DeepSeek and the Humility of Moats

On Monday, Nvidia lost nearly $600 billion in market value in a single day. The largest one-day loss for any company in history, as far as anyone can tell.

The cause was a model from a Chinese lab called DeepSeek. It reasons about as well as the best American models, it’s open, and the company reports that training its base model cost a few million dollars of compute - a small fraction of what everyone assumed was required. The app went to the top of the App Store over the weekend. By Monday morning, a lot of very confident people were reconsidering.

I don’t know what DeepSeek means for Nvidia’s stock. Nobody does, including the people on television explaining it. But I think there’s a quieter lesson here, and it’s an old one.

Every moat is temporary

A week ago, the story was simple. Frontier AI requires enormous capital, therefore only a handful of companies can do it, therefore those companies have a moat. Last week, at the White House, a joint venture announced plans to spend up to half a trillion dollars on AI infrastructure. The moat was money, and the money was flowing.

Then a relatively small team did something clever, and the moat looked a lot shallower.

This has happened before, many times. The mainframe had a moat until the PC. The PC had a moat until the phone. Search had a moat until… well, we’ll see. Every generation believes its advantage is structural. Most advantages turn out to be temporary, and the ones that last are usually not the ones people were measuring.

What actually lasts

When I think about the companies and people I’ve seen endure over twenty-five years, what they had in common was rarely a single advantage. It was something harder to name:

A habit of learning faster than the world changed. Relationships built on trust over many years. Taste - knowing what’s worth building. A willingness to let go of the thing that made them successful when it stopped working.

None of those show up on a balance sheet. None of them can be bought with a GPU cluster. And none of them can be copied over a weekend by a clever lab on the other side of the world.

For those of us who build

If you’re building a product on AI, the lesson isn’t to panic about which model is winning this week. It’s the opposite: assume the models will keep getting better, cheaper, and more interchangeable. Assume that whatever you’re doing with today’s model, someone will be able to do more cheaply tomorrow.

Then ask: what’s left? What do I bring that doesn’t get cheaper when intelligence does?

Usually the answer is: understanding a specific problem deeply, caring about specific people, and being trustworthy over time. Those were always the real moats. The technology just keeps reminding us.

On certainty

What strikes me most about this week isn’t the technology. It’s how certain everyone was, a week ago, about a future that changed in a weekend. Analysts, investors, executives - all very smart people, all extrapolating confidently from what they could see.

Kohelet wrote, a very long time ago, that there is nothing new under the sun. I used to read that as pessimism. I read it now as comfort. The specific surprises are new. The pattern of surprise is ancient. Things that seem permanent turn out not to be, and the people who do well are the ones who hold their certainties a little more loosely.

I’ve been wrong about technology many times. I expect to be wrong again, probably this year. The best I can do is be wrong quickly, admit it, and keep learning.

The moat you can see is rarely the one that protects you.