuable thing your senior engineers do? It’s not typing code. It’s knowing which of the three plausible solutions will fall over in production at 3am, knowing which customer will scream when you change that API, and knowing what “done” actually means for this particular system. That’s exactly the kind of judgement Mercor is selling to the labs.

So the question I’d ask as a CTO is: how much of that judgement lives only in people’s heads? Because an agent can’t read heads (yet!). It can read specs, acceptance criteria, architecture decision records and test suites. Every time a senior engineer writes down a rule as a test, an eval or a crisp spec, they’re turning their judgement into something that scales across a hundred agent runs. Every time it stays in a code review comment that nobody reads again, it doesn’t.

What I’d Do on Monday

  • Make writing specs and evals a first-class part of senior engineers’ jobs, and recognise it in reviews and promotions, the same way we once recognised the person who wrote the clever code.
  • Turn recurring review comments into checks. If you’ve written “we don’t call that service synchronously” three times, it should be a test or a lint rule, not a fourth comment.
  • Hire for judgement. The interview question isn’t “can you implement this?” any more, the agent can. It’s “here are two implementations that both pass the tests: which one would you ship, and why?”
  • Don’t stop hiring juniors. Judgement is learnt, and somebody has to be your senior engineers in ten years’ time.

Bottom line: the market just put a $20 billion price tag on human judgement, written down. You probably have a fair amount of it in your own team already. The question is whether it’s written down anywhere an agent can find it.