I saw this coming in 2018, and I moved my whole life because of it

Last week Anthropic published an economic model of what AI does to work between now and 2030. Three scenarios. I read all three, and then I read them again, because I have been waiting eight years for someone to put numbers on the thing I have been saying at dinner tables since 2018.

Here is what the model says.

In the modest scenario, AI ends up mattering about as much as the internet did. GDP grows 1.6%. Almost nobody’s life changes structurally.

In the substantial scenario — AI handling roughly half of knowledge work — GDP grows 8.3%. Unemployment goes to 4.6%. White-collar employment drops about 4%.

In the extreme scenario, GDP grows 32%, unemployment hits 12%, white-collar employment falls more than 20%, and unemployment among knowledge workers specifically reaches almost 18%.

But the number that actually matters is not in any of those headlines. It is this one: in every single scenario, income shifts away from workers and toward the people who own the assets. Labor’s share of the economy goes from 60% down to 56% in the middle case and down to 45% in the extreme one. Capital takes the difference. The economy gets bigger in all three, and in two of the three the people who work for a living get a thinner slice of a fatter pie.

And underneath that, the model shows a split that nobody wants to say out loud: knowledge workers get automated and displaced, while construction, nursing and the manual trades see demand and wages go up.

Now let me tell you where I was in 2018.

I was General Manager for Africa at RTB House, a programmatic real-time-bidding agency. At the time we were the biggest agency of our kind on the continent. Our client list was, roughly, the top five e-commerce companies in Africa, the top five online travel agencies, and the top five airlines. Before that I had built software for hotels, then e-commerce with Jumia, then HotelOnline. I had spent my entire career in software.

At RTB House we were using deep learning to predict customer behaviour. Not “AI” in the way people say it now. Neural networks, trained on behavioural data, deciding in under 100 milliseconds what to show a specific human being and how much to pay for the privilege. We were doing this at continental scale, every day, for years.

And here is the thing that changed my life: it was the same technology. The architecture running inside the models everyone is losing their mind about today is the direct descendant of what we had sitting in production in 2018. Different scale, different data, vastly more compute. Same fundamental machine.

So I had a very unusual seat. I got to watch, up close and daily, what these systems were actually good at. And what they were good at was not “creativity” or “consciousness” or any of the things people were arguing about on panels. What they were good at was this: taking a large amount of information, finding the pattern in it, and making a judgement call faster and cheaper than a person.

Which is, if you are honest about it, a fairly complete description of what most well-paid office jobs consist of.

That was the moment. Not a prediction about AGI. A much more boring observation: a machine does not need to be smart to take your income. It only needs to be cheaper than you at the specific task somebody is paying you to do. And I was watching it get cheaper every quarter.

So I made a decision that most of my peers thought was insane. I got out of software.

I went into hardware. In my case, that meant land and building. I moved into real estate development — acquiring land on the Samaná Peninsula in the Dominican Republic and building on it. From programmatic advertising to pouring concrete. People asked me if I was having a midlife crisis.

The logic was simple and it has not changed. A model can write the marketing plan for a building. It cannot pour the foundation. It cannot survey the parcel, clear the title, negotiate with the municipality, hire the crew, or hand the buyer a set of keys. The value in a physical asset is anchored to physics, to a specific location, and to legal ownership of a thing that exists in one place on Earth. You cannot inference your way into more beachfront.

And notice what happens in Anthropic’s model. In the scenario where knowledge-worker wages go flat or negative, wages for physical work go up. Because when the abundant thing gets cheap, the scarce thing gets expensive. If AI makes cognitive labour abundant, then everything that is not cognitive labour — land, materials, trades, energy, people who can actually build — gets repriced upward.

Two honest caveats, because I would rather be right than loud.

First, the extreme scenario is the one everyone will quote and it is also the least likely one. Anthropic’s own model puts Dario Amodei’s famous warnings about entry-level office jobs in that extreme bucket. The middle scenario is more plausible and it is much less dramatic. A 4% drop in white-collar employment is not the apocalypse. It is a bad decade for a lot of specific people, which is a different thing.

Second — and this cuts against me — the model deliberately excludes advances in robotics. If robots get good faster than expected, the trades are not the safe harbour I am describing. I do not think that happens on the same timeline, because moving atoms is harder than moving tokens, but I would be lying if I said it was not a risk to my own thesis.

What I will not walk back is the underlying point, and it is not really about AI at all.

If you make your money sitting in front of a computer as an employee, your income depends on a task staying expensive. That is the whole bet. And that bet is getting worse every year, not because anyone is coming for you, but because the cost curve on that specific kind of work is collapsing.

The people who will be fine in every one of those three scenarios are the ones on the capital side of the labour-share line. Not because they are smarter. Because they own something.

That was the conclusion I reached in 2018, sitting in an office in Africa watching neural networks outperform media buyers. I did not have a model with three scenarios and a chart. I had a hunch and a seat close to the machine.

Eight years later, somebody finally did the maths. It says the avalanche is real, the only question is how fast it comes down the mountain, and the ground shifts under the people who work for a living either way.

Own something that has to exist in physical space.

Anthropic’s model and the three scenarios: https://www.anthropic.com/institute/econ-scenarios Their thread on it: https://x.com/AnthropicAI/status/2097679796687769689