20VC: Mercor CPO Oswald Nitski on AI Value, Data Moats, and Hiring

**Harry Stebbings** (0:01)
So, I've been thinking about the big question everyone's asking right now. If frontier models, open source and synthetic data keep evolving at this pace, who actually captures the value?
Harry Stebbings just talked with Oswald Nitski, Mercor's CPO, and Oswald's take is pretty nuanced.

**Oswald Nitski** (0:21)
Yeah, and he pushes back hard on this idea that open source is cannibalizing their business.
His argument is that data matters most at the edge of model capability. Customers aren't just buying random data sets, they're buying eval and training data to fill very specific gaps.

**Harry Stebbings** (0:40)
Right, he says open source models just raise the floor of what people are interested in.
So you get more baseline capability, but that actually creates more room for specialized demand. That's counterintuitive.

**Oswald Nitski** (0:54)
It is. And building on that point, Oswald thinks the market is seriously underestimating latent demand. He's not buying the claim that 90% of enterprise workflows can already be handled by existing models.
Instead, he points to long-horizon tasks, like automating an entire procurement team for months, as examples of work that's just beginning to emerge.

**Harry Stebbings** (1:19)
That connects to how enterprises actually behave with these tools.
Sensitive work flows, like generating legal memos, are still harder to place into frontier systems. But general tasks, like HR or procurement, feel more acceptable.

**Oswald Nitski** (1:34)
What's interesting is Oswald's point about open-weight models offering more control because inference can happen in multiple places.
Exactly. Privacy, utility, and risk. Companies are balancing all three. On the product strategy side, Oswald believes specialized models for every company are real. But every single one needs enterprise-specific eval training, data. He doesn't see a major ROI crisis yet because teams are still in exploration mode. That's fascinating. And he talks about how AI is changing product management itself. Engineering is less of a bottleneck now. So the real challenge is understanding workflows and choosing what drives revenue.
He identifies two big shifts for PMs, fewer tools to learn, and much more emphasis on business impact. Yeah, and his warning is blunt. Do not delegate your actual job to a model. He says AI can turn your brain into goop if you outsource judgment too far.
Mercor learned this lesson internally after trying to support too many workflows at once. The result was flexibility without clarity.

**Harry Stebbings** (2:45)
So now they focus on guardrails, simpler services, and pods that own distinct product areas. As engineering leverage rises, the PM to engineer ratio keeps changing.
What about his views on hiring and where the market's headed?

**Oswald Nitski** (3:01)
Mercor now favors senior candidates, AI fluency tests and whiteboarding over take-home work. They want people who can run experiments, read statistics, and make sound calls.
He's bullish on cyber, calling it deeply adversarial, where goalposts are always moving, and sees strong growth in environments, simulation data for agents, and robotics over the next three years. His final advice? Get a real internship fast, because school knowledge ages quickly, and the frontier never stops moving.