The Unfakeable Layer

When generation gets cheap enough to be effectively free, what holds value starts to shift.

I spent two days at Startup Olé Marbella. Same venue, same pitch competition, same investors. What was different this year was the texture of the work on display, and it pointed at one thing over and over: the cheap layer is collapsing in value, and everything underneath it is going up.

Start with where this ends. Right now I talk to my agent. Soon my agent talks to your agent. Then your agent talks to you. Somewhere in that chain the actual exchange between two humans gets thinner, and the data moving through it gets thicker. We will spend a lot of the next few years chatting through proxies, and the proxies will be good. Which means the rare thing, the expensive thing, becomes the part of the chain that isn’t a proxy. A real conversation. A verified human. An idea that wasn’t interpolated from everything that came before it.

Marbella was a preview of that sorting. Here’s what I saw.

Timelines are compressed

There was a guy making a film about himself, alone, with the production value that used to need a crew and a budget. One person, one pipeline. A few years ago a single pilot episode was a fundraising event. Now it’s a weekend. That’s the good news and the trap in the same sentence: the cost of producing something that looks finished has gone to nearly zero. Looking finished no longer signals anything.

Polish stopped being proof

A lot of what gets pitched now is a half-polished demo built on Claude or Cursor. Not childish, but not valuable either, because anyone in the room could have generated the same thing on the train over. When the artifact is free, the artifact isn’t the proof. The proof is whether you can show the work behind it: the constraint you solved, the thing that broke, the measurement. No proof of work, no value. This is the whole reason I write down experiments before I run them.

The demo was never the company

Here’s what Marbella itself proves, just by existing. If software were the job, nobody would fly to the Costa del Sol to stand in a room for two days. But they do, because building was never the hard part. You still have to sell it, through people who look right, speak well, and show up in person. You still need the machinery underneath — compliance, data privacy, the regulated layer that takes years to earn trust in. A model writes a working app in a weekend. It does not write the reason a customer hands you their data. On privacy it’s actively worse than neutral: the easy path with these tools leaks data, so doing it right is a deliberate choice that cuts against the grain of the tool. The weekend demo is the cheap part. What makes it a company is the part a model can’t do for you.

Cybersecurity needs humans, more than ever

A team from Madrid pitched security and clearly knew their stuff. They made the final, and they made a point I keep coming back to: if there’s no red team, you’re not safe. Automated vulnerability scans are now cheap and everywhere, which is exactly why they’re no longer enough. The scan is the commodity layer. The human looking at the system, thinking like an attacker, is the part that can’t be automated away. That’s the unfakeable layer in security: not the detection, but the intent behind the probe. They’re growing, and they should be.

Selling AI at a fixed price is a bet on the layer you can’t see

One company was confident enough in its orchestration to sell AI at a fixed price. That’s a bold move when your own costs are variable and the models underneath you change monthly. It only works if the orchestration layer is genuinely a moat, not a wrapper. I don’t know if it pays off. But it’s the right kind of bet: pricing your own engineering, not the model’s output. Pricing the model’s output is a race to zero. Pricing the judgment on top of it is at least a bet worth making.

The most original thing I saw

A team wanted true answers to questions about their newborn. Important questions, the kind where being wrong matters. They asked frontier models and noticed the questions were fine but the answers diverged. So they asked all of them. Then they put the models in one room and let them argue with each other. Emergent behavior showed up, including the models describing themselves as confident parrots.

That was the most original idea at the event, and it’s original precisely because it doesn’t trust any single cheap answer. It treats disagreement between models as signal — and disagreement is the one thing a single cheap answer can’t produce. The founders want to build their R&D lab here in Spain. I hope they do — not in generating answers, but in knowing when to distrust them.

The pattern

None of these are AI stories exactly. They’re stories about what survives when AI makes the surface free. All of it — the compressed timelines, the commodity scans, the polished demos, the agent-mediated everything — pushes value down to the same place. Proof of work. A real conversation. An idea that’s actually yours.

The mistake is reading “AI builds it in a weekend” as “the thing has no value.” The demo was never the company. AI just stripped away the part that was hiding everything else — the selling, the trust, the compliance, the privacy you have to choose deliberately because the cheap path leaks. That was always the real work. Now it’s just more visible.

Cheap floods the zone. It always has. What changes is how much is suddenly cheap. The response isn’t to compete with the flood. It’s to be the thing the flood can’t produce.

(Two days, maybe a dozen conversations. These are observations, not conclusions. And yes, I used a frontier model to help shape this one. The thinking is mine. The help was rented.)